{"schema_version":1,"checked":"2026-10-17","counts":{"total":774,"main":451,"supplement_only":323,"technical":287,"fulltext":761,"resource_stage_checked":357,"no_weight_update":41,"comparable_ft":251,"known_ft_gpu_count":222,"known_ft_device_hours":93,"status":{"已核对资源阶段":357,"未定位可比微调配置":371,"无需本文权重训练":33,"全文获取受限":13},"reasoning":109},"rules":["去重 774 篇：Topics 主库451篇、技术标注池287篇、R1/R2/R3推理清单109篇，集合之间有重叠，不能直接相加。会议趋势的批量标题筛选库不属于逐篇资源核查清单。","逐篇进行全文资源关键词检索与上下文筛查；仅阶段明确的记录进入比较。未定位不等于原文未报告，获取受限不等于无训练。","保留原 taxonomy；新增资源字段不改变任务、贡献、condition、目标、生成范式、论文类型、证据类型与任务定义。","硬件分档为 [推断]：单卡、2–8 卡、9–64 卡、65+ 卡。按一篇论文已披露的相关微调阶段最高并行卡数归档，每篇只计一次；不是最低可复现门槛或完整 pipeline 成本。","微调／适配含 full fine-tuning、LoRA、adapter 与学习控制向量；预训练、辅助模型、逐实例 3D 优化、推理、数据生成和最低配置声明分开列出，不混入主微调统计。","GPU-hours 只有设备数和时长属于同一阶段时才相乘，标记 [推算]；原文直接报告总数另保留。不把不同模型变体相加，不将多阶段的某一小阶段称作全流程成本。","GPU 型号和显存差异很大；GPU-hours 未归一化，不是费用或 FLOPs 排名。TPU / PPU 单列。卡容量不等于实测峰值；缺失保持为空，不能按 GPU 型号补填。","零新增权重训练仍可能有 API、搜索、渲染、rollout、数据采集和用户实验成本；完整研究的调参/多 seeds 也可能远高于单次 run。","现有论文池是研究选题样本，不是领域随机样本；按 topic 的占比描述本论文池内的已披露证据，不能推出整个领域的成本分布。"],"role_labels":{"ft":"微调／适配训练","aux":"辅助模型训练","pre":"预训练／从零训练","opt":"逐实例表示优化","inf":"推理／评测／rollout","claim":"作者声明的最低可运行配置"},"records":[{"id":"R0001","title":"Video-to-Video Synthesis","year":"2018","venue":"NeurIPS","arxiv_id":"1808.06601","aliases":["pool:P001"],"technical_key":"arxiv-1808.06601","catalog_id":"P001","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/1808.06601","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"V100","count_min":8.0,"count_max":8.0,"card_memory_gb":16.0,"peak_memory_gb":null,"wall_hours_min":240.0,"wall_hours_max":240.0,"time_bound":"approximate","gpu_hours_min":1920.0,"gpu_hours_max":1920.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"vid2vid 从零训练；2K 分辨率；约 10 天","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/1808.06601","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"vid2vid；多尺度条件 GAN","training_datasets":"Cityscapes、Apolloscape；人脸与舞蹈分支数据名称待核","existing_method":"","training_signal":"图像／视频判别器损失、光流监督、特征匹配","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1dc84003f9cc28deacaa52da6561a84c00ca9ed00ef849b74f5504ae33f13185","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0002","title":"Dreamix: Video Diffusion Models are General Video Editors","year":"2023","venue":"不确定（arXiv preprint）","arxiv_id":"2302.01329","aliases":["pool:P002"],"technical_key":"arxiv-2302.01329","catalog_id":"P002","library_id":"","topic_code":"","topic":"","task":"编辑与修改","directions":[],"source_url":"https://arxiv.org/html/2302.01329","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Imagen Video 级联视频扩散模型","training_datasets":"每次编辑的输入视频／主体图像；底座预训练语料另计","existing_method":"","training_signal":"视频重建损失＋屏蔽时间模块的逐帧重建损失","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"cfdd21813b8e7d161589c983735473eff8033c19789ed656caa3246f8ef2a0c5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0003","title":"VideoComposer: Compositional Video Synthesis with Motion Controllability","year":"2023","venue":"NeurIPS","arxiv_id":"2306.02018","aliases":["pool:P003"],"technical_key":"arxiv-2306.02018","catalog_id":"P003","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2306.02018","status":"未定位可比微调配置","resource_note":"只定位到 400K steps 和图像/视频 GPU 配比，不能由配比推算设备总数。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"VideoComposer","training_datasets":"具体训练数据清单待本轮核实","existing_method":"","training_signal":"多条件随机组合下的扩散去噪训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"62093fc123652bafdaa71e9d75047c82dee0a3b0cc205e15e76184c1b2e828d7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0004","title":"SparseCtrl: Adding Sparse Controls to Text-to-Video Diffusion Models","year":"2023","venue":"ECCV","arxiv_id":"2311.16933","aliases":["pool:P004"],"technical_key":"arxiv-2311.16933","catalog_id":"P004","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2311.16933","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"AnimateDiff／个性化 Stable Diffusion 底座","training_datasets":"WebVid-10M","existing_method":"","training_signal":"冻结生成底座，训练稀疏条件编码器","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b4fa61c9893bd41b859c92516e1761fd412091df5afbbeec143cc6877e253945","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0005","title":"MotionCtrl: A Unified and Flexible Motion Controller for Video Generation","year":"2023","venue":"SIGGRAPH Conference Proceedings","arxiv_id":"2312.03641","aliases":["pool:P005"],"technical_key":"arxiv-2312.03641","catalog_id":"P005","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2312.03641","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"V100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"相机控制器 50K steps；对象控制器 20K+20K steps；未定位耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2312.03641","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"LVDM；另有 SVD 版本","training_datasets":"Augmented-RealEstate10K；合成物体轨迹数据","existing_method":"","training_signal":"分阶段学习相机模块与物体运动模块","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"4167c8ca81a9a862942dc3d7602bffd9da9d92a2e24cbd79e524f42549ad12bd","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0006","title":"CameraCtrl: Enabling Camera Control for Video Diffusion Models","year":"2024","venue":"ICLR","arxiv_id":"2404.02101","aliases":["pool:P006"],"technical_key":"arxiv-2404.02101","catalog_id":"P006","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2404.02101","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":16.0,"count_max":16.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":25.0,"wall_hours_max":25.0,"time_bound":"approximate","gpu_hours_min":400.0,"gpu_hours_max":400.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"T2V camera adapter；50K steps","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2404.02101","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":32.0,"count_max":32.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":40.0,"wall_hours_max":40.0,"time_bound":"approximate","gpu_hours_min":1280.0,"gpu_hours_max":1280.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"I2V/SVD camera encoder + merging linear；50K steps；与 T2V 是不同配置，不相加","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2404.02101","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"AnimateDiff；LoRA 适配底座","training_datasets":"RealEstate10K","existing_method":"","training_signal":"冻结／适配底座后训练 camera encoder 与 attention 模块","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"dd11f0a307d5a73fd29b2a434f5c08384c6623732b5f5c66bfcb23e4dae63e52","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0007","title":"PhysGen: Rigid-Body Physics-Grounded Image-to-Video Generation","year":"2024","venue":"ECCV","arxiv_id":"2409.18964","aliases":["pool:P007"],"technical_key":"arxiv-2409.18964","catalog_id":"P007","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2409.18964","status":"已核对资源阶段","resource_note":"PhysGen 明确采用 training-free 的物理模拟与视频生成流程。","stages":[{"role":"inf","device":"A40","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Training-free 物理模拟与生成；单次推理约 3 分钟，不能记作训练时间","comparable":false,"is_gpu":true,"role_label":"推理／评测／rollout","source_url":"https://arxiv.org/html/2409.18964","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"PhysGen；物理模拟＋视频扩散细化","training_datasets":"无新增训练；使用现成模型与物理引擎","existing_method":"","training_signal":"不适用：training-free 系统","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"096ad49fd71c8daee7ca837249b49adba32c13a5e767f3357e1c502e2be2fe08","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0008","title":"MIMO: Controllable Character Video Synthesis with Spatial Decomposed Modeling","year":"2024","venue":"CVPR","arxiv_id":"2409.16160","aliases":["pool:P008"],"technical_key":"arxiv-2409.16160","catalog_id":"P008","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2409.16160","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"SD1.5 UNet、pose encoder、reference net；冻结 VAE/CLIP；约 50K steps","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2409.16160","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"SD 1.5＋AnimateDiff；reference-net","training_datasets":"HUD-7K：5K 真实视频＋2K 合成动画","existing_method":"","training_signal":"分解内容条件下的视频去噪训练","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"4abec11377fbfde72f3878096e6f6404df85fefdd67f8734677234d78c462549","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0009","title":"EchoMimic: Lifelike Audio-Driven Portrait Animations through Editable Landmark Conditions","year":"2024","venue":"AAAI","arxiv_id":"2407.08136","aliases":["pool:P009"],"technical_key":"arxiv-2407.08136","catalog_id":"P009","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2407.08136","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"两阶段各 30K steps；batch 4；512²；未定位训练时长","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2407.08136","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"EchoMimic；SD v1.5 Reference U-Net；Wav2Vec 音频编码","training_datasets":"约 540 小时互联网视频（130K clips）；HDTF、CelebV-HQ","existing_method":"","training_signal":"音频增强、随机 landmark 选择与分阶段去噪训练","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"9cfe26a3aef61739aecb15a4c020c64023877c5adaf2a01424310bc3c68c7d94","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0010","title":"HunyuanVideo: A Systematic Framework For Large Video Generative Models","year":"2024","venue":"不确定（arXiv preprint）","arxiv_id":"2412.03603","aliases":["pool:P010"],"technical_key":"arxiv-2412.03603","catalog_id":"P010","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2412.03603","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"HunyuanVideo 13B","training_datasets":"分层筛选的五级自建训练集；完整来源与规模待本轮核实","existing_method":"","training_signal":"图像／视频分阶段训练；具体语料配比另核","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"78c401281e2a3a9440d0921ab88f196b9077867ada4464fb7f3a832ae39b2a6d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0011","title":"VACE: All-in-One Video Creation and Editing","year":"2025","venue":"ICCV","arxiv_id":"2503.07598","aliases":["pool:P011"],"technical_key":"arxiv-2503.07598","catalog_id":"P011","library_id":"","topic_code":"","topic":"","task":"不确定","directions":[],"source_url":"https://arxiv.org/html/2503.07598","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VACE LTX-2B 配置；不要与 Wan-14B 配置相加","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2503.07598","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":128.0,"count_max":128.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VACE Wan2.1-14B 配置；推理延迟不能作为训练耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2503.07598","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"LTX-Video；Wan-T2V","training_datasets":"多任务视频数据；完整来源与各任务配比待核实","existing_method":"","training_signal":"从补全／延长逐步扩展到多参考、多任务训练","hardware_bin":"large","max_reported_ft_gpus":128.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"dc8cd0634a6896202b02bc507030e14433ed42a87e8f685c1b426a3f23f2e9d7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0012","title":"MotionAgent: Fine-grained Controllable Video Generation via Motion Field Agent","year":"2025","venue":"ICCV","arxiv_id":"2502.03207","aliases":["pool:P012","reasoning:ICCV-2025-1614"],"technical_key":"arxiv-2502.03207","catalog_id":"P012","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2502.03207","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A800","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"冻结 SVD、训练 optical-flow adapter；24 frames，512²","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2502.03207","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Stable Video Diffusion；GPT-4o motion agent","training_datasets":"光流 adapter 微调数据的完整来源待核实","existing_method":"","training_signal":"冻结 SVD，训练光流 adapter","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"bf2b06f44e5ebc21acbf64df10a6142047ff96de5d21e3d71207b53bb5f15207","reasoning_id":"ICCV-2025-1614","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0013","title":"OmniHuman-1: Rethinking the Scaling-Up of One-Stage Conditioned Human Animation Models","year":"2025","venue":"ICCV","arxiv_id":"2502.01061","aliases":["pool:P013"],"technical_key":"arxiv-2502.01061","catalog_id":"P013","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2502.01061","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":400.0,"count_max":400.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":240.0,"wall_hours_max":240.0,"time_bound":"approximate","gpu_hours_min":96000.0,"gpu_hours_max":96000.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"OmniHuman 生成模型训练；每个 phase 约 10 天；不推算未知 phase 的总和","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2502.01061","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"OmniHuman","training_datasets":"混合人物视频语料；具体数据来源清单待核实","existing_method":"","training_signal":"弱／强条件混合与分阶段训练；降低强条件任务采样比例","hardware_bin":"large","max_reported_ft_gpus":400.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"858b8054d368d6fa6539cd1fc5d7ddd7bff700fad2c03c4dc5b3fbfeab7beac9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0014","title":"Any2Caption:Interpreting Any Condition to Caption for Controllable Video Generation","year":"2025","venue":"不确定（arXiv preprint）","arxiv_id":"2503.24379","aliases":["pool:P014"],"technical_key":"arxiv-2503.24379","catalog_id":"P014","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2503.24379","status":"已核对资源阶段","resource_note":"","stages":[{"role":"aux","device":"GPU（型号未定位）","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Qwen2-VL captioner 适配；多阶段；不是视频生成器训练","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2503.24379","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen2-VL-7B 条件解释器；下游视频模型可替换","training_datasets":"Any2CapIns：多种输入条件与指令配对数据；各底层来源待核","existing_method":"","training_signal":"训练多模态条件到 caption 的映射，不等于训练视频底座","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e62dd9848b388540ee51e8e87af2c72f5212d5a36dd3db8edee087379f13c877","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0015","title":"GEN3C: 3D-Informed World-Consistent Video Generation with Precise Camera Control","year":"2025","venue":"CVPR","arxiv_id":"2503.03751","aliases":["pool:P015"],"technical_key":"arxiv-2503.03751","catalog_id":"P015","library_id":"","topic_code":"","topic":"","task":"3D/4D 一致生成","directions":[],"source_url":"https://arxiv.org/html/2503.03751","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":96.0,"wall_hours_max":96.0,"time_bound":"approximate","gpu_hours_min":3072.0,"gpu_hours_max":3072.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"GEN3C 从 SVD 适配 temporal layers；全训练约 4 天；HTML 数学双重渲染已按原数字去重","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2503.03751","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Stable Video Diffusion；扩展实验另用 Cosmos","training_datasets":"RealEstate10K、DL3DV、Kubric4D；第 4 个数据集待核实","existing_method":"","training_signal":"用 cache 渲染与真实目标视频的配对数据微调","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"73db52a9ad318734dba385ad07e19ed75e0ee7a5e6ed37b18fbdafb93301207a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0016","title":"VideoMage: Multi-Subject and Motion Customization of Text-to-Video Diffusion Models","year":"2025","venue":"CVPR","arxiv_id":"2503.21781","aliases":["pool:P016"],"technical_key":"arxiv-2503.21781","catalog_id":"P016","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2503.21781","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"VideoMage；视频底座版本待核实","training_datasets":"每个主体的参考图及参考运动视频；逐实例 LoRA 训练","existing_method":"","training_signal":"主体重建＋appearance-agnostic motion loss；学习文本 token","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"25ec9874bfc57307a20fef2cc41c31fa118e95bfa86c96221eb08289139ae758","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0017","title":"Hand2World: Autoregressive Egocentric Interaction Generation via Free-Space Hand Gestures","year":"2026","venue":"不确定（arXiv preprint）","arxiv_id":"2602.09600","aliases":["pool:P017"],"technical_key":"arxiv-2602.09600","catalog_id":"P017","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2602.09600","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Hand2World camera adapter 10K steps；adapter + rank-256 LoRA 100K steps；蒸馏阶段耗时另缺","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2602.09600","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.1-1.3B-Control","training_datasets":"训练集名称与来源待核实","existing_method":"","training_signal":"先训练 camera adapter，再联合 LoRA 微调","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"c29c7512844d415796b67caa72555611ce61a4308c611ad63111dc093438425f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0018","title":"TrajectoryMover: Generative Movement of Object Trajectories in Videos","year":"2026","venue":"不确定（arXiv preprint）","arxiv_id":"2603.29092","aliases":["pool:P018"],"technical_key":"arxiv-2603.29092","catalog_id":"P018","library_id":"","topic_code":"","topic":"","task":"编辑与修改","directions":[],"source_url":"https://arxiv.org/html/2603.29092","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"TrajectoryMover 3,200 steps；batch 16；31,629 训练 pairs","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.29092","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.1-T2V-1.3B","training_datasets":"TrajectorySynth 合成配对数据＋真实视频交错训练","existing_method":"","training_signal":"合成轨迹编辑监督与普通真实视频生成交替训练","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"a205588aef7f7e6ee4105c05852a4e6bfd6bd62d368fb02c048003b01599e8ea","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0019","title":"Tune-A-Video: One-Shot Tuning of Image Diffusion Models for Text-to-Video Generation","year":"2022","venue":"ICCV","arxiv_id":"2212.11565","aliases":["pool:P019","library:P166"],"technical_key":"arxiv-2212.11565","catalog_id":"P019","library_id":"P166","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"编辑与修改","directions":[],"source_url":"https://arxiv.org/html/2212.11565","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":0.166667,"wall_hours_max":0.166667,"time_bound":"approximate","gpu_hours_min":0.1667,"gpu_hours_max":0.1667,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Tune-A-Video 每个输入视频 300 steps、约 10 分钟；batch 1；不是通用模型从零训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2212.11565","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Stable Diffusion v1.4","training_datasets":"单个输入视频—文本对；DAVIS 的 42 段视频用于实例测试","existing_method":"Tune-A-Video 通过时间注意力和单样本微调学习运动，使用反演提供结构条件。 本条方法定位来自摘要，训练细节待核。","training_signal":"单视频去噪重建；只更新部分 attention 参数","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"c9fbb48cf3670dbeb65b2dd1ea2c3f2987419345cad1e7ed703d177798029545","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0020","title":"VBench: Comprehensive Benchmark Suite for Video Generative Models","year":"2023","venue":"CVPR","arxiv_id":"2311.17982","aliases":["pool:P020"],"technical_key":"arxiv-2311.17982","catalog_id":"P020","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2311.17982","status":"已核对资源阶段","resource_note":"","stages":[{"role":"aux","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":1.0,"wall_hours_max":1.0,"time_bound":"approximate","gpu_hours_min":8.0,"gpu_hours_max":8.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"VBench 为评测适配 VideoChat；30K instruction pairs，3 epochs；不是生成器 fine-tuning","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2311.17982","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"多种被测 T2V 模型；RAFT／ViCLIP／MUSIQ 等评价器","training_datasets":"不新增训练生成器；评价器的预训练集另计","existing_method":"","training_signal":"不适用：评测贡献","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0d161fae4faaddbfcddad63e660a1226ccde54d2675a5d7fee9516a09a4c3462","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0021","title":"SafeSora: Towards Safety Alignment of Text2Video Generation via a Human Preference Dataset","year":"2024","venue":"NeurIPS Datasets and Benchmarks","arxiv_id":"2406.14477","aliases":["pool:P021"],"technical_key":"arxiv-2406.14477","catalog_id":"P021","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2406.14477","status":"已核对资源阶段","resource_note":"","stages":[{"role":"aux","device":"H800","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":2.0,"wall_hours_max":2.0,"time_bound":"approximate","gpu_hours_min":16.0,"gpu_hours_max":16.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"SafeSora moderation VLM；3 epochs","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2406.14477","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"H800","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":2.0,"wall_hours_max":2.0,"time_bound":"approximate","gpu_hours_min":16.0,"gpu_hours_max":16.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"SafeSora preference/reward model；3 epochs","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2406.14477","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H800","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":1.0,"wall_hours_max":1.0,"time_bound":"approximate","gpu_hours_min":8.0,"gpu_hours_max":8.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"SafeSora video diffusion refiner；1 epoch；生成偏好数据的成本另计","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2406.14477","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Video-LLaVA；Vicuna-7B v1.5＋LanguageBind","training_datasets":"SafeSora 多标签与偏好数据；审核器使用其训练划分","existing_method":"","training_signal":"多标签安全监督与人类偏好标注","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"5845b404a449ecd94f6bcf7a701975670b7b4e07307f1b35bf42ff0b961aaa3d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0022","title":"T2V-Turbo: Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback","year":"2024","venue":"NeurIPS","arxiv_id":"2405.18750","aliases":["pool:P022"],"technical_key":"arxiv-2405.18750","catalog_id":"P022","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2405.18750","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":6.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"T2V-Turbo 蒸馏；10K steps；正文与附录给 8 / 6–8 GPUs；不硬选单一数值","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2405.18750","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"VideoCrafter2；ModelScopeT2V","training_datasets":"WebVid-10M","existing_method":"","training_signal":"一致性蒸馏＋图像／视频 reward 梯度","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"0bc5046903b25bd4e3c5c49df85320dd89560148951a700de822d401153b4f1d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0023","title":"Video Diffusion Alignment via Reward Gradients","year":"2024","venue":"不确定（arXiv preprint）","arxiv_id":"2407.08737","aliases":["pool:P023"],"technical_key":"arxiv-2407.08737","catalog_id":"P023","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2407.08737","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A6000","count_min":2.0,"count_max":2.0,"card_memory_gb":48.0,"peak_memory_gb":null,"wall_hours_min":12.0,"wall_hours_max":12.0,"time_bound":"approximate","gpu_hours_min":24.0,"gpu_hours_max":24.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"VADER 实际实验平均耗时；LoRA + 截断反传；不包括基础模型预训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2407.08737","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"claim","device":"GPU（型号未定位）","count_min":1.0,"count_max":1.0,"card_memory_gb":16.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"作者声称代码可在单卡 16GB 运行；不是上行实际实验的实测配置","comparable":false,"is_gpu":true,"role_label":"作者声明的最低可运行配置","source_url":"https://arxiv.org/html/2407.08737","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"VideoCrafter／ModelScope／OpenSora／SVD","training_datasets":"生成采样与 prompt；具体训练 prompt 来源待核实","existing_method":"","training_signal":"可微 reward 的截断反向传播；LoRA 更新","hardware_bin":"small","max_reported_ft_gpus":2.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"634bc1dd128f6ff990c9a5d6dc388f9421fed5530ce7e7957e3bba85a5bd543d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0024","title":"DOLLAR: Few-Step Video Generation via Distillation and Latent Reward Optimization","year":"2024","venue":"ICCV","arxiv_id":"2412.15689","aliases":["pool:P024"],"technical_key":"arxiv-2412.15689","catalog_id":"P024","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2412.15689","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"DOLLAR 每个蒸馏实验；最多 40K iterations；128 frames，192×320","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2412.15689","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"改造的 CogVideoX 式双向视频 DiT；非直接声明使用官方 checkpoint","training_datasets":"教师生成／蒸馏训练样本；具体语料清单待核实","existing_method":"","training_signal":"分布匹配／一致性蒸馏＋latent reward 优化","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"9819a95337e0226424f6a7844ce1c8bac30ed94ffe87aba36b40ab2810fb3f49","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0025","title":"Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion","year":"2025","venue":"NeurIPS","arxiv_id":"2506.08009","aliases":["pool:P025"],"technical_key":"arxiv-2506.08009","catalog_id":"P025","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2506.08009","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":1.5,"wall_hours_max":1.5,"time_bound":"approximate","gpu_hours_min":96.0,"gpu_hours_max":96.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Self Forcing 的 DMD 收敛阶段；约 96 GPU-hours；并行硬件门槛仍高","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2506.08009","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.1 系列；1.3B 自回归学生","training_datasets":"DMD／SiD 分支无需真实视频；GAN 分支用 14B 教师生成的 70k 视频","existing_method":"","training_signal":"自身 rollout 上的序列分布匹配；DMD／SiD／GAN 为不同设置","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"50fe4a43acb9deaf0fd9b7a4b969957d5e2e86bcbcb2259de211ec75958dbbd0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0026","title":"VideoREPA: Learning Physics for Video Generation through Relational Alignment with Foundation Models","year":"2025","venue":"NeurIPS","arxiv_id":"2505.23656","aliases":["pool:P026","external:XR-REF-VIDEOREPA"],"technical_key":"arxiv-2505.23656","catalog_id":"P026","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2505.23656","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VideoREPA：2B full fine-tuning 4K steps / 5B LoRA 2K steps；OpenVid","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2505.23656","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CogVideoX-2B／5B；VideoMAEv2 teacher","training_datasets":"OpenVid 子集：2B 用 32k，5B 用 64k","existing_method":"","training_signal":"Token Relation Distillation＋原去噪损失","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"c90e4ec7907023526a9effbe47926e8a48c35b3ecebe0c5388498020befd488e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0027","title":"RewardDance: Reward Scaling in Visual Generation","year":"2025","venue":"不确定（arXiv preprint）","arxiv_id":"2509.08826","aliases":["pool:P027"],"technical_key":"arxiv-2509.08826","catalog_id":"P027","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2509.08826","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"RewardDance；1B–26B 多规模 reward model","training_datasets":"任务感知 CoT／偏好指令数据；完整来源待核实","existing_method":"","training_signal":"CoT 偏好训练；下游 BoN／ReFL 使用评分","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"82b44e679446d52369ea9ebb2fa75771eeae6a7c4f0fe45aa31c97c63d843aee","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0028","title":"Improving Video Generation with Human Feedback","year":"2025","venue":"NeurIPS","arxiv_id":"2501.13918","aliases":["pool:P028"],"technical_key":"arxiv-2501.13918","catalog_id":"P028","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2501.13918","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A800","count_min":16.0,"count_max":16.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"SFT / Flow-RWR / Flow-DPO；LoRA rank 64；1 epoch；不同方法配置不相加","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2501.13918","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"A800","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"奖励 VLM：vision 全参数 + language LoRA；2 epochs；与生成器训练分开","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2501.13918","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"VideoAlign／VideoReward；生成底座按实验区分","training_datasets":"16k prompts、108k 生成视频、182k 人类偏好 triplets","existing_method":"","training_signal":"视觉质量、运动质量与文本对齐的人类偏好监督","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"07ca926f9a70f4994fa900bfd14255d368d5cd6ea6cfa911c8793a574f74162d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0029","title":"Video Generation Models Are Good Latent Reward Models","year":"2025","venue":"不确定（arXiv preprint）","arxiv_id":"2511.21541","aliases":["pool:P029"],"technical_key":"arxiv-2511.21541","catalog_id":"P029","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2511.21541","status":"未定位可比微调配置","resource_note":"正文命中的单 GPU 30 分钟描述数据生成，不能当作 PAVRM 训练成本。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"PAVRM；使用视频 DiT 前 8 个 blocks","training_datasets":"二元运动质量偏好数据；完整来源与规模待核实","existing_method":"","training_signal":"带噪 latent 上的偏好／运动质量分类监督","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4b3a3d29861943c1ffe13d1f0d05d93394f5ce7940b3498d547a54a46bb10b55","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0030","title":"Respond Beyond Language: A Benchmark for Video Generation in Response to Realistic User Intents","year":"2025","venue":"不确定（arXiv preprint）","arxiv_id":"2506.01689","aliases":["pool:P030"],"technical_key":"arxiv-2506.01689","catalog_id":"P030","library_id":"","topic_code":"","topic":"","task":"视频作为推理或回答","directions":[],"source_url":"https://arxiv.org/html/2506.01689","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"多种被测视频生成器；GPT-3.5-Turbo 做意图识别","training_datasets":"无新增生成器训练；用户意图问题经筛选构建评测","existing_method":"","training_signal":"不适用：benchmark","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f8c04c249fb5060302127bb0cbcb21e3326ac0a03c2f264f471332245c174396","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0031","title":"Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation","year":"2026","venue":"ICML","arxiv_id":"2602.02214","aliases":["pool:P031"],"technical_key":"arxiv-2602.02214","catalog_id":"P031","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2602.02214","status":"未定位可比微调配置","resource_note":"硬件命中包含 Self Forcing baseline 的推理速度；未据此给 Causal Forcing 赋予训练配置。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.1-T2V-1.3B","training_datasets":"3k 双向模型合成样本；因果 ODE 轨迹数据","existing_method":"","training_signal":"Teacher forcing 因果训练＋因果 ODE 初始化＋consistency distillation","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"8528378ddfacca91051f33811544ac8ed15623cc11bd4d5b0f33bbc249db266e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0032","title":"Learning to Credit the Right Steps: Objective-aware Process Optimization for Visual Generation","year":"2026","venue":"不确定（arXiv preprint）","arxiv_id":"2604.19234","aliases":["pool:P032"],"technical_key":"arxiv-2604.19234","catalog_id":"P032","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2604.19234","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"OPO 视频 Wan2.2-14B 配置；Flux 图像配置另用 8 H100，不计入视频配置","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2604.19234","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-T2V；图像分支 FLUX.1-dev","training_datasets":"训练 prompt 与在线生成 rollout；完整 prompt 来源待核实","existing_method":"","training_signal":"目标相关 timestep credit＋多目标 advantage／梯度组合","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"f94affdb3bd3c3f1e8913b4474f119f397ba4b6fa5240b4260f6f1c02b751d35","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0033","title":"Manifold-Aware Exploration for Reinforcement Learning in Video Generation","year":"2026","venue":"不确定（arXiv preprint）","arxiv_id":"2603.21872","aliases":["pool:P033"],"technical_key":"arxiv-2603.21872","catalog_id":"P033","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2603.21872","status":"未定位可比微调配置","resource_note":"已定位 per-GPU batch 和梯度累积，但未定位 GPU 数；不能用有效 batch 反推数量。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"HunyuanVideo 1.5；VideoAlign reward","training_datasets":"在线生成 rollout；训练 prompt 来源待核实","existing_method":"","training_signal":"流形感知探索＋GRPO；VQ／MQ／TA 组合 reward","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"aee8960392b2be765230c0c0949e84604a0c1696599dece6025f2ed5e4cb2776","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0034","title":"Wan-R1: Verifiable-Reinforcement Learning for Video Reasoning","year":"2026","venue":"不确定（arXiv preprint）","arxiv_id":"2603.27866","aliases":["pool:P034"],"technical_key":"arxiv-2603.27866","catalog_id":"P034","library_id":"","topic_code":"","topic":"","task":"视频作为推理或回答","directions":[],"source_url":"https://arxiv.org/html/2603.27866","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Wan-R1 LoRA rank 32，1 epoch，group 8，193 frames；未定位 GPU 数和耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.27866","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-TI2V-5B","training_datasets":"VR-Bench 简单任务训练划分；Target-Bench 导航设置","existing_method":"","training_signal":"SFT＋可验证任务 reward 的 GRPO","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"7352065150f5af5aa40b0a8b016c2bac5c1d846474d0bad0ad582ab61ff19b7b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0035","title":"CaC: Advancing Video Reward Models via Hierarchical Spatiotemporal Concentrating","year":"2026","venue":"NeurIPS","arxiv_id":"2605.11723","aliases":["pool:P035"],"technical_key":"arxiv-2605.11723","catalog_id":"P035","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2605.11723","status":"已核对资源阶段","resource_note":"","stages":[{"role":"aux","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":11.5,"wall_hours_max":11.5,"time_bound":"approximate","gpu_hours_min":92.0,"gpu_hours_max":92.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"CaC reward model LoRA warm-up；9,330 steps","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2605.11723","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":30.5,"wall_hours_max":30.5,"time_bound":"approximate","gpu_hours_min":244.0,"gpu_hours_max":244.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"CaC reward model multi-frame full SFT；7,000 steps","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2605.11723","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"H200","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":42.0,"wall_hours_max":42.0,"time_bound":"approximate","gpu_hours_min":672.0,"gpu_hours_max":672.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"CaC reward model GRPO；250 steps","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2605.11723","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H200","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"CaC 下游 video generator Flow-GRPO；LoRA rank 32；120 steps；耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.11723","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen3-VL-8B","training_datasets":"三阶段互不重叠的异常定位／视频判别数据；完整来源待核实","existing_method":"","training_signal":"单帧定位 warmup＋多帧 SFT＋多轮 RL","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"fa9b622028646f9e251e02257946f70b96f8e7bf1561ad60fad1b9adbd29bc62","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0036","title":"Video Models Can Reason with Verifiable Rewards","year":"2026","venue":"NeurIPS","arxiv_id":"2605.15458","aliases":["pool:P036"],"technical_key":"arxiv-2605.15458","catalog_id":"P036","library_id":"","topic_code":"","topic":"","task":"视频作为推理或回答","directions":[],"source_url":"https://arxiv.org/html/2605.15458","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"B200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VideoRLVR Wan2.2-TI2V-5B；30K synthetic tasks；原文 HTML 的 88 是 8 的重复渲染","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.15458","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-TI2V-5B","training_datasets":"30k 合成规则任务样本；3k 独立测试样本","existing_method":"","training_signal":"终态正确 reward＋过程合法 reward；先 SFT 后 RL","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"bf5569a4fc07e9e7db7048e73157fbf26fda1efed2250afeb03dd74ae9f40b19","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0037","title":"Learning to Simulate Dynamic Environments with GameGAN","year":"2020","venue":"CVPR","arxiv_id":"2005.12126","aliases":["pool:P037"],"technical_key":"arxiv-2005.12126","catalog_id":"P037","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2005.12126","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"GameGAN；动态模块＋神经渲染器","training_datasets":"游戏代理采集的 Pacman／Pacman Maze 游戏轨迹；完整实验变体见原文","existing_method":"","training_signal":"对抗训练＋重建／动态约束","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4cd51344a3113a9e9f5dcd296d105d8f469e568b5a84d5dba5017f09e77f048a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0038","title":"Learning Interactive Real-World Simulators","year":"2023","venue":"ICLR","arxiv_id":"2310.06114","aliases":["pool:P038"],"technical_key":"arxiv-2310.06114","catalog_id":"P038","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2310.06114","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"TPU v3","count_min":512.0,"count_max":512.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":480.0,"wall_hours_max":480.0,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"UniSim 5.6B 世界模型训练；TPU chip-hours 单列，不换算 GPU-hours","comparable":false,"is_gpu":false,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2310.06114","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"TPU v3","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":24.0,"wall_hours_max":24.0,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"UniSim 下游 PaLM-E 12B policy 适配；与世界模型训练分开","comparable":false,"is_gpu":false,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2310.06114","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"UniSim","training_datasets":"网络视频、机器人轨迹、Habitat／HM3D 与模拟交互数据；全清单待核实","existing_method":"","training_signal":"统一动作条件的视频去噪监督","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"bd1b146cd45a8c107eef7b9cd736db0da64083e6c24cd04202283b973de0d45b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0039","title":"GAIA-1: A Generative World Model for Autonomous Driving","year":"2023","venue":"不确定（arXiv preprint）","arxiv_id":"2309.17080","aliases":["pool:P039"],"technical_key":"arxiv-2309.17080","catalog_id":"P039","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2309.17080","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":32.0,"count_max":32.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":96.0,"wall_hours_max":96.0,"time_bound":"approximate","gpu_hours_min":3072.0,"gpu_hours_max":3072.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"GAIA-1 tokenizer；200K steps","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2309.17080","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"pre","device":"A100","count_min":64.0,"count_max":64.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":360.0,"wall_hours_max":360.0,"time_bound":"approximate","gpu_hours_min":23040.0,"gpu_hours_max":23040.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"GAIA-1 6.5B world model；100K steps；decoder 成本不包含在本项","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2309.17080","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"GAIA-1","training_datasets":"Wayve 自采驾驶视频与动作／文本；规模待核实","existing_method":"","training_signal":"视频 token 预测＋条件视频解码训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f3250f00ce86a52f50ccad21bfb3d9ca532e7e6995dd30bdfce09e076bd783cd","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0040","title":"Genie: Generative Interactive Environments","year":"2024","venue":"ICML","arxiv_id":"2402.15391","aliases":["pool:P040","library:P123"],"technical_key":"arxiv-2402.15391","catalog_id":"P040","library_id":"P123","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2402.15391","status":"未定位可比微调配置","resource_note":"附录单 GPU / TPU 的 CoinRun 小规模示例不是完整 11B Genie 的训练成本。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Genie","training_datasets":"Platformers 网络平台游戏视频；另有机器人验证分支","existing_method":"潜动作模型从相邻帧推断离散控制，ST-transformer编码视频token，动态模型根据历史与动作预测下一帧；不需真实动作标签训练。","training_signal":"视频 tokenizer、latent action model 与 dynamics 联合流程","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3ae1055e8ed7c9e81addf34546c377c097e846d291cce1ea4a6a1467216154d4","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0041","title":"Diffusion Models Are Real-Time Game Engines","year":"2024","venue":"ICLR","arxiv_id":"2408.14837","aliases":["pool:P041"],"technical_key":"arxiv-2408.14837","catalog_id":"P041","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2408.14837","status":"已核对资源阶段","resource_note":"单 TPU 20/50 FPS 是 GameNGen 推理配置，不能据此描述训练资源。","stages":[{"role":"ft","device":"TPU v5e devices","count_min":128.0,"count_max":128.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"GameNGen 生成模型使用 128 TPU-v5e data parallel；不是单 TPU inference","comparable":true,"is_gpu":false,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2408.14837","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"GameNGen；Stable Diffusion 系列初始化","training_datasets":"DOOM 中 RL agent 收集的游戏轨迹","existing_method":"","training_signal":"动作条件去噪＋历史噪声增强","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"45dc22648809a21d8909114e9a8dd31b7afd5f01c2d92c5dee72ddbba77b21ce","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0042","title":"Diffusion for World Modeling: Visual Details Matter in Atari","year":"2024","venue":"NeurIPS","arxiv_id":"2405.12399","aliases":["pool:P042"],"technical_key":"arxiv-2405.12399","catalog_id":"P042","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2405.12399","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 4090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":12.0,"wall_hours_min":69.6,"wall_hours_max":69.6,"time_bound":"approximate","gpu_hours_min":69.6,"gpu_hours_max":69.6,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"DIAMOND Atari：单次训练 2.9 天；26 games × 5 seeds 的完整实验集另为 1.03 GPU-years","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2405.12399","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"pre","device":"RTX 4090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":288.0,"wall_hours_max":288.0,"time_bound":"approximate","gpu_hours_min":288.0,"gpu_hours_max":288.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"DIAMOND CSGO 381M 配置；12 天；不同任务配置不相加","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2405.12399","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"DIAMOND","training_datasets":"Atari 100k 交互；独立 CSGO 分支使用 87 小时录像","existing_method":"","training_signal":"环境交互数据上的扩散预测；另训练 reward／termination","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"84482bd6f6d34e5f2f9fd64322e8a79111e1f73c42afba736efbb336333d007d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0043","title":"AdaWorld: Learning Adaptable World Models with Latent Actions","year":"2025","venue":"ICML","arxiv_id":"2503.18938","aliases":["pool:P043","reasoning:ICML-2025-0008"],"technical_key":"arxiv-2503.18938","catalog_id":"P043","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2503.18938","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":0.008333,"wall_hours_max":0.008333,"time_bound":"upper","gpu_hours_min":0.0083,"gpu_hours_max":0.0083,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"AdaWorld nuScenes 的两层 action-interface MLP 适配；原文 <30 秒，是上界；不代表整个世界模型训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2503.18938","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"AdaWorld 基础 autoregressive world model 适配训练；80K steps；batch 64","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2503.18938","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Stable Video Diffusion 1.5B＋latent-action autoencoder","training_datasets":"Something-Something、Ego4D、Open X-Embodiment 等四个公开视频／机器人来源（第四项待核）；1,016 个 Gym Retro 环境轨迹","existing_method":"","training_signal":"latent action 表示学习＋动作条件世界模型训练","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"8bbcf78a2c956767d56fc2a1f61bc7e9da2ada3c88666f65f8ce728fa8cfb018","reasoning_id":"ICML-2025-0008","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0044","title":"WorldMem: Long-term Consistent World Simulation with Memory","year":2025,"venue":"NeurIPS 2025","arxiv_id":"2504.12369","aliases":["pool:P044","library:P135","lineage:WORLD"],"technical_key":"arxiv-2504.12369","catalog_id":"P044","library_id":"P135","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"动作条件的预测与交互世界模型","directions":["AIGC-B","XR-S"],"source_url":"https://arxiv.org/html/2504.12369v3","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"WorldMem Minecraft memory model；500K steps；每 GPU batch 4；型号和耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2504.12369v3","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Oasis／WorldMem","training_datasets":"Minecraft 交互数据；其他环境分支另核","existing_method":"为帧存姿态/时间，按当前目标状态检索并注意历史视图；扩散模型结合当前上下文与记忆生成。","training_signal":"带历史记忆的视频预测监督","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"1f35f4b1a4425c060ac91e6ecab0eeb853ec9b6a596e1f2bc00620ae4143e4f1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0045","title":"Ctrl-World: A Controllable Generative World Model for Robot Manipulation","year":"2025","venue":"ICLR","arxiv_id":"2510.10125","aliases":["pool:P045"],"technical_key":"arxiv-2510.10125","catalog_id":"P045","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2510.10125","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":768.0,"gpu_hours_max":1152.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"CtrlWorld 世界模型 fine-tuning；100K steps；batch 64","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2510.10125","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"H100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"CtrlWorld 下游 π0.5-DROID policy；2K steps；不是世界模型成本","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2510.10125","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Stable Video Diffusion 1.5B＋三层 action MLP","training_datasets":"DROID","existing_method":"","training_signal":"机器人动作—多视角视频配对监督","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"5b26abac842444715349fa7dde65e4d25b883ccfc2f6c5bd3796362cd743690e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0046","title":"Matrix-game 2.0: An open-source, real-time, and streaming interactive world model","year":"2025","venue":"不确定（arXiv preprint）","arxiv_id":"2508.13009","aliases":["pool:P046"],"technical_key":"arxiv-2508.13009","catalog_id":"P046","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2508.13009","status":"未定位可比微调配置","resource_note":"单 H100 25 FPS 是 Matrix-Game 2.0 推理；RTX 3090 出现在数据采集阶段。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"SkyReels-V2-I2V-1.3B／Wan2.1 架构","training_datasets":"自建多场景交互数据；来源与规模待核实","existing_method":"","training_signal":"动作条件基础训练＋自回归少步蒸馏","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a96de96a2fa4cf5ea762fcafddf540a786c8ede09b674fe91c0cb502ecdc2da6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0047","title":"The World is Your Canvas: Painting Promptable Events with Reference Images, Trajectories, and Text","year":"2025","venue":"不确定（arXiv preprint）","arxiv_id":"2512.16924","aliases":["pool:P047"],"technical_key":"arxiv-2512.16924","catalog_id":"P047","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2512.16924","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H800","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"WorldCanvas；9K steps；batch 64；280K triplets；480×832","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2512.16924","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-I2V-14B","training_datasets":"自建 280k trajectory–video–text triplets","existing_method":"","training_signal":"轨迹—视频—文本配对监督","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"d408304fa6492ba00be4f361ccd4001ba07a50ef66f28ac0b2ff3c33f36b03fc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0048","title":"Training Agents Inside of Scalable World Models","year":"2025","venue":"不确定（arXiv preprint）","arxiv_id":"2509.24527","aliases":["pool:P048"],"technical_key":"arxiv-2509.24527","catalog_id":"P048","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2509.24527","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"TPU v5p","count_min":256.0,"count_max":1024.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Dreamer 4 世界模型训练；tokenizer 400M + dynamics 1.6B；不以单卡推理描述训练成本","comparable":false,"is_gpu":false,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2509.24527","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Dreamer 4","training_datasets":"视频与少量动作配对的环境轨迹；完整环境数据清单待核实","existing_method":"","training_signal":"无动作视频预训练＋动作适配＋想象轨迹上的策略学习","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"dacd528df56fee96d33c7593ffbbd471dc5b71d8f11e68526caf50dd96fe2f93","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0049","title":"ActionSplice: In-Flight Action Editing for Interactive World Models","year":"2026","venue":"不确定（arXiv preprint）","arxiv_id":"2609.08230","aliases":["pool:P049"],"technical_key":"arxiv-2609.08230","catalog_id":"P049","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2609.08230","status":"无需本文权重训练","resource_note":"ActionSplice 在去噪过程中替换动作条件并干预当前生成，不新增模型训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"minWM–WanAction2V；HY-WM1.5（分别训练 corrector）","training_datasets":"自建可中断交互数据：120 train／30 validation／30 test scenes；训练 3 次、测试 5 次 interrupt","existing_method":"","training_signal":"训练轻量 counterfactual state corrector","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d298bac924770b16634171d84e60616a8542a80a17e158d473615d8de4ae4e66","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0050","title":"Echo-Memory: A Controlled Study of Memory in Action World Models","year":"2026","venue":"不确定（arXiv preprint）","arxiv_id":"2606.09803","aliases":["pool:P050"],"technical_key":"arxiv-2606.09803","catalog_id":"P050","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2606.09803","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"统一 video DiT substrate","training_datasets":"Context-as-Memory","existing_method":"","training_signal":"统一数据与训练预算下的受控记忆消融","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"88c33beb1dade5a470c68b65d383fd6780759a348e93466f4eb685f74673ce20","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0051","title":"WorldDirector: Building Controllable World Simulators with Persistent Dynamic Memory","year":"2026","venue":"不确定（arXiv preprint）","arxiv_id":"2607.02517","aliases":["pool:P051"],"technical_key":"arxiv-2607.02517","catalog_id":"P051","library_id":"","topic_code":"","topic":"","task":"不确定","directions":[],"source_url":"https://arxiv.org/html/2607.02517","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":64.0,"count_max":64.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":4608.0,"gpu_hours_max":4608.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"WorldDirector 完整 post-training pipeline；3K steps；8 nodes × 8 GPUs","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2607.02517","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"LingBot-World-Base","training_datasets":"自建动态场景数据；完整来源待核实","existing_method":"","training_signal":"带动态上下文的世界模型训练","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"ce83c26481358b6d5a60f071b4d81f73cb39b2ae373db9d0b360c34c33725291","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0052","title":"Precise Editing and Flexible Referencing for Interactable Worlds","year":"2026","venue":"不确定（arXiv preprint）","arxiv_id":"2609.34470","aliases":["pool:P052"],"technical_key":"arxiv-2609.34470","catalog_id":"P052","library_id":"","topic_code":"","topic":"","task":"编辑与修改","directions":[],"source_url":"https://arxiv.org/html/2609.34470","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"LingBot-World-Base","training_datasets":"自建全局／局部编辑与参考数据；具体来源待核实","existing_method":"","training_signal":"Teacher forcing 逐渐过渡到 self-resampling 历史训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"38009ec082f8195fbdf13b4f4034604288fee987d5c95561d011d6793cb6dfca","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0053","title":"WorldAttention: An Efficient Attention Architecture for Interactive Video World Models","year":"2026","venue":"不确定（arXiv preprint）","arxiv_id":"2609.34606","aliases":["pool:P053"],"technical_key":"arxiv-2609.34606","catalog_id":"P053","library_id":"","topic_code":"","topic":"","task":"长视频与多镜头叙事","directions":[],"source_url":"https://arxiv.org/pdf/2609.34606","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":30.0,"wall_hours_max":30.0,"time_bound":"approximate","gpu_hours_min":960.0,"gpu_hours_max":960.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"WorldAttention 完整训练过程；另需 192 CPU cores / 960GB CPU RAM","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2609.34606","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.1-T2V-1.3B","training_datasets":"Self-Forcing／LongLive 式训练；具体 prompt 与 rollout 来源待核实","existing_method":"","training_signal":"沿用自回归少步训练，主改动在 attention 结构","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"c719a89cb1eb094b2abd0726a7d596a8edc3bf7791d41546508e739fef5735e2","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0054","title":"WorldReward: Reward Modeling for Camera-Conditioned World Models","year":"2026","venue":"不确定（arXiv preprint）","arxiv_id":"2609.03952","aliases":["pool:P054"],"technical_key":"arxiv-2609.03952","catalog_id":"P054","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2609.03952","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"WorldReward","training_datasets":"相机条件视频偏好数据；完整来源待核实","existing_method":"","training_signal":"相机遵循与生成质量的 reward 学习","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9f0cc8a7d76247176bb1c482cc53127790fd29b44521a9031d24b085579cc0dc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0055","title":"From Slow Bidirectional to Fast Autoregressive Video Diffusion Models","year":"2024","venue":"CVPR","arxiv_id":"2412.07772","aliases":["pool:P055"],"technical_key":"arxiv-2412.07772","catalog_id":"P055","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2412.07772","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":3072.0,"gpu_hours_max":3072.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"从双向到 causal video diffusion 的初始化与蒸馏训练；约 2 天","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2412.07772","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CausVid；具体预训练版本待核实","training_datasets":"教师视频／ODE 轨迹；完整训练来源待核实","existing_method":"","training_signal":"双向教师指导因果学生的分布匹配蒸馏","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"97fa5e18eb0bf1a228416821302f8cab41b731b7971bc095d6c506d245fffdfb","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0056","title":"Mind the Time: Temporally-Controlled Multi-Event Video Generation","year":"2024","venue":"CVPR","arxiv_id":"2412.05263","aliases":["pool:P056"],"technical_key":"arxiv-2412.05263","catalog_id":"P056","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":[],"source_url":"https://arxiv.org/html/2412.05263","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Mind the Time 使用 A100；未定位数量或耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2412.05263","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"MinT；预训练 latent video DiT","training_datasets":"带时间事件与镜头切分的训练视频；数据来源待核实","existing_method":"","training_signal":"事件时间／scene cut 标注下的全模型微调","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"2a2e044b2a612c64895d8b709623c5ff542d7c799d297246ec338dfc1c64f6b1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0057","title":"T2V-CompBench: A Comprehensive Benchmark for Compositional Text-to-video Generation","year":"2024","venue":"CVPR","arxiv_id":"2407.14505","aliases":["pool:P057"],"technical_key":"arxiv-2407.14505","catalog_id":"P057","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2407.14505","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"多种被测 T2V；含 ModelScope 版 LVD","training_datasets":"无新增生成器训练；T2V-CompBench 为评测集","existing_method":"","training_signal":"不适用：评测贡献","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5331bbce1e0f86a23599c9d121b865b5363279da317723f61e074761236e521b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0058","title":"FFP-300K: Scaling First-Frame Propagation for Generalizable Video Editing","year":"2026","venue":"CVPR","arxiv_id":"2601.01720","aliases":["pool:P058"],"technical_key":"arxiv-2601.01720","catalog_id":"P058","library_id":"","topic_code":"","topic":"","task":"编辑与修改","directions":[],"source_url":"https://arxiv.org/html/2601.01720","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"FFP 编辑模型；具体 backbone 待核实","training_datasets":"FFP-300K","existing_method":"","training_signal":"编辑首帧—源视频—目标视频配对训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"29b7e3f6911fc85c82a731f4290662cf516697883cee128625523a68f7c978b0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0059","title":"VideoCoF: Unified Video Editing with Temporal Reasoner","year":"2025","venue":"CVPR","arxiv_id":"2512.07469","aliases":["pool:P059","reasoning:CVPR-2026-0435"],"technical_key":"arxiv-2512.07469","catalog_id":"P059","library_id":"","topic_code":"","topic":"","task":"编辑与修改","directions":[],"source_url":"https://arxiv.org/html/2512.07469","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VideoCoF；50K curated videos；8K iterations；推理 10 秒不计作训练时间","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2512.07469","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan-14B；DMD-LoRA 少步推理","training_datasets":"50k 编辑视频对；Pexels 构建数据＋Señorita-2M 筛选子集","existing_method":"","training_signal":"编辑配对监督＋区域 reasoning latent 学习","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"edc4778df09a1f3e4bda264d3715a22e8e2f2ba368833278426f509ab06eb102","reasoning_id":"CVPR-2026-0435","reasoning_role":"R1","reasoning_topic":"Editing / 修改已有视频"},{"id":"R0060","title":"Physical Simulator In-the-Loop Video Generation","year":"2026","venue":"CVPR","arxiv_id":"2603.06408","aliases":["pool:P060"],"technical_key":"arxiv-2603.06408","catalog_id":"P060","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2603.06408","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Go-with-the-Flow 视频生成器＋物理模拟器","training_datasets":"无新增生成器训练","existing_method":"","training_signal":"不适用：推理时物理校验与修订","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2abf07df372d04bf304ab6c04eeb14a08a70b84c630f811c9c05d3600c78384e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0061","title":"ReCamMaster: Camera-Controlled Generative Rendering from A Single Video","year":"2025","venue":"ICCV","arxiv_id":"2503.11647","aliases":["pool:P061"],"technical_key":"arxiv-2503.11647","catalog_id":"P061","library_id":"","topic_code":"","topic":"","task":"编辑与修改","directions":[],"source_url":"https://arxiv.org/html/2503.11647","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"内部预训练 T2V 模型；具体 checkpoint 未在已核对段落命名","training_datasets":"UE5 渲染的多相机同步视频数据","existing_method":"","training_signal":"源／目标相机视频配对监督","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4a29a4ca358280e0abca85a5f5dc5c9de637f33607e72169a859cb1a8fad6168","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0062","title":"VMem: Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory","year":"2025","venue":"ICCV","arxiv_id":"2506.18903","aliases":["pool:P062"],"technical_key":"arxiv-2506.18903","catalog_id":"P062","library_id":"","topic_code":"","topic":"","task":"3D/4D 一致生成","directions":[],"source_url":"https://arxiv.org/html/2506.18903","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A40","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VMem LoRA rank 256；600K iterations；batch 24 / GPU；未定位小时数","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2506.18903","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"SEVA","training_datasets":"RealEstate10K；记忆检索微调，训练 K=4、M=4","existing_method":"","training_signal":"主贡献为记忆检索；特定配置额外微调","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"04091b3c2ea0ed99aa2ba6b6d296fcfc59427edb4311cec3342829b466707851","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0063","title":"WorldScore: A Unified Evaluation Benchmark for World Generation","year":"2025","venue":"ICCV","arxiv_id":"2504.00983","aliases":["pool:P063","library:P287"],"technical_key":"arxiv-2504.00983","catalog_id":"P063","library_id":"P287","topic_code":"B1","topic":"能力诊断与评测基准","task":"3D/4D 一致生成","directions":[],"source_url":"https://arxiv.org/html/2504.00983","status":"未定位可比微调配置","resource_note":"WorldScore 的 H100 / L40S 描述评测生成硬件，未算入生成器 fine-tuning。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"多种被测 3D／视频世界生成模型","training_datasets":"无新增生成器训练；WorldScore 评测设置","existing_method":"WorldScore 以明确相机轨迹组织下一场景生成，分开评价控制、质量和动态。 本条方法定位来自摘要，训练细节待核。","training_signal":"不适用：评测贡献","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"14e721f8cf24e20598b6367ad701ac4c89c485ab9ddd76da1d14a35364deccba","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0064","title":"CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer","year":"2024","venue":"ICLR","arxiv_id":"2408.06072","aliases":["pool:P064"],"technical_key":"arxiv-2408.06072","catalog_id":"P064","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2408.06072","status":"未定位可比微调配置","resource_note":"CogVideoX 的 18–76GB 显存表是 inference memory，不计作训练显存。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CogVideoX-2B／5B","training_datasets":"自建过滤视频／图像语料；完整来源未在本记录核实","existing_method":"","training_signal":"混合时长／分辨率视频去噪训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"34d08944cc0eae4277d317dd9a9e708b3274c48ba7a42325ab823b63a63cde88","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0065","title":"OpenVid-1M: A Large-Scale High-Quality Dataset for Text-to-video Generation","year":"2024","venue":"ICLR","arxiv_id":"2407.02371","aliases":["pool:P065"],"technical_key":"arxiv-2407.02371","catalog_id":"P065","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2407.02371","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":336.0,"wall_hours_max":336.0,"time_bound":"lower","gpu_hours_min":10752.0,"gpu_hours_max":10752.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"OpenVid1M 的 256² 生成模型实验；至少 14 天，为下界；数据清洗 GPU 成本另计","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2407.02371","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"MVDiT","training_datasets":"OpenVid-1M；高清阶段 OpenVidHD-0.4M；WebVid／Panda 为数据对照","existing_method":"","training_signal":"高质量视频—文本配对去噪训练","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"0de65b9d956a2e356e2e4340d75671f55ff8b86c9db16b33879435c55e979137","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0066","title":"Pyramidal Flow Matching for Efficient Video Generative Modeling","year":"2024","venue":"ICLR","arxiv_id":"2410.05954","aliases":["pool:P066"],"technical_key":"arxiv-2410.05954","catalog_id":"P066","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2410.05954","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Pyramid Flow 报告模型训练合计 20,700 A100 GPU-hours；GPU 数和墙钟时长未据此反推","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2410.05954","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"pre","device":"A100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":20700,"gpu_hours_max":20700,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"模型训练合计","comparable":true,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2410.05954","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Pyramid Flow","training_datasets":"LAION-5B 美学子集、CC-12M、SA-1B、JourneyDB 等图像；视频来源另核","existing_method":"","training_signal":"跨分辨率 flow matching；历史低分辨率条件训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d5f0b7c82faf4933427b7c380e45e7439dacc902724b8bbe31ea3fcae5eec3b5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0067","title":"Unified In-Context Video Editing","year":"2025","venue":"ICLR","arxiv_id":"2506.04216","aliases":["pool:P067"],"technical_key":"arxiv-2506.04216","catalog_id":"P067","library_id":"","topic_code":"","topic":"","task":"编辑与修改","directions":[],"source_url":"https://arxiv.org/html/2506.04216","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H800","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"UNIC fine-tune DiT 与新 tokenizer；冻结 VAE/T5；16K iterations","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2506.04216","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"预训练 1B 视频生成器（28 个 DiT blocks）；名称待核","training_datasets":"自建 ID／风格编辑数据；Multi-Cam Video Dataset 等","existing_method":"","training_signal":"多任务配对视频训练","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"efeaf24234203dee6a0c62a03c7978e26f561f594201e5afdc80237a7df86b14","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0068","title":"Horizon Imagination: Efficient On-Policy Rollout in Diffusion World Models","year":"2026","venue":"ICLR","arxiv_id":"2602.08032","aliases":["pool:P068"],"technical_key":"arxiv-2602.08032","catalog_id":"P068","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":[],"source_url":"https://arxiv.org/html/2602.08032","status":"未定位可比微调配置","resource_note":"Horizon 的单 RTX 4090 profiling 与训练时长不能未经确认合并为单卡训练配置。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"基于 Cosmos 实现的动作条件 DiT","training_datasets":"Atari100K 与 Craftium 在线交互轨迹","existing_method":"","training_signal":"环境轨迹上的世界模型训练＋想象轨迹上的 actor–critic","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d0d3f02c0564007d81d6971e717437fba67abcc2fc23218379ff4ab78e1ea188","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0069","title":"LongLive: Real-time Interactive Long Video Generation","year":"2025","venue":"ICLR","arxiv_id":"2509.22622","aliases":["pool:P069"],"technical_key":"arxiv-2509.22622","catalog_id":"P069","library_id":"","topic_code":"","topic":"","task":"长视频与多镜头叙事","directions":[],"source_url":"https://arxiv.org/html/2509.22622","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":12.0,"wall_hours_max":12.0,"time_bound":"approximate","gpu_hours_min":768.0,"gpu_hours_max":768.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"LongLive 60 秒 streaming long tuning；3K iterations；总 batch 64","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2509.22622","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.1-T2V-1.3B","training_datasets":"长视频 rollout 微调数据；具体来源待核实","existing_method":"","training_signal":"在自身长视频 rollout 上微调，减轻误差累积","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"0edbba1ba832e3fd3e051778dde2bcca3b23a619e9e9ce4d6c0849d2087dc218","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0070","title":"Señorita-2M: A High-Quality Instruction-based Dataset for General Video Editing by Video Specialists","year":"2025","venue":"NeurIPS","arxiv_id":"2502.06734","aliases":["pool:P070"],"technical_key":"arxiv-2502.06734","catalog_id":"P070","library_id":"","topic_code":"","topic":"","task":"编辑与修改","directions":[],"source_url":"https://arxiv.org/pdf/2502.06734","status":"未定位可比微调配置","resource_note":"Señorita 的单 RTX 4090、1–2 分钟为推理，不计作训练配置。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CogVideoX 编辑专家","training_datasets":"Señorita-2M；由多个编辑专家生成配对样本","existing_method":"","training_signal":"专家合成编辑配对监督","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0b60757c021814fb2de4a5a173e700e68902cba390835ae84a78904dae102b7b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0071","title":"Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search","year":"2025","venue":"NeurIPS","arxiv_id":"2501.19252","aliases":["pool:P071"],"technical_key":"arxiv-2501.19252","catalog_id":"P071","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2501.19252","status":"已核对资源阶段","resource_note":"DLBS 是 training-free 的 inference-time search；搜索与 API 成本仍需单列。","stages":[{"role":"inf","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":40.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"DLBS training-free inference-time search；Latte/CogVideoX 配置；搜索/API 有额外成本","comparable":false,"is_gpu":true,"role_label":"推理／评测／rollout","source_url":"https://arxiv.org/html/2501.19252","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"inf","device":"H100","count_min":4.0,"count_max":4.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"DLBS Wan2.1 inference 配置；不能当作 fine-tuning 硬件","comparable":false,"is_gpu":true,"role_label":"推理／评测／rollout","source_url":"https://arxiv.org/html/2501.19252","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"多种视频扩散底座；含 Wan2.1","training_datasets":"无新增生成器训练；搜索使用 prompt 与 reward","existing_method":"","training_signal":"不适用：不更新生成器权重","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"75feb90d67e0a53151ae6d0520d0d0868f283481554be4205af6c95c98f8f4ba","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0072","title":"DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models","year":"2025","venue":"NeurIPS","arxiv_id":"2506.03517","aliases":["pool:P072"],"technical_key":"arxiv-2506.03517","catalog_id":"P072","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":[],"source_url":"https://arxiv.org/html/2506.03517","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":16.0,"wall_hours_max":16.0,"time_bound":"approximate","gpu_hours_min":1024.0,"gpu_hours_max":1024.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"DenseDPO 附录总述 training configuration 约 16h；SFT/DPO 各阶段时间未拆分，不声称是 DPO 独占耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2506.03517","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"DenseDPO；底座具体版本待核实","training_datasets":"分段视频偏好；含人类标注与 VLM 自动标注设置","existing_method":"","training_signal":"时间片段级 DPO；人类或 VLM 偏好标签","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"7ad02dbed29e1f8b557f6d114e0179d95f48d57a358d5e77ec057cfc866712d4","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0073","title":"WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors","year":"2026","venue":"NeurIPS","arxiv_id":"2605.10434","aliases":["pool:P073","reasoning:NeurIPS-2026-5021"],"technical_key":"arxiv-2605.10434","catalog_id":"P073","library_id":"","topic_code":"","topic":"","task":"不确定","directions":[],"source_url":"https://arxiv.org/html/2605.10434","status":"未定位可比微调配置","resource_note":"WorldReasonBench 的 H100 描述生成和 judge 评测，不属于 fine-tuning。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"多种被测生成器；Qwen3.5／GPT-5.4 评价设置","training_datasets":"WorldReasonBench 评测与人类偏好；生成训练设置另核","existing_method":"","training_signal":"主贡献为评测；不能把评测偏好自动当作训练集","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3b3620ee395b1f8c40b1547ae9562d73f12878ed3f82f8e6b6e6571ff0e440a3","reasoning_id":"NeurIPS-2026-5021","reasoning_role":"R3","reasoning_topic":"评测 / 数据 / 分析 / 安全"},{"id":"R0074","title":"PermaVid: Consistent Video Generation Across Edits via Disentangled Context Memory","year":2026,"venue":"NeurIPS 2026","arxiv_id":"2606.16449","aliases":["pool:P074","lineage:PERMA","recent:PERMA"],"technical_key":"arxiv-2606.16449","catalog_id":"P074","library_id":"","topic_code":"","topic":"","task":"编辑与修改","directions":["AIGC-B","XR-S"],"source_url":"https://arxiv.org/html/2606.16449v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"PermaVid stage 1 camera activation；SpatialVid，10K steps；stage 2 资源不能自动视为相同","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.16449v2","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.1-14B；VACE 初始化的冻结记忆分支","training_datasets":"可编辑场景／导航训练视频；完整来源待核实","existing_method":"","training_signal":"混合 RGB／depth 记忆条件下的视频生成监督","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"a3373c268233a075a0a7b35a49a21dc866abd496dea0b01f110dd1bf85b8001f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0075","title":"EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation","year":"2026","venue":"NeurIPS","arxiv_id":"2605.15199","aliases":["pool:P075"],"technical_key":"arxiv-2605.15199","catalog_id":"P075","library_id":"","topic_code":"","topic":"","task":"长视频与多镜头叙事","directions":[],"source_url":"https://arxiv.org/html/2605.15199","status":"未定位可比微调配置","resource_note":"EntityBench 的 L20 cluster 是评测设备，不能计为训练成本。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"EntityBench；EntityMem memory baseline","training_datasets":"无新增生成器训练的 EntityMem；EntityBench 为评测数据","existing_method":"","training_signal":"主贡献为评测；EntityMem 是推理时记忆基线","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5f86ba7e8b896c3630ff24b1c7247c17bed8654089ff95f2c1a44ad20a65abb7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0076","title":"DrivingGen: A Comprehensive Benchmark for Generative Video World Models in Autonomous Driving","year":"2026","venue":"ICLR","arxiv_id":"2601.01528","aliases":["pool:P076"],"technical_key":"arxiv-2601.01528","catalog_id":"P076","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.01528","status":"未定位可比微调配置","resource_note":"DrivingGen 的单 GPU 1–2 天是评测 400 个视频的成本，不是微调。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2326eb4191f9e8b6cf08f662e27d3da271c8a4708604887c00a4e61d7f0cbff3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0077","title":"RISE-Video: Can Video Generators Decode Implicit World Rules?","year":"2026","venue":"NeurIPS","arxiv_id":"2602.05986","aliases":["pool:P077","reasoning:NeurIPS-2026-1384"],"technical_key":"arxiv-2602.05986","catalog_id":"P077","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.05986","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"被测 11 模型：Hailuo2.3、Wan2.6、Sora2、Veo3.1、Kling2.6、Seedance1.5-Pro；Wan2.2-I2V-A14B/TI2V-5B、HunyuanVideo1.5-720P-I2V 及 cfg-distill、CogVideoX1.5-5B","training_datasets":"无新增生成模型训练；使用现成模型推理","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ecb7cc4029cedd24e6700808c610ca2bba0c04efa2f8b29b03f06e8bfbaa78e1","reasoning_id":"NeurIPS-2026-1384","reasoning_role":"R3","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0078","title":"FlashMotion: Few-Step Controllable Video Generation with Trajectory Guidance","year":"2026","venue":"CVPR","arxiv_id":"2603.12146","aliases":["pool:P078"],"technical_key":"arxiv-2603.12146","catalog_id":"P078","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.12146","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"FlashMotion stage 1 SlowAdapter 4.6K+5.4K steps；stage 2 FastGenerator 5.5K steps；耗时未知","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.12146","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"FlashMotion 最后 alignment 阶段 1K steps；不代表包含前两阶段的完整复现成本","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.12146","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"02a569b8eeb07c232be64dde21aa82e37451886e45abf09275a6bdfe7b04a6e1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0079","title":"GenHOI: Towards Object-Consistent Hand-Object Interaction with Temporally Balanced and Spatially Selective Object Injection","year":"2026","venue":"CVPR","arxiv_id":"2603.06048","aliases":["pool:P079"],"technical_key":"arxiv-2603.06048","catalog_id":"P079","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.06048","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":16.0,"count_max":16.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":1152.0,"gpu_hours_max":1152.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"GenHOI Wan-14B-I2V；约 19K videos；3 天","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.06048","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"0c9232e28f8c2fb4275bf1378e664ad57d41928370b8b74e1bb5f9e2be785e2e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0080","title":"Plenoptic Video Generation","year":"2026","venue":"CVPR","arxiv_id":"2601.05239","aliases":["pool:P080"],"technical_key":"arxiv-2601.05239","catalog_id":"P080","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.05239","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Plenoptic Video Generation：Cosmos-Predict2.5-2B；只更新 self-attention 与 camera encoder；93 frames","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2601.05239","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"46e9b9ed6ee3450f6ec4055516fa251f9d17a9764b80015530bc96938cad7819","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0081","title":"Ultra Flash: Scaling Real-Time Streaming Video Generation to High Resolutions","year":"2026","venue":"NeurIPS","arxiv_id":"2606.09150","aliases":["pool:P081"],"technical_key":"arxiv-2606.09150","catalog_id":"P081","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.09150","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Ultra Flash 分阶段训练；原文合计约 2,176 GPU-hours；多阶段 2.8 天顺序 / 2.5 天并行","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.09150","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H200","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":2176,"gpu_hours_max":2176,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"表 7 全部训练阶段合计","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.09150","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"7b809c7dbb7ed1b5412999c9239f4f387d5df4dfcd7c0b8225319f4c967a8324","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0082","title":"Activation Steering of Video Generation Models via Reduced-Order Linear Optimal Control","year":"2026","venue":"NeurIPS","arxiv_id":"2606.04775","aliases":["pool:P082"],"technical_key":"arxiv-2606.04775","catalog_id":"P082","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.04775","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b02d59eb89ccd6f9c0975c2cbd4abba2e138741db68058d5e8b8cb7312897e93","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0083","title":"Physics-Informed Video Generation via Mixture-of-Experts Latent Alignment","year":"2026","venue":"NeurIPS","arxiv_id":"2606.04737","aliases":["pool:P083"],"technical_key":"arxiv-2606.04737","catalog_id":"P083","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.04737","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":40.0,"wall_hours_max":40.0,"time_bound":"approximate","gpu_hours_min":320.0,"gpu_hours_max":320.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"PILA Wan2.1-1.3B adapter；WISA-80K；Wan2.2-14B 是直接 transfer，无额外 14B 训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.04737","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"2ffb1aa454c7511812fc023aadefdb260b7f0960c769d654d4da1126afe59fcc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0084","title":"RefDecoder: Enhancing Visual Generation with Conditional Video Decoding","year":"2026","venue":"NeurIPS","arxiv_id":"2605.15196","aliases":["pool:P084"],"technical_key":"arxiv-2605.15196","catalog_id":"P084","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.15196","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":120.0,"wall_hours_max":120.0,"time_bound":"approximate","gpu_hours_min":960.0,"gpu_hours_max":960.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"RefDecoder attention/reference encoder + pretrained VAE decoder；5 天","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.15196","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"321cbccd6609aac295caea98a36cbcdafee6d2dfe2ce394ebb9f99f076bf0642","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0085","title":"RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO","year":"2026","venue":"NeurIPS","arxiv_id":"2605.15190","aliases":["pool:P085"],"technical_key":"arxiv-2605.15190","catalog_id":"P085","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.15190","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"RAVEN 和 CM-GRPO 分别约 70 / 170 H200 GPU-hours；GPU 数未定位，不反推墙钟时长","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.15190","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H200","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":70,"gpu_hours_max":70,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"RAVEN","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.15190","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H200","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":170,"gpu_hours_max":170,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"CM-GRPO","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.15190","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"7148f91f5eeae4682569eca3b8f524bfe8edbfb3c6ff07d7ddea4d5d120ec0c1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0086","title":"EverAnimate: Minute-Scale Human Animation via Latent Flow Restoration","year":"2026","venue":"NeurIPS","arxiv_id":"2605.15042","aliases":["pool:P086"],"technical_key":"arxiv-2605.15042","catalog_id":"P086","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.15042","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"EverAnimate 两阶段 4K / 1K iterations；LoRA rank 128；型号和耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.15042","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"7c0ac2fd90b636d1a86bd95f259255e7c786f237e03e26e6920d0d32db31ee66","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0087","title":"PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation","year":"2026","venue":"NeurIPS","arxiv_id":"2605.14269","aliases":["pool:P087"],"technical_key":"arxiv-2605.14269","catalog_id":"P087","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.14269","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":60.9,"wall_hours_max":60.9,"time_bound":"approximate","gpu_hours_min":487.2,"gpu_hours_max":487.2,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"PhyMotion LoRA RL；330 steps；主实验附录 60.9h，正文约 2.5 天；拥挤设备对照约 66h 不混入","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.14269","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"da38a3ade528b6de26ceafb9d80a29dc441bd42c95907b75a1ace1a91abe69be","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0088","title":"CRePE: Curved Ray Expectation Positional Encoding for Unified-Camera-Controlled Video Generation","year":"2026","venue":"NeurIPS","arxiv_id":"2605.12938","aliases":["pool:P088"],"technical_key":"arxiv-2605.12938","catalog_id":"P088","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.12938","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"af969edc1ced81cf88c9e8e04c1ef7bca3de4e61d7b026848bdb28e27ba172ce","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0089","title":"From Articulated Kinematics to Routed Visual Control for Action-Conditioned Surgical Video Generation","year":"2026","venue":"NeurIPS","arxiv_id":"2605.08712","aliases":["pool:P089"],"technical_key":"arxiv-2605.08712","catalog_id":"P089","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.08712","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H20","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"手术视频模型 KVLR；batch 24；两阶段 fine-tuning；数量和耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.08712","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"fde9522fc6e149733bbe639f0a38f9acde4aade869208475f23c3f2b18868b41","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0090","title":"PhyCo: Learning Controllable Physical Priors for Generative Motion","year":"2026","venue":"CVPR","arxiv_id":"2604.28169","aliases":["pool:P090"],"technical_key":"arxiv-2604.28169","catalog_id":"P090","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2604.28169","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":45.0,"wall_hours_min":12.0,"wall_hours_max":12.0,"time_bound":"approximate","gpu_hours_min":48.0,"gpu_hours_max":48.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"PhyCo 每个属性 ControlNet 分支；10K steps；约半天；不是所有属性分支的总和","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2604.28169","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":115.0,"wall_hours_min":1.166667,"wall_hours_max":1.166667,"time_bound":"approximate","gpu_hours_min":9.3333,"gpu_hours_max":9.3333,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"PhyCo 每个分支额外 VLM-guided reward 优化；100 iterations；约 70 分钟","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2604.28169","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"c2f14539625a8fc303964288c9e97015061db37587325eee3c8ca56ca402b353","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0091","title":"SPIRAL: Self-Evolving Action-Conditioned Video Generation via Reflective Planning Agents","year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","arxiv_id":"2603.08403","aliases":["pool:P091","lineage:SPIRAL","reasoning:NeurIPS-2026-3987"],"technical_key":"arxiv-2603.08403","catalog_id":"P091","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":["XR-E"],"source_url":"https://arxiv.org/html/2603.08403v3","status":"未定位可比微调配置","resource_note":"SPIRAL 有 LoRA rank 16、学习率、epoch 等设置，但当前检索未定位明确 GPU 数量与训练耗时。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Spiral；Wan2.1／Wan2.2／CogVideoX／SVI；另有闭源工具适配","training_datasets":"训练数据沿用既有正文标注","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"acf8afae3269cb150d59da2560f0c3efd912bb33ff29b1d6cf47cf6a86fc45ce","reasoning_id":"NeurIPS-2026-3987","reasoning_role":"R1","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0092","title":"PISCO: Precise Video Instance Insertion with Sparse Control","year":"2026","venue":"NeurIPS","arxiv_id":"2602.08277","aliases":["pool:P092"],"technical_key":"arxiv-2602.08277","catalog_id":"P092","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.08277","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"PISCO 多阶段训练；14B 高/低噪声 denoiser 分别训练；数量和耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2602.08277","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"cf37bb1e69dfad290cff4c3ef66c997aea6baf7c4a77153aa5b67b15de8a1681","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0093","title":"Inference-time Physics Alignment of Video Generative Models with Latent World Models","year":"2026","venue":"CVPR","arxiv_id":"2601.10553","aliases":["pool:P093"],"technical_key":"arxiv-2601.10553","catalog_id":"P093","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.10553","status":"已核对资源阶段","resource_note":"WMReward 使用冻结的 latent world model 在 inference time 引导生成。","stages":[{"role":"inf","device":"H200","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"WMReward 的 vLDM 推理配置；基于 frozen VJEPA2 的 inference-time search/guidance","comparable":false,"is_gpu":true,"role_label":"推理／评测／rollout","source_url":"https://arxiv.org/html/2601.10553","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"inf","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"WMReward 的 MAGI-1 24B 推理配置；不是训练资源","comparable":false,"is_gpu":true,"role_label":"推理／评测／rollout","source_url":"https://arxiv.org/html/2601.10553","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f3d90ca550a93d77e2088c8157627b7c353a8b69cf2616f2dde6f2783646d632","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0094","title":"Mind the Generative Details: Direct Localized Detail Preference Optimization for Video Diffusion Models","year":"2026","venue":"CVPR","arxiv_id":"2601.04068","aliases":["pool:P094"],"technical_key":"arxiv-2601.04068","catalog_id":"P094","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.04068","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c6c2dedce986f97701ec73ee0b91f8e93daedcf5506929b7a9519ad6b23224cf","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0095","title":"Thinking with Frames: Generative Video Distortion Evaluation via Frame Reward Model","year":"2026","venue":"CVPR","arxiv_id":"2601.04033","aliases":["pool:P095"],"technical_key":"arxiv-2601.04033","catalog_id":"P095","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.04033","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"eeee2133e8460b2cb309e2e6ba56de74678c22d502da592465206523b4817ff9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0096","title":"ActWorld: From Explorable to Interactive World Model via Action-Aware Memory","year":"2026","venue":"NeurIPS","arxiv_id":"2606.17730","aliases":["pool:P096"],"technical_key":"arxiv-2606.17730","catalog_id":"P096","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.17730","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":48.0,"count_max":48.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ActWorld stage 1 和 stage 2 各 5K steps；stage 3 另有 ODE warm-up + GAN distillation，未将硬件外推至该阶段","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.17730","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":48.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"cb38dde28575b8de6406b65d2b92eaeeccc6bfc5ca4fa0743b79b9233c92f4b0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0097","title":"AnchorWorld: Embodied Egocentric World Simulation with View-based Evolution Customization","year":"2026","venue":"NeurIPS","arxiv_id":"2606.07326","aliases":["pool:P097"],"technical_key":"arxiv-2606.07326","catalog_id":"P097","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.07326","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":16.0,"count_max":16.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"AnchorWorld 四阶段 GPU-hours 分别 600 / 300 / 253 / 253；不是墙钟小时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.07326","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"GPU（型号未定位）","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":600,"gpu_hours_max":600,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"Stage I","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.07326","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"GPU（型号未定位）","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":300,"gpu_hours_max":300,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"Stage II","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.07326","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"GPU（型号未定位）","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":253,"gpu_hours_max":253,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"Stage III","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.07326","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"GPU（型号未定位）","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":253,"gpu_hours_max":253,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"Stage IV","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.07326","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"6dfdcf6486e3f9c67dbae82946d6a0065c11416f02d92cf467ba7d8a15cb9af3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0098","title":"SonoWorld: From One Image to a 3D Audio-Visual Scene","year":"2026","venue":"CVPR","arxiv_id":"2603.28757","aliases":["pool:P098"],"technical_key":"arxiv-2603.28757","catalog_id":"P098","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.28757","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"959f1bd4ce5ed2ce4e2411fd2f8841af988da0daca544195758b9bf6a545012c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0099","title":"PerpetualWonder: Long-Horizon Action-Conditioned 4D Scene Generation","year":"2026","venue":"CVPR","arxiv_id":"2602.04876","aliases":["pool:P099"],"technical_key":"arxiv-2602.04876","catalog_id":"P099","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.04876","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2270fe397a28c2da888e5863d690a97aa95db27f10064ecdd01d7c886d59d456","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0100","title":"VerseCrafter: Dynamic Realistic Video World Model with 4D Geometric Control","year":"2026","venue":"CVPR","arxiv_id":"2601.05138","aliases":["pool:P100"],"technical_key":"arxiv-2601.05138","catalog_id":"P100","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.05138","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":16.0,"count_max":16.0,"card_memory_gb":96.0,"peak_memory_gb":null,"wall_hours_min":380.0,"wall_hours_max":380.0,"time_bound":"approximate","gpu_hours_min":6080.0,"gpu_hours_max":6080.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"VerseCrafter 两阶段各 2,500 iterations；原文总墙钟约 380 小时；不能猜测 96GB GPU 的型号","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2601.05138","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"4434c761bfeaefd8c84f5f27d62856eb5d2ca6df69117055fe74769398995d6f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0101","title":"NeoVerse: Enhancing 4D World Model with in-the-wild Monocular Videos","year":"2026","venue":"CVPR","arxiv_id":"2601.00393","aliases":["pool:P101"],"technical_key":"arxiv-2601.00393","catalog_id":"P101","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.00393","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A800","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"NeoVerse 两阶段 150K / 50K iterations；81 frames，336×560","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2601.00393","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"e4d15628ed6421befac89642a0fae53f424352462d06e03b3bddc76dc85fbf72","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0102","title":"Solaris: Building a Multiplayer Video World Model in Minecraft","year":"2026","venue":"NeurIPS","arxiv_id":"2602.22208","aliases":["pool:P102"],"technical_key":"arxiv-2602.22208","catalog_id":"P102","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.22208","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"TPU v5p-64","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Solaris multiplayer / causal / Self Forcing 使用 v5p-64 配置；不把 TPU 拓扑名直接当 GPU 数；single-player 为 v5p-128","comparable":true,"is_gpu":false,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2602.22208","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9ea0da64112171c0c068a0e26e6f79ef861a82839f12b2cbf8706605573b0f14","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0103","title":"Fast-WAM: Do World Action Models Need Test-time Future Imagination?","year":"2026","venue":"NeurIPS","arxiv_id":"2603.16666","aliases":["pool:P103"],"technical_key":"arxiv-2603.16666","catalog_id":"P103","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.16666","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c74f72ab469ad1e0a2d1583bbecbd98bdf495d204bd33be147ac13d735a8279f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0104","title":"SwitchCraft: Training-Free Multi-Event Video Generation with Attention Controls","year":"2026","venue":"CVPR","arxiv_id":"2602.23956","aliases":["pool:P104"],"technical_key":"arxiv-2602.23956","catalog_id":"P104","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.23956","status":"已核对资源阶段","resource_note":"SwitchCraft 明确为 training-free 的 attention control。","stages":[{"role":"inf","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"SwitchCraft training-free attention control；832×480，81 frames；推理硬件","comparable":false,"is_gpu":true,"role_label":"推理／评测／rollout","source_url":"https://arxiv.org/html/2602.23956","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c37a2ea6691ed0d6ad48d1d0597edf4547bdfea78500f790da8ed9ffa1413e09","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0105","title":"NOVA: Sparse Control, Dense Synthesis for Pair-Free Video Editing","year":"2026","venue":"CVPR","arxiv_id":"2603.02802","aliases":["pool:P105"],"technical_key":"arxiv-2603.02802","catalog_id":"P105","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.02802","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"NOVA 训练只报告 aggregate GPU memory 320GB；不是单卡显存；5K Pexels clips，8K steps，冻结 WAN VACE","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.02802","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"8e13c1dc5e4349f5f2eadd64ad4ce693b5b2b74f5821f5cb0a23922dc0806920","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0106","title":"EditCtrl: Disentangled Local and Global Control for Real-Time Generative Video Editing","year":"2026","venue":"CVPR","arxiv_id":"2602.15031","aliases":["pool:P106"],"technical_key":"arxiv-2602.15031","catalog_id":"P106","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.15031","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":24.0,"wall_hours_max":24.0,"time_bound":"approximate","gpu_hours_min":192.0,"gpu_hours_max":192.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"EditCtrl 约 1 天；49 frames，480×720；batch 8 + gradient accumulation 8","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2602.15031","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"47b8fed6b05ddbcc70b1fd22834ce20cc410656919248398dcc92d289aa6ebfb","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0107","title":"Chain of Event-Centric Causal Thought for Physically Plausible Video Generation","year":"2026","venue":"CVPR","arxiv_id":"2603.09094","aliases":["pool:P107","reasoning:CVPR-2026-3739"],"technical_key":"arxiv-2603.09094","catalog_id":"P107","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.09094","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"CogVideoX-5B","training_datasets":"主文方法是基于现有模型的推理流程，未报告新增训练集","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"56e0700bd5a6604b4836c8f78b49a880a879c9460634f7792a260de3a4824df4","reasoning_id":"CVPR-2026-3739","reasoning_role":"R1","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0108","title":"Video Generation Models: A Survey of Post-Training and Alignment","year":"2026","venue":"TMLR（作者自报）","arxiv_id":"2610.00812","aliases":["pool:P108"],"technical_key":"arxiv-2610.00812","catalog_id":"P108","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2610.00812","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c98b06de6758414b3eeb72fe21d875bed42d7d82f9e5a105048528d903810fe6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0109","title":"MuSS: A Large-Scale Dataset and Cinematic Narrative Benchmark for Multi-Shot Subject-to-Video Generation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2604.23789","aliases":["pool:P109"],"technical_key":"arxiv-2604.23789","catalog_id":"P109","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2604.23789","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H20","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":84.0,"wall_hours_max":84.0,"time_bound":"approximate","gpu_hours_min":2688.0,"gpu_hours_max":2688.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"MuSS 的训练 baseline；50K steps；832×480；161 frames；3.5 天","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2604.23789","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"31c5395f06efc0eed554a0ec0288b4838eda9a82aeb31d1730df8b4fd79beeda","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0110","title":"MTAVG-Bench: A Diagnostic Benchmark for Multi-Talker Dialogue-Centric Audio-Video Generation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2602.00607","aliases":["pool:P110"],"technical_key":"arxiv-2602.00607","catalog_id":"P110","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.00607","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5624057596a15bd7c6a8c1f1cdb8487b0e3f0d57ad7a86b1325818ceee0d0d86","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0111","title":"MSVBench: Towards Human-Level Evaluation of Multi-Shot Video Generation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2602.23969","aliases":["pool:P111"],"technical_key":"arxiv-2602.23969","catalog_id":"P111","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.23969","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c6475176a8f44de6bdafa7b11810a31377d94e005dac64582ab421ac0756c2e6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0112","title":"MechVerse: Evaluating Physical Motion Consistency in Video Generation Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2605.14843","aliases":["pool:P112"],"technical_key":"arxiv-2605.14843","catalog_id":"P112","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.14843","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GH200","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"MechVerse 两个生成模型 LoRA rank 32；2 epochs；batch 1；耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.14843","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"b3372c0db80ccb1bf077aba4d2646f93669e0bfc347fb404d572263ae127ba70","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0113","title":"BRITE: A Benchmark for Reliable and Interpretable T2V Evaluation on Implausible Scenarios","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2605.00873","aliases":["pool:P113"],"technical_key":"arxiv-2605.00873","catalog_id":"P113","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.00873","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f1554ed9f46a771178dca8691b52b99698d2d8304171b91b2ab69f3611e4bbf3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0114","title":"VIGOR: VIdeo Geometry-Oriented Reward for Temporal Generative Alignment","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.16271","aliases":["pool:P114"],"technical_key":"arxiv-2603.16271","catalog_id":"P114","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.16271","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"876c38ad975a83f1118688e2b36354495bf0a582696d8583bc3f6e23f96c9c41","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0115","title":"World-R1: Reinforcing 3D Constraints for Text-to-Video Generation","year":"2026","venue":"ICML（作者自报）","arxiv_id":"2604.24764","aliases":["pool:P115"],"technical_key":"arxiv-2604.24764","catalog_id":"P115","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2604.24764","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":48.0,"count_max":48.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"World-R1-Small Wan2.1-1.3B；Flow-GRPO-Fast；832×480","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2604.24764","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H200","count_min":96.0,"count_max":96.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"World-R1-Large Wan2.1-14B；与 Small 是不同配置，不相加","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2604.24764","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"large","max_reported_ft_gpus":96.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"68f0cf842944f993acfa165042a61883f6193a58eb05d17c6be09a9b6ab7fd4f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0116","title":"TAGRPO: Boosting GRPO on Image-to-Video Generation with Direct Trajectory Alignment","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2601.05729","aliases":["pool:P116"],"technical_key":"arxiv-2601.05729","catalog_id":"P116","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.05729","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"385ab5408d50b091beb02e0ced98e7cf3d237386dca02a8ea41678ada0ac33f9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0117","title":"LIFT: Layout-In-Future Video Generation under Large Viewpoint Change via On-Policy Self-Distillation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2609.38146","aliases":["pool:P117"],"technical_key":"arxiv-2609.38146","catalog_id":"P117","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2609.38146","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"LIFT 三阶段 8K / 4K / 500 steps；352×640，81 frames；总耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2609.38146","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"678c0679dfa218cae5ae986979e752822365d4068a355d2ee1db966765db80e3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0118","title":"Delta Forcing: Trust Region Steering for Interactive Autoregressive Video Generation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2605.14382","aliases":["pool:P118"],"technical_key":"arxiv-2605.14382","catalog_id":"P118","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.14382","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Delta Forcing + streaming long tuning；阶段 2 额外 3K steps；GPU 数量未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.14382","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"3d389b148e71f195fd0e22766d3f825d82485c8afa990099a53e75f40d7128c1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0119","title":"CreFlow: Corrective Reflow for Sparse-Reward Embodied Video Diffusion RL","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2605.14274","aliases":["pool:P119"],"technical_key":"arxiv-2605.14274","catalog_id":"P119","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.14274","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0420d64f0b25a14169ff4c9519b0bbac20606c1cf0e6fe073f2034ae0c3b0f1e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0120","title":"TIE: Time Interval Encoding for Video Generation over Events","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2605.10543","aliases":["pool:P120"],"technical_key":"arxiv-2605.10543","catalog_id":"P120","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.10543","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f949c5f1dc47acaa646354d794c248ebc06215a5685fec6a93dbf448bd1c04ec","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0121","title":"Diffusion-APO: Trajectory-Aware Direct Preference Alignment for Video Diffusion Transformers","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2605.07503","aliases":["pool:P121"],"technical_key":"arxiv-2605.07503","catalog_id":"P121","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.07503","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":192.0,"count_max":192.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Diffusion-APO LoRA preference alignment；五阶段 pipeline 约 64,000 GPU-hours；该合计不能视为单独 LoRA 阶段","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.07503","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H200","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":64000,"gpu_hours_max":64000,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"五阶段 pipeline 合计；不是单独 alignment 阶段","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.07503","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"large","max_reported_ft_gpus":192.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"fe5ce723133d9c45c4fb9ecf5aaf3365ff0220df65e7d17acf022a10d748817a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0122","title":"OP-GRPO: Efficient Off-Policy GRPO for Flow-Matching Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2604.04142","aliases":["pool:P122"],"technical_key":"arxiv-2604.04142","catalog_id":"P122","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2604.04142","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"84c05ba4d69606385a522c79fb97f33b1cbec516919251a0e1a4739cc9c5aae4","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0123","title":"AesRM: Improving Video Aesthetics with Expert-Level Feedback","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2604.28078","aliases":["pool:P123"],"technical_key":"arxiv-2604.28078","catalog_id":"P123","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2604.28078","status":"已核对资源阶段","resource_note":"","stages":[{"role":"aux","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":5.0,"wall_hours_max":5.0,"time_bound":"approximate","gpu_hours_min":40.0,"gpu_hours_max":40.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"AesRM stage 1 reward model；3 epochs；生成器训练另计","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2604.28078","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"A100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":2304.0,"gpu_hours_max":2304.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"AesRM-CoT GRPO reward model；5 epochs；Gemini 合成 CoT 的 API 成本另计","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2604.28078","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a7fa5c7ce07eeb5bfae6c201ceab823be96f2d2228cb92bfa2629f41cdc4c974","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0124","title":"CoMoGen: COntrollable MOtion Dynamics and Interactions with Mask-Guided Video GENeration","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2605.22996","aliases":["pool:P124"],"technical_key":"arxiv-2605.22996","catalog_id":"P124","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.22996","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"CoMoGen：MaskAdapter 1 epoch + motion LoRA/adapter 2 epochs；batch 2；单卡 H100 不代表消费卡可复现","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.22996","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"bf902c33797a290c815d5eb581dd382ac644b716ef83d95b1372dcfb5a314194","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0125","title":"CogOmniControl: Reasoning-Driven Controllable Video Generation via Creative Intent Cognition","year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","arxiv_id":"2605.19995","aliases":["pool:P125","lineage:COG"],"technical_key":"arxiv-2605.19995","catalog_id":"P125","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":["AIGC-D"],"source_url":"https://arxiv.org/html/2605.19995v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H20","count_min":32.0,"count_max":32.0,"card_memory_gb":96.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"CogOmniDiT SFT 三阶段；LoRA rank 256；10K / 500 / 10K steps","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.19995v1","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H20","count_min":16.0,"count_max":16.0,"card_memory_gb":96.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"CogOmniDiT RFT；200 steps；不与 VLM 训练混记","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.19995v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CogVLM＋视频 DiT；具体 checkpoint 待核","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"a19ad1da896cd101d327db041cc46e8049284b2bbc25ae15b212726f7d4acc98","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0126","title":"MMControl: Unified Multi-Modal Control for Joint Audio-Video Generation","year":"2026","venue":"ECCV（作者自报）","arxiv_id":"2604.19679","aliases":["pool:P126"],"technical_key":"arxiv-2604.19679","catalog_id":"P126","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2604.19679","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":12.0,"wall_hours_max":12.0,"time_bound":"approximate","gpu_hours_min":48.0,"gpu_hours_max":48.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"MMControl LTX-2 19B bypass；7,200 steps；每 GPU batch 2、accumulation 2","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2604.19679","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"289b835e5ba2c161bfc156467f3902a5623c99d725675ed967205dec0f4f57e2","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0127","title":"Control-DINO: Feature Space Conditioning for Controllable Image-to-Video Diffusion","year":"2026","venue":"ECCV（作者自报）","arxiv_id":"2604.01761","aliases":["pool:P127"],"technical_key":"arxiv-2604.01761","catalog_id":"P127","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2604.01761","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3c7ab35a01456a3013453c93cfde65c93991174ce76e62a948fd861561ce5ba1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0128","title":"TokenDial: Continuous Attribute Control for Text-to-Video Generation in Visual Dial Space","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.27520","aliases":["pool:P128"],"technical_key":"arxiv-2603.27520","catalog_id":"P128","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/pdf/2603.27520","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":1.0,"count_max":1.0,"card_memory_gb":80.0,"peak_memory_gb":58.0,"wall_hours_min":0.2,"wall_hours_max":0.2,"time_bound":"approximate","gpu_hours_min":0.2,"gpu_hours_max":0.2,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"TokenDial Wan2.1-1.3B 每个 visual dial 优化；12 分钟；属于参数化控制向量优化，不是 full backbone fine-tuning","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2603.27520","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":2.0,"count_max":2.0,"card_memory_gb":80.0,"peak_memory_gb":44.0,"wall_hours_min":0.3,"wall_hours_max":0.3,"time_bound":"approximate","gpu_hours_min":0.6,"gpu_hours_max":0.6,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"TokenDial Wan2.1-14B 每个 dial；18 分钟；两卡 peak 分别 44 / 42GB；不能代表全模型训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2603.27520","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":2.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"0f2ef35c2d1d85534585dc2f9492701a462fe30506849dfdf40ab89bd3f72245","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0129","title":"VGGRPO: Towards World-Consistent Video Generation with 4D Latent Reward","year":"2026","venue":"ECCV（作者自报）","arxiv_id":"2603.26599","aliases":["pool:P129"],"technical_key":"arxiv-2603.26599","catalog_id":"P129","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.26599","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VGGRPO LoRA rank 32；总训练约 1,536 GPU-hours；reward 计算的单步显存不能当成完整训练峰值","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.26599","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"GPU（型号未定位）","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":1536,"gpu_hours_max":1536,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"VGGRPO 训练合计","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.26599","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"cedd346d1e1b94fe949cee2aa98244263f0c3ca9f782abba8f936e3819cad7be","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0130","title":"DiReCT: Disentangled Regularization of Contrastive Trajectories for Physics-Refined Video Generation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.25931","aliases":["pool:P130"],"technical_key":"arxiv-2603.25931","catalog_id":"P130","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.25931","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"eb5399fc906566c54b40a0b9e8008a533a17578835c455b95968b076aaf1eeec","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0131","title":"Helios: Real Real-Time Long Video Generation Model","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.04379","aliases":["pool:P131"],"technical_key":"arxiv-2603.04379","catalog_id":"P131","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.04379","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Helios stage 1 / 2 LoRA 配置表均为 64 H100；各阶段 5.5K / 7.5K / 16K / 20K steps；单 H100 19.5 FPS 是推理","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.04379","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"53e688e381acf3eff8fbc01496506538687815141675ce13e6290717bdd9de23","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0132","title":"AR-CoPO: Align Autoregressive Video Generation with Contrastive Policy Optimization","year":"2026","venue":"ECCV（作者自报）","arxiv_id":"2603.17461","aliases":["pool:P132"],"technical_key":"arxiv-2603.17461","catalog_id":"P132","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.17461","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":24.0,"count_max":24.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"AR-CoPO LoRA rank 64；group 12；型号与耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.17461","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":24.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"e4402fe549e2b8883667ddf605a7c696a93472a1768ce616883a021654479c74","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0133","title":"DreamVideo-Omni: Omni-Motion Controlled Multi-Subject Video Customization with Latent Identity Reinforcement Learning","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.12257","aliases":["pool:P133"],"technical_key":"arxiv-2603.12257","catalog_id":"P133","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.12257","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"DreamVideo-Omni 第一阶段 omni-motion / identity SFT；40K iterations；RL 后续成本另缺","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.12257","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"67289c4d0eec15edbe4e8549e42612b8b03862d9bac586ec133a9fa16f591017","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0134","title":"EVA: Aligning Video World Models with Executable Robot Actions via Inverse Dynamics Rewards","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.17808","aliases":["pool:P134"],"technical_key":"arxiv-2603.17808","catalog_id":"P134","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.17808","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A800","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"EVA video generator SFT 与 LoRA rank-32 GRPO；batch 32；49 frames；未定位总耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.17808","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"9b01b79f1968b03f3595d126b23b4dace18dce0045a443c305ad7ce76f469318","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0135","title":"FlowPortrait: Reinforcement Learning for Audio-Driven Portrait Video Generation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.00159","aliases":["pool:P135"],"technical_key":"arxiv-2603.00159","catalog_id":"P135","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.00159","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ae889da6a3f288e342bbc74ba7fbb01eae4f86715a6d65b25dac239fda3b69bb","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0136","title":"InfinityStory: Unlimited Video Generation with World Consistency and Character-Aware Shot Transitions","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.03646","aliases":["pool:P136"],"technical_key":"arxiv-2603.03646","catalog_id":"P136","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.03646","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d5f7eff608be5a4565ab5da0aa1dd14ca8c38d6f028418129e6f8125b33cf8d2","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0137","title":"Making Video Models Adhere to User Intent with Minor Adjustments","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.19672","aliases":["pool:P137"],"technical_key":"arxiv-2603.19672","catalog_id":"P137","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.19672","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"20d0cbf32d17c8656afab196b2cb1f8185abc065fbaa0fce24a52e232f863fc5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0138","title":"ShotVerse: Advancing Cinematic Camera Control for Text-Driven Multi-Shot Video Creation","year":"2026","venue":"SIGGRAPH Asia（作者自报）","arxiv_id":"2603.11421","aliases":["pool:P138"],"technical_key":"arxiv-2603.11421","catalog_id":"P138","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.11421","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H20","count_min":96.0,"count_max":96.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ShotVerse camera encoder + LoRA；FSDP；未定位总耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.11421","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"large","max_reported_ft_gpus":96.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"542a5df9725a6bf87e19f7451de14aab943328800577bc5d7d4300656b49abf6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0139","title":"ViFeEdit: A Video-Free Tuner of Your Video Diffusion Transformer","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.15478","aliases":["pool:P139"],"technical_key":"arxiv-2603.15478","catalog_id":"P139","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.15478","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 6000 Ada","count_min":3.0,"count_max":3.0,"card_memory_gb":null,"peak_memory_gb":18.0,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ViFeEdit LoRA rank 32；peak 18 GiB；每 epoch 风格任务 5 分钟、其他编辑 9 分钟；不是完整训练总时长","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.15478","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":3.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"bded5e94a8e2a6cb024957a3594b21429d43984d6041e4ee34b1e4d992fb6d6e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0140","title":"Tele-Omni: a Unified Multimodal Framework for Video Generation and Editing","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2602.09609","aliases":["pool:P140"],"technical_key":"arxiv-2602.09609","catalog_id":"P140","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.09609","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d6f747861cc98517bb1aa28e55db799a5d9a173fa103d9e5662451c5bb99d723","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0141","title":"Omni-Video 2: Scaling MLLM-Conditioned Diffusion for Unified Video Generation and Editing","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2602.08820","aliases":["pool:P141"],"technical_key":"arxiv-2602.08820","catalog_id":"P141","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.08820","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"Alibaba PPU","count_min":1600.0,"count_max":1600.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Omni-Video 2 两个 14B DiT + adapter；大规模训练；原文 PPU 设备不与 NVIDIA GPU-hours 等价","comparable":true,"is_gpu":false,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2602.08820","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d104a4323767c99786f44efb63d67bb4e94b38aff638de4c3f9bd2cb46868b51","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0142","title":"PISCES: Annotation-free Text-to-Video Post-Training via Optimal Transport-Aligned Rewards","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2602.01624","aliases":["pool:P142"],"technical_key":"arxiv-2602.01624","catalog_id":"P142","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.01624","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":30.0,"wall_hours_max":30.0,"time_bound":"approximate","gpu_hours_min":240.0,"gpu_hours_max":240.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"PISCES direct backprop 实际训练约 30h；含推理可视化的周转时间约 2 天","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2602.01624","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":78.0,"wall_hours_max":78.0,"time_bound":"approximate","gpu_hours_min":624.0,"gpu_hours_max":624.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"PISCES GRPO 实际训练约 78h；含推理可视化约 4 天；不同方法不相加","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2602.01624","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"59ed49a9f774b4bf6c95d7d4c8d3ac412492efe2d9ba5b1b352b4c8295de94af","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0143","title":"Reward-Forcing: Autoregressive Video Generation with Reward Feedback","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2601.16933","aliases":["pool:P143"],"technical_key":"arxiv-2601.16933","catalog_id":"P143","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.16933","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Reward-Forcing 训练基础设施；ODE 初始化、DMD 与 reward-only 配置耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2601.16933","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"7007878d17d32ad186b4da82c2f2788258a1db2a3fbbf4aa5a8c0468bdf93f1b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0144","title":"PhysRVG: Physics-Aware Unified Reinforcement Learning for Video Generative Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2601.11087","aliases":["pool:P144"],"technical_key":"arxiv-2601.11087","catalog_id":"P144","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.11087","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H20","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"PhysRVG 4 nodes × 8 H20；stage 1 16K steps / stage 2 250 steps","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2601.11087","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"0e5953c53861941853282d5c8d6c4821f5bd56d479fb5aa72494fbe0e233edc0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0145","title":"Focal Guidance: Unlocking Controllability from Semantic-Weak Layers in Video Diffusion Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2601.07287","aliases":["pool:P145"],"technical_key":"arxiv-2601.07287","catalog_id":"P145","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.07287","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ee16ea0ac46b6dc7d96a84451ab5193376b821caa242ed997a7dff41565228ae","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0146","title":"Slot-ID: Identity-Preserving Video Generation from Reference Videos via Slot-Based Temporal Identity Encoding","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2601.01352","aliases":["pool:P146"],"technical_key":"arxiv-2601.01352","catalog_id":"P146","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.01352","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9add9aaab4f8bec076dd8be661d9c86673e29e93268617155c925e64a53ac486","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0147","title":"PROBE: Diagnosing Residual Concept Capacity in Erased Text-to-Video Diffusion Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.21547","aliases":["pool:P147"],"technical_key":"arxiv-2603.21547","catalog_id":"P147","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.21547","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":3.0,"wall_hours_max":10.0,"time_bound":"approximate","gpu_hours_min":3.0,"gpu_hours_max":10.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"PROBE 优化 pseudo-tokens；CogVideoX-2B 各概念 3 / 10h；5B 和 Wan2.2-5B 各类约 7h；不是全模型 fine-tuning","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.21547","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"88005061c2df2289f56dad65a63c88fa91f70b3b3162d716e399f0eb95a37a47","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0148","title":"DirectorBench: Diagnosing Long-Form Video Generation with Personalized Multi-Agent Evaluation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2605.30090","aliases":["pool:P148"],"technical_key":"arxiv-2605.30090","catalog_id":"P148","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.30090","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a9dd733131568cd4e17bd5452dbc6e42d37f0a3cf5fcc8bfbcaae80feeed3336","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0149","title":"EvalVerse: Pipeline-Aware and Expert-Calibrated Benchmarking for Professional Cinematic Video Generation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2605.23271","aliases":["pool:P149"],"technical_key":"arxiv-2605.23271","catalog_id":"P149","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.23271","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"55eeb16161a2afe1173016f6a2442c155a31b62730d695dd07252ee7b4502743","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0150","title":"Stream-R1: Reliability-Perplexity Aware Reward Distillation for Streaming Video Generation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2605.03849","aliases":["pool:P150"],"technical_key":"arxiv-2605.03849","catalog_id":"P150","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.03849","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Stream-R1；1K optimizer steps；per-GPU batch 1，accumulation 8","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.03849","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"77bd6e3c73e1b622b231288faa76a63e4121c5fcd0b5b8966726390eb526a498","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0151","title":"Not all tokens contribute equally to diffusion learning","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2604.07026","aliases":["pool:P151"],"technical_key":"arxiv-2604.07026","catalog_id":"P151","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2604.07026","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H20","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Token 贡献加权 diffusion learning；30K iterations；推理另为 8 H20","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2604.07026","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"3a0c045805dfb13c160ee8c4c28bf9290750edd4c523ebec249b1ed5e20f9b16","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0152","title":"Towards Interactive Video World Modeling: Frontiers, Challenges, Benchmarks, and Future Trends","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2606.01164","aliases":["pool:P152"],"technical_key":"arxiv-2606.01164","catalog_id":"P152","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.01164","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1b04048ff0e0274bee91cd1876a213fef2fcf2e188bdccc9a6234ce94a9a8798","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0153","title":"HelixWorld: A Real-time Interactive Audio-Visual World Model","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2609.38123","aliases":["pool:P153"],"technical_key":"arxiv-2609.38123","catalog_id":"P153","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2609.38123","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":48.0,"count_max":48.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"HelixWorld training hardware：6 nodes × 8 H200；batch 48；单 H800 为 inference","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2609.38123","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":48.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"e7639bc0f724aaa11aeb8ccda4d63f1c95053933679abfcfca63eaa09d9344d3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0154","title":"WorldPlay2: Extending Real-Time Interactive World Models in Control and Horizon","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2609.35560","aliases":["pool:P154"],"technical_key":"arxiv-2609.35560","catalog_id":"P154","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2609.35560","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"392c8b68f81ec204076e8b63650fe16223d5fee265a9db7165c3b9f05aedf576","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0155","title":"Geometry as Address: Routing Attention to Visual Memory for Long-Horizon Camera-Controlled Video Generation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2609.34722","aliases":["pool:P155"],"technical_key":"arxiv-2609.34722","catalog_id":"P155","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2609.34722","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"GEAR；10K iterations；DL3DV-10K；型号和耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2609.34722","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"39b95e6ffb9541f8bfa1565f82b42513b99e7deef56c466d520fcf6282d5c4f3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0156","title":"GameDirector: Decoupling Gameplay Logic from Rendering for Player-Configurable Game World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2609.25652","aliases":["pool:P156"],"technical_key":"arxiv-2609.25652","catalog_id":"P156","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2609.25652","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"GameDirector DiT adaptation；60K steps；三个游戏共用；distillation 与辅助网络成本另缺","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2609.25652","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"daa2a8230ac01b991285ba0242f194a9599a3c10423fc91650f6a9def14e46ab","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0157","title":"ConsistWorld: Evidence Routing for Consistent Multi-Agent World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2609.22641","aliases":["pool:P157"],"technical_key":"arxiv-2609.22641","catalog_id":"P157","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2609.22641","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"6b63f7c56c7f4f275091251308bb94f4e8bc555ad03bbe869908bd2e0dff34c8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0158","title":"Programmable World Model","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2609.10540","aliases":["pool:P158"],"technical_key":"arxiv-2609.10540","catalog_id":"P158","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2609.10540","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4ba5deacad3c1ccf42841d25ab7f90b2c8167491563eabafeaca083fcad748b6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0159","title":"FlashRender: Few-Step Generative Rendering via Camera-Controlled Video MeanFlow","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2609.03563","aliases":["pool:P159"],"technical_key":"arxiv-2609.03563","catalog_id":"P159","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2609.03563","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX PRO 6000","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":336.0,"wall_hours_max":336.0,"time_bound":"approximate","gpu_hours_min":2688.0,"gpu_hours_max":2688.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"FlashRender 全部三个训练阶段约两周；81 frames，480×832；不等同于推理 B200 配置","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2609.03563","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"68590baaecb8b21dd38533c2ef471049883f17f266a861e0f9fb643834d46633","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0160","title":"Matrix-Game 3.5: Enhancing Real-Time Streaming Interactive World Models with Patch Memory","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2608.29910","aliases":["pool:P160"],"technical_key":"arxiv-2608.29910","catalog_id":"P160","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2608.29910","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e349f965ef4214abba6d5f57936f980443163b29c36f648bc89a43e9a5428c5e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0161","title":"VectorWorld: Efficient Streaming World Model via Diffusion Flow on Vector Graphs","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.17652","aliases":["pool:P161"],"technical_key":"arxiv-2603.17652","catalog_id":"P161","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.17652","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H20","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VectorWorld latent generator 与 autoencoder；ΔSim 另用 4 H20；非预训练视频底座 fine-tuning","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2603.17652","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"97e106eff60fb0abbfec52281ed669662ef9dece78db34da02b41397c2465769","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0162","title":"Holo-World: Unified Camera, Object and Weather Control for Video World Model","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2606.20083","aliases":["pool:P162"],"technical_key":"arxiv-2606.20083","catalog_id":"P162","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.20083","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"82c53a779b83d9c1e42c9e381a725036f0537ab5a98fe8237be5e8ccbb106026","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0163","title":"MoVerse: Real-Time Video World Modeling with Panoramic Gaussian Scaffold","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2606.13376","aliases":["pool:P163"],"technical_key":"arxiv-2606.13376","catalog_id":"P163","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.13376","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a890ef958f26335087f0238e7647d02f40795604a0104689fbede4742482d35c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0164","title":"BiWM: Advancing Open-Source Interactive Video World Models with Bidirectional Autoregression","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2606.10135","aliases":["pool:P164"],"technical_key":"arxiv-2606.10135","catalog_id":"P164","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.10135","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"BiWM camera control 约 100 steps + DMD 约 200 steps；原文仅称数小时，无可计算准确耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.10135","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"70890b6051cbe2c12f032ba97530f96b6971bf3229b8a55184454be464b6e06b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0165","title":"Prisma-World: Camera-Controllable Multi-Agent Video World Model","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2606.09507","aliases":["pool:P165"],"technical_key":"arxiv-2606.09507","catalog_id":"P165","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.09507","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"98cd40eb05aee1470761a16081d2167ac4d8186d6e014d99c1cdb305ca7215ef","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0166","title":"MosaicMem: Hybrid Spatial Memory for Controllable Video World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.17117","aliases":["pool:P166"],"technical_key":"arxiv-2603.17117","catalog_id":"P166","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.17117","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"MosaicMem Wan2.2-5B；250K steps；effective batch 64；未定位总时长","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.17117","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"45e2e71be09322ef90de41200651766457eceb64f955bcd8aa70c97a8a946e43","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0167","title":"LiveWorld: Simulating Out-of-Sight Dynamics in Generative Video World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.07145","aliases":["pool:P167"],"technical_key":"arxiv-2603.07145","catalog_id":"P167","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.07145","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"LiveWorld state adapter 10K steps + rank-64 LoRA 5K steps；batch 16","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.07145","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"f18a4d83eb3c4c0707f622416e3fa4becbfda1bbca726b3e833e6713ff9d84a9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0168","title":"RealWonder: Real-Time Physical Action-Conditioned Video Generation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.05449","aliases":["pool:P168","library:VG10"],"technical_key":"arxiv-2603.05449","catalog_id":"P168","library_id":"VG10","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"动作条件的预测与交互世界模型","directions":["XR-C"],"source_url":"https://arxiv.org/html/2603.05449","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"RealWonder flow-conditioned teacher + causal student；原文总训练 128 A100 GPU-days = 3,072 GPU-hours","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.05449","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":3072,"gpu_hours_max":3072,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"128 A100 GPU-days 换算为 GPU-hours","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.05449","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"VideoXFun Wan2.1-1.3B-InP；Genesis；Franka 模型","training_datasets":"模拟交互训练数据，完整数据源待核","existing_method":"从图像建立三维场景，以物理仿真将力或动作转为光流／RGB 条件，再由视频生成模型实现视觉输出。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"d31a81c331ab94a17deda7cc3c02b86cc240605ba06503030420c645a61d3cf1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0169","title":"Olaf-World: Orienting Latent Actions for Video World Modeling","year":"2026","venue":"ICML（作者自报）","arxiv_id":"2602.10104","aliases":["pool:P169","library:P081"],"technical_key":"arxiv-2602.10104","catalog_id":"P169","library_id":"P081","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/abs/2602.10104v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"aux","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":108.0,"wall_hours_max":108.0,"time_bound":"approximate","gpu_hours_min":864.0,"gpu_hours_max":864.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Olaf-World latent-action model；100 epochs，约 146K steps；4.5 天；视频 world model 训练另需核对","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/abs/2602.10104v2","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.5（同文多阶段请结合上方范围）"},{"role":"ft","device":"H200","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Olaf-World video generator 从 SkyReels 1.3B 适配；10K steps，每卡 batch 4；下游 LoRA rank 16，耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2602.10104v2","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.5（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"β-VAE逆/前向动力学学习潜动作；累计动作表征对齐冻结自监督视频编码器特征差；训练潜动作条件视频生成并适配新控制接口。","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"25e6b028f1a209d5eb964d665109e6d248a5018b5d4b317c8a78e54e1f96c97b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0170","title":"WorldCompass: Reinforcement Learning for Long-Horizon World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2602.09022","aliases":["pool:P170"],"technical_key":"arxiv-2602.09022","catalog_id":"P170","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.09022","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H20","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":4608.0,"gpu_hours_max":4608.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"WorldCompass long-horizon RL；3 天","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2602.09022","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"54c469d2aa409869edc754157552d8d6929349a39617af1479e55bfe5440d76f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0171","title":"MIND: Benchmarking Memory Consistency and Action Control in World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2602.08025","aliases":["pool:P171"],"technical_key":"arxiv-2602.08025","catalog_id":"P171","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.08025","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"MIND 的方法 baseline 蒸馏；ODE 3K steps + DMD 2K steps；benchmark 也有训练实验","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2602.08025","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"dfa2b173092bb134243625048c51efa144db394de0491db3ed570b94fb2fd16a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0172","title":"Infinite-World: Scaling Interactive World Models to 1000-Frame Horizons via Pose-Free Hierarchical Memory","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2602.02393","aliases":["pool:P172"],"technical_key":"arxiv-2602.02393","catalog_id":"P172","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.02393","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H800","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Infinite-World 各实验配置；30h 第一视角数据 + 30min revisit fine-tuning 数据；小时数是数据时长，不是训练耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2602.02393","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"e7eee08fb48a3a93141dd08ec9ea58a3953a54c816477a92bc4574c2b1c9628a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0173","title":"Game2World Engine: Unlocking In-the-Wild Gameplay Videos for World Model Training","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2608.24680","aliases":["pool:P173"],"technical_key":"arxiv-2608.24680","catalog_id":"P173","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2608.24680","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Game2World 正文 GameCleaner：LoRA rank 64，3K steps，batch 32","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2608.24680","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Game2World 附录另一训练配置：1,400 steps，720p；与正文 H100/3K 的阶段对应不充分，不合成单一成本","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2608.24680","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"a7315f2d7f4f5595d0ec94e8d221689ece630f0c078e7992b2f8bb03dcaedc34","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0174","title":"From Pixels to States: Rethinking Interactive World Models as Game Engines","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2607.14076","aliases":["pool:P174"],"technical_key":"arxiv-2607.14076","catalog_id":"P174","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2607.14076","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"8ec8df756a86b9882b8509c4809f35e228d51711936888b45499c6eeed675b04","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0175","title":"SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2605.15178","aliases":["pool:P175"],"technical_key":"arxiv-2605.15178","catalog_id":"P175","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.15178","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":84.0,"wall_hours_max":84.0,"time_bound":"approximate","gpu_hours_min":5376.0,"gpu_hours_max":5376.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"SANA-WM LTX2 VAE 适配；50K steps；3.5 天","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.15178","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":360.0,"wall_hours_max":360.0,"time_bound":"approximate","gpu_hours_min":23040.0,"gpu_hours_max":23040.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"SANA-WM 主 DiT 四阶段训练约 15 天；不含前述 VAE 适配","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.15178","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"0988659aa041128a3bca790c51f108a5ffbba03f02214cef5aa505eda832f90e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0176","title":"ActionParty: Multi-Subject Action Binding in Generative Video Games","year":"2026","venue":"ECCV（作者自报）","arxiv_id":"2604.02330","aliases":["pool:P176"],"technical_key":"arxiv-2604.02330","catalog_id":"P176","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2604.02330","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c5a4a6ae770f23555f494bfd5d85cfd79f34e05180bc859861bc48f60b6a6bea","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0177","title":"Beyond Pixel Histories: World Models with Persistent 3D State","year":"2026","venue":"ICML（作者自报）","arxiv_id":"2603.03482","aliases":["pool:P177"],"technical_key":"arxiv-2603.03482","catalog_id":"P177","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.03482","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":96.0,"wall_hours_max":96.0,"time_bound":"approximate","gpu_hours_min":768.0,"gpu_hours_max":768.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Persistent 3D State 的 pixel VAE；4 天","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2603.03482","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"pre","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":288.0,"wall_hours_max":288.0,"time_bound":"approximate","gpu_hours_min":2304.0,"gpu_hours_max":2304.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Persistent 3D State 的 3D VAE；12 天","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2603.03482","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"pre","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":576.0,"gpu_hours_max":576.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Persistent 3D State 的 3D-S denoiser；3 天；3D-XL / pixel denoiser 各 10 天，未合并变体","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2603.03482","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"pre","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":16.0,"wall_hours_max":16.0,"time_bound":"approximate","gpu_hours_min":16.0,"gpu_hours_max":16.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Persistent 3D State 的 camera model；只是多模块系统中的小模块","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2603.03482","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f26001ebf51e2b773d0e83f15fee3b64dd1bac5b05603bca3837b1613bde45d0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0178","title":"M$^\\text{4}$World: A Multi-view Multimodal Driving World Model for Interactive Object Manipulation and Minute-long Streaming","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2607.14005","aliases":["pool:P178"],"technical_key":"arxiv-2607.14005","catalog_id":"P178","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2607.14005","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c67839deffc865a3980b1c96f9abba868098c601680714145d5187c3336bf632","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0179","title":"DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2608.13489","aliases":["pool:P179"],"technical_key":"arxiv-2608.13489","catalog_id":"P179","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2608.13489","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4e4badcaf8c087d090ae6985518a4808395bf61ae10a0d43b118f36ee72ceeb7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0180","title":"IOI: Decoupling Kinematics and Physics for Interactive World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2606.23296","aliases":["pool:P180"],"technical_key":"arxiv-2606.23296","catalog_id":"P180","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.23296","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"46230f604b51ed940e00028b9deb8d485f4f35dc3270e7bb2ff2be7ba519f770","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0181","title":"Mem-World: Memory-Augmented Action-Conditioned World Models for Persistent Robot Manipulation","year":"2026","venue":"CoRL（作者自报）","arxiv_id":"2606.18960","aliases":["pool:P181"],"technical_key":"arxiv-2606.18960","catalog_id":"P181","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.18960","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":384.0,"gpu_hours_max":384.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Mem-World temporal attention fine-tuning；batch 32；约 2 天","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.18960","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Mem-World 针对机器人 domain shift 的 world model 5K steps；policy 另 20K steps","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.18960","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"424c88f1c9fde0d2ea70faaae8819cb902750f1a59ae83a6dce2d28c88e60b88","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0182","title":"$\\omega$-EVA: Envision, Verify, and Act with Latent Interactive World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2606.09457","aliases":["pool:P182"],"technical_key":"arxiv-2606.09457","catalog_id":"P182","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.09457","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":16.0,"count_max":16.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ω-EVA 三阶段 50 / 30 / 20 epochs；batch 1024；耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.09457","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"a1907fd4103dbebc5738257ffb3636e538d5a9f0c6d542436fb1ccf8d1dcd60a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0183","title":"Causal World Modeling for Robot Control","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2601.21998","aliases":["pool:P183"],"technical_key":"arxiv-2601.21998","catalog_id":"P183","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2601.21998","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"38a26c8dc39c99b58b0badd41c51279f5fd37108419cf47afdfa2cee13900d9a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0184","title":"R2M-Bench: Evaluating Revisit Memory via Relative Consistency in Interactive Video World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2608.27328","aliases":["pool:P184"],"technical_key":"arxiv-2608.27328","catalog_id":"P184","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2608.27328","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"546f36b4f6361c561450edebe2ba71c414dd1a61dfd30cc614a93d301998be3e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0185","title":"LivingWorld: Interactive 4D World Generation with Environmental Dynamics","year":"2026","venue":"ECCV（作者自报）","arxiv_id":"2604.01641","aliases":["pool:P185"],"technical_key":"arxiv-2604.01641","catalog_id":"P185","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2604.01641","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d1b2cd12fa13d67a9978de8bbdf9348ebd7e13ff95c51520bdb9190bf815c66d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0186","title":"Out of Sight but Not Out of Mind: Hybrid Memory for Dynamic Video World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.25716","aliases":["pool:P186"],"technical_key":"arxiv-2603.25716","catalog_id":"P186","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.25716","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"HM-World；10K iterations，batch 32；GPU 型号和时长未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.25716","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"d822108f685882346db106649869be48b32ca4811000ad065249b2ad5e241949","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0187","title":"MultiGen: Level-Design for Editable Multiplayer Worlds in Diffusion Game Engines","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2603.06679","aliases":["pool:P187"],"technical_key":"arxiv-2603.06679","catalog_id":"P187","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2603.06679","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0e9f68283cd89985f724f6b8322e1f291390bdf4d41e39026f47eb5353390c2b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0188","title":"ReWorld: An Interactive World Model with Long-Horizon Memory","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2608.23565","aliases":["pool:P188"],"technical_key":"arxiv-2608.23565","catalog_id":"P188","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2608.23565","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"12fbb650de036f8e8d78e2ab534cd9748ed2a4b4985eccd9d602678cf3ec99cb","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0189","title":"ForgeWM: Progressive Causal Training for Few-Step Action-Conditioned Video World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2608.14022","aliases":["pool:P189"],"technical_key":"arxiv-2608.14022","catalog_id":"P189","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2608.14022","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ForgeWM 多阶段训练使用 8 GPUs + FSDP；型号和时长未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2608.14022","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"63f259c37ccb5e8ddf51615a4ff04c072d4478a54cd9333e4b1c9f45d8ecb1a9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0190","title":"Cycle-World: Mitigating Error Accumulation in Long-term Video World Models via Reverse-Prediction Cycle Consistency","year":"2026","venue":"ECCV（作者自报）","arxiv_id":"2607.11836","aliases":["pool:P190"],"technical_key":"arxiv-2607.11836","catalog_id":"P190","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2607.11836","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f1f24727221ffa0e674575b0db87e478b161af191dae9941934c9fb28ceeeb7b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0191","title":"PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2608.13552","aliases":["pool:P191"],"technical_key":"arxiv-2608.13552","catalog_id":"P191","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2608.13552","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"651c33e2bc1d6fa5ed35dd9c8496bc70d6ed2629f159d25a1586559cf171b849","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0192","title":"WorldArena: A Unified Benchmark for Evaluating Perception and Functional Utility of Embodied World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2602.08971","aliases":["pool:P192"],"technical_key":"arxiv-2602.08971","catalog_id":"P192","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2602.08971","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"523d0eb1759d0d2e7c8887da8757e267e789d31594199188286f16a8ebae6cee","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0193","title":"WBench: A Comprehensive Multi-turn Benchmark for Interactive Video World Model Evaluation","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2605.25874","aliases":["pool:P193"],"technical_key":"arxiv-2605.25874","catalog_id":"P193","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2605.25874","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"fb08b7b5f36aa2380fbd335f82f509e7b50d6a8e60097a7d2d9a3cdbbdaade01","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0194","title":"MemLearner: Learning to Query Context memory for Video World Models","year":"2026","venue":"ECCV（作者自报）","arxiv_id":"2606.31734","aliases":["pool:P194"],"technical_key":"arxiv-2606.31734","catalog_id":"P194","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.31734","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9fb8123c55fb89b11262625ec07e8d065e675c405544b03f9b3f540079b32422","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0195","title":"Causal-rCM: A Unified Teacher-Forcing and Self-Forcing Open Recipe for Autoregressive Diffusion Distillation in Streaming Video Generation and Interactive World Models","year":"2026","venue":"不确定（arXiv；正式 venue 未核实）","arxiv_id":"2606.25473","aliases":["pool:P195"],"technical_key":"arxiv-2606.25473","catalog_id":"P195","library_id":"","topic_code":"","topic":"","task":"不确定：本轮尚未核实","directions":[],"source_url":"https://arxiv.org/html/2606.25473","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"89564cc3e36909bdb7194ebd63b159990d8e9b2da1fd5769c1787f89c0681d1f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0196","title":"Reflexion: Language Agents with Verbal Reinforcement Learning","year":"2023","venue":"未核实","arxiv_id":"","aliases":["library:P012"],"technical_key":"library-P012","catalog_id":"X196","library_id":"P012","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"其他（说明）","directions":["XR-E"],"source_url":"https://github.com/noahshinn/reflexion","status":"无需本文权重训练","resource_note":"Reflexion 用 verbal feedback 更新外部记忆，不更新语言模型权重。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"GPT-3（ALFWorld）；GPT-4（代码实验）","training_datasets":"无参数训练；ALFWorld、HotpotQA、HumanEval／MBPP 用于交互与评价","existing_method":"执行任务后由外部或内部评价器给出反馈，生成反思并写入有限情节记忆，下次提示时检索/拼接；在 ALFWorld、HotpotQA 和代码任务验证。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0f041332615b331a3ee6eabea223a60b7b2585ff7d5ae3d8bb8c2dd3765acc8d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0197","title":"Black Swan: Abductive and Defeasible Video Reasoning in Unpredictable Events","year":2025,"venue":"CVPR 2025","arxiv_id":"","aliases":["library:P021","lineage:BLACK"],"technical_key":"library-P021","catalog_id":"X197","library_id":"P021","topic_code":"B1","topic":"能力诊断与评测基准","task":"其他（说明）","directions":["AIGC-B","AIGC-C","XR-S"],"source_url":"https://arxiv.org/html/2412.05725v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"多种被测 VLM；Llama-3.1-8B-Instruct 用于语义评价","training_datasets":"不训练视频生成器；Oops! 派生 BlackSwanSuite 1,655 videos","existing_method":"分阶段给出视觉证据，评估溯因和可撤销推理；使用选择题准确率、CLIP/LLM 生成指标及人类评价。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d95c345832911a94f72532e581a0e73a1ba60bbd63c366e38cfacfb6b6669661","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0198","title":"Enabling Self-Improving Agents to Learn at Test Time With Human-In-The-Loop Guidance","year":"2025","venue":"未核实","arxiv_id":"2507.17131","aliases":["library:P022"],"technical_key":"arxiv-2507.17131","catalog_id":"X198","library_id":"P022","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"其他（说明）","directions":["XR-E"],"source_url":"https://arxiv.org/abs/2507.17131v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"GPT-4o；Qwen2.5-7B","training_datasets":"无新增权重训练；TikTok Pay 11,846 个顺序案例用于在线适应／评价","existing_method":"自对话评估不确定性，在查询预算内请求解释/纠错；抽取知识、加时间戳，检测冲突并澄清后检索用于后续决策。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"efd7dbe6109b600a4cd4d3bb108b01170874b92f12f8e9b21567364198040208","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0199","title":"Nested Learning: The Illusion of Deep Learning Architecture","year":"2025","venue":"未核实","arxiv_id":"2512.24695","aliases":["library:P024"],"technical_key":"arxiv-2512.24695","catalog_id":"X199","library_id":"P024","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"其他（说明）","directions":["XR-E"],"source_url":"https://arxiv.org/abs/2512.24695v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Hope；Titans；M3 optimizer","training_datasets":"具体语言模型训练语料待核","existing_method":"将梯度/上下文压缩解释为关联记忆；设计多个时间尺度更新的 MLP 记忆、M3 优化器，以及结合自指 Titans 的 Hope。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"11d3ba00ca8425bbe0e41910b8387521ca361ce48514e8594a0e3ac82734804e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0200","title":"Gen-Searcher: Reinforcing Agentic Search for Image Generation","year":"2026","venue":"未核实","arxiv_id":"2603.28767","aliases":["library:P033"],"technical_key":"arxiv-2603.28767","catalog_id":"X200","library_id":"P033","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"其他（说明）","directions":["AIGC-D"],"source_url":"https://arxiv.org/abs/2603.28767v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H800","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Gen-Searcher Qwen3-VL-8B SFT + agentic RL；SFT-10K / RL-6K；训练 rollout 另部署 16 H800 的图像工具","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2603.28767v2","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"inf","device":"H800","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Gen-Searcher RL 期间的 Qwen-Image-Edit rollout 服务；是额外运行资源，不是 optimizer GPU","comparable":false,"is_gpu":true,"role_label":"推理／评测／rollout","source_url":"https://arxiv.org/abs/2603.28767v2","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen3-VL-8B 搜索 agent；Qwen-Image 生成器","training_datasets":"Gen-Searcher-SFT-10k；Gen-Searcher-RL-6k","existing_method":"Qwen3-VL-8B 先 SFT 再 agentic RL；工具执行搜索/浏览，双奖励评价知识正确性与生成图的实际符合度。","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"5421550e1fc385d427b8ec597795754ef4efe59951e8b6af50d780a51bc89967","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0201","title":"SimStep: Human-in-the-Loop Authoring of Interactive Educational Simulations Through Task-Level Abstractions","year":"2026","venue":"未核实","arxiv_id":"","aliases":["library:P036"],"technical_key":"library-P036","catalog_id":"X201","library_id":"P036","topic_code":"X0","topic":"教程、技能反馈与空间指导的应用近邻","task":"其他（说明）","directions":["XR-C","XR-E"],"source_url":"https://doi.org/10.1145/3772318.3791514","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Claude 3.5 Sonnet；Puppeteer 测试","training_datasets":"无新增权重训练；教师创作任务用于系统评价","existing_method":"教师在抽象图上验证和编辑，Claude 生成/修改代码；自动测试和图形回溯定位隐含假设；进行教师使用与抽象保真度研究。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c3eaddbabe4ec583e23faff50177f4ad1e6bafceb7b468a4ffce55f365ca32b4","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0202","title":"Sensible Agent: A Framework for Unobtrusive Interaction with Proactive AR Agents","year":"2025","venue":"未核实","arxiv_id":"","aliases":["library:P059"],"technical_key":"library-P059","catalog_id":"X202","library_id":"P059","topic_code":"H1","topic":"主动交互、协作与用户需求","task":"其他（说明）","directions":["XR-L"],"source_url":"https://doi.org/10.1145/3746059.3747748","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Sensible Agent；GPT-4o 用于研究中的意图格式化，运行时 LMM 版本待核","training_datasets":"未核实有无参数训练；专家工作坊 N=12、标注研究 N=40、原型研究 N=10 是研究样本","existing_method":"第一视角相机/传感器与LMM推断活动和约束，选择图标、二选一或多选建议及可用输入模态；与语音提问基线做被试内比较。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f16e490f9e69a91188f4ed49bd8cbe95020cb31d93fded1fa0d70be0b608b1f6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0203","title":"AgentHands: Generating Interactive Hand Gestures for Spatially Grounded Agent Conversations in XR","year":"2026","venue":"未核实","arxiv_id":"","aliases":["library:P061"],"technical_key":"library-P061","catalog_id":"X203","library_id":"P061","topic_code":"X0","topic":"教程、技能反馈与空间指导的应用近邻","task":"其他（说明）","directions":["XR-S"],"source_url":"https://doi.org/10.1145/3772318.3790938","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Gemini 2.5 Flash；Unity／OpenXR","training_datasets":"无新增权重训练；N=10 形成性研究","existing_method":"N=10形成性研究→LLM生成按词对齐的手势类型/参数→解析为时间戳姿态→Unity/OpenXR渲染；Galaxy XR、Gemini2.5Flash与语音识别/合成。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"737b5960d30bda2d0d4733f4c2e586fc3e1bca0a4c7de63fd81e7fde2448df22","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0204","title":"Interactive Visual Reasoning under Uncertainty","year":2023,"venue":"NeurIPS 2023 Datasets and Benchmarks","arxiv_id":"2206.09203","aliases":["library:P143","lineage:IVR"],"technical_key":"arxiv-2206.09203","catalog_id":"X204","library_id":"P143","topic_code":"B1","topic":"能力诊断与评测基准","task":"其他（说明）","directions":["AIGC-D","XR-C"],"source_url":"https://arxiv.org/html/2206.09203v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 3090 / V100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Interactive Visual Reasoning 的 RL agents 10^7 steps；GPU 数和耗时未定位","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2206.09203v2","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"DDPG／TD3／PPO、启发式、LLM 等对照","training_datasets":"CLEVR 风格 IVRE 合成环境；RL 训练／评测各自划分","existing_method":"智能体同时输出下一实验的物体选择与各物体Blicketness belief；观测为二进制向量或160×120合成图。奖励联合全体识别成功、每步成本及与搜索oracle belief的Jensen–Shannon距离。对比随机/Bayes/逐个测试/NOTEARS/搜索组合、actor-critic RL与GPT-3.5/4。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"83a0073efc6beeeed98fe0669c0f5483957742526eee185d8aae360cab9c693a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0205","title":"ExpertAF: Expert Actionable Feedback from Video","year":"2025 / 2024","venue":"未核实","arxiv_id":"","aliases":["library:P146"],"technical_key":"library-P146","catalog_id":"X205","library_id":"P146","topic_code":"X0","topic":"教程、技能反馈与空间指导的应用近邻","task":"其他（说明）","directions":["XR-C","XR-E"],"source_url":"https://openaccess.thecvf.com/content/CVPR2025/papers/Ashutosh_ExpertAF_Expert_Actionable_Feedback_from_Video_CVPR_2025_paper.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"V100","count_min":8.0,"count_max":8.0,"card_memory_gb":32.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ExpertAF 可训练模块与语言模型适配；video / pose encoders 冻结；5 epochs","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://openaccess.thecvf.com/content/CVPR2025/papers/Ashutosh_ExpertAF_Expert_Actionable_Feedback_from_Video_CVPR_2025_paper.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Llama3-8B；InternVideo2；Llama3-70B 数据整理","training_datasets":"25,505 个弱监督训练 tuple；测试 1,272 个 tuple 不计入训练","existing_method":"Llama3-70B压缩专家评论并标注身体部位/正确性；按技能等级匹配错误与正确演示，用PA-MPJPE对齐时间并过滤。InternVideo2视频特征、姿态codebook和文本映射至Llama3-8B，联合学习评论、示范检索与姿态token生成。训练25,505个tuple，人工核验测试1,272个。","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"49f6f1dd9bfc50a692f8d26c31e599400360d0d5e6b2da91203001916929c37f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0206","title":"Learning to Clarify: Multi-turn Conversations with Action-Based Contrastive Self-Training","year":2025,"venue":"ICLR 2025","arxiv_id":"","aliases":["library:P149","lineage:ACT"],"technical_key":"library-P149","catalog_id":"X206","library_id":"P149","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"其他（说明）","directions":["AIGC-D","XR-L"],"source_url":"https://arxiv.org/html/2406.00222v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":16.0,"wall_hours_max":16.0,"time_bound":"approximate","gpu_hours_min":128.0,"gpu_hours_max":128.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Learning to Clarify：full ACT 最长配置，PACIFIC 250 conversations / 1,923 examples；12 epochs；其他实验不能默认同耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2406.00222v2","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"澄清对话 LLM；具体 checkpoint 待核","training_datasets":"PACIFIC、machine reading comprehension、AmbigSQL","existing_method":"先根据正确/错误对话动作产生winning/losing响应；训练中从当前模型采样，识别动作是否正确。对正确动作继续模拟用户和助手对话，按最终语义/SQL执行结果更新偏好对，再进行DPO式优化。主模型Zephyr-7B，比较50/100/250条对话下的SFT、IRPO及提示基线。","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"23ab273b7130dfd3841da96a979cccde0a002c0df2f572a23968f5842775602d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0207","title":"Exploring Exploration with Foundation Agents in Interactive Environments","year":"2025 / 2024","venue":"未核实","arxiv_id":"2412.06438","aliases":["library:P157"],"technical_key":"arxiv-2412.06438","catalog_id":"X207","library_id":"P157","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"其他（说明）","directions":["XR-L"],"source_url":"https://arxiv.org/abs/2412.06438v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"GPT-4o、o4-mini、Claude 3.7、Gemini 2.5 Pro 等被测模型","training_datasets":"无新模型训练；Feature World、TextAlchemy 是交互测试环境","existing_method":"Feature World中寻找产生奖励的颜色/形状属性或属性组合，比较模型探索与随机及理想策略；将Gemini扩展到三维视频输入和人类执行。TextAlchemy比较GPT-4o、o4-mini、Claude 3.7与Gemini 2.5 Pro，控制是否给结构先验、是否每轮自总结，并在后续试次静默更换环境机制。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"14ff0d4d38a5f542203826c6a4c72eac447e4e3a7052a60d3832adf256d1f0ba","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0208","title":"Counterfactual World Models via Digital Twin-conditioned Video Diffusion","year":2025,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","arxiv_id":"2511.17481","aliases":["library:P071","lineage:CWMDT"],"technical_key":"arxiv-2511.17481","catalog_id":"X208","library_id":"P071","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"编辑与修改","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2511.17481v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 4090（原文标注）","count_min":1.0,"count_max":1.0,"card_memory_gb":48.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Counterfactual Digital Twin：95 paired samples，LTX-Video 适配；原文写 RTX 4090 48GB，与常见标准规格不符，需作者核对；不作为 24GB 消费卡推荐依据","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2511.17481v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"LTX-Video；Qwen3-VL-8B；Qwen-Image-Edit-2509","training_datasets":"RVTBench 95 对 digital-twin／视频样本","existing_method":"对象身份、类别、属性、质心／深度／大小与 mask 等来自感知模块；数字孪生是结构化表示。由 LLM 预测未来，生成器接受其条件；每次干预可采样多个合理未来。","training_signal":"不确定：本轮尚未核实","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"d40d929c329be9ada2b14db34da6c90643d291b2f65e6adec42dad0067a183e1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0209","title":"GPS: Graph-guided Proactive Information Seeking in Large Language Models","year":2026,"venue":"ICLR 2026","arxiv_id":"","aliases":["library:P159","lineage:GPS","recent:GPS"],"technical_key":"library-P159","catalog_id":"X209","library_id":"P159","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"其他（说明）","directions":["AIGC-D","XR-L"],"source_url":"https://proceedings.iclr.cc/paper_files/paper/2026/file/1f49b7fbd7c7afc52c1db0d4ed1a338d-Paper-Conference.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A800","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"GPS GRPO LoRA rank 64；1 epoch，batch 32；耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://proceedings.iclr.cc/paper_files/paper/2026/file/1f49b7fbd7c7afc52c1db0d4ed1a338d-Paper-Conference.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"GPS Reasoner；具体 LLM checkpoint 待核","training_datasets":"ConditionalQA 构造的 3,250 条 synthetic train；Synthetic test、ConditionalQA、ShARC 另作评测","existing_method":"Reasoner输出条件变量、值分支与答案叶节点；Clarifier按可达路径代价选择问题，收到回答后剪枝。用ConditionalQA中的欠明确样本及DeepSeek-R1合成条件路径，筛选完整输入可答而缺失输入不可答的数据。GRPO训练Reasoner，结合正确性、轮数惩罚与结构质量奖励；在Qwen2.5-7B和Llama3-8B骨干上验证。","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"e8be97fb57eb9b157928b62500cc511052ecf15098158e30bd0f6a3340a8ede6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0210","title":"Useful Memories Become Faulty When Continuously Updated by LLMs","year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","arxiv_id":"2605.12978","aliases":["library:P097","lineage:MEMFAULT"],"technical_key":"arxiv-2605.12978","catalog_id":"X210","library_id":"P097","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"其他（说明）","directions":["XR-E"],"source_url":"https://arxiv.org/html/2605.12978v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"GPT-4.1／GPT-5-nano 轨迹；GPT-5.4 abstraction；其他配置见正文","training_datasets":"ALFWorld、ScienceWorld、WebShop、AppWorld、Mind2Web、ARC-AGI Stream 轨迹；不等于 LLM 权重训练集","existing_method":"固定经验池比较Static-All、Static-Group、Stream与只保留episodic轨迹；允许Retain/Delete/Consolidate；分析错分组、过度概括和窄分布过拟合。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0b396498366e6c036506ffe5357afc32655acf5270f64f5eb2b499dfbe6ff583","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0211","title":"Learning Long-form Movie Prior via Large Language Models","year":"2026","venue":"未核实","arxiv_id":"","aliases":["library:P161"],"technical_key":"library-P161","catalog_id":"X211","library_id":"P161","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"其他（说明）","directions":["AIGC-C"],"source_url":"https://doi.org/10.1109/TPAMI.2026.3729756","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"LLaMA-3／3.2 1B、3B、8B；ROICtrl 用于可视化","training_datasets":"Storyboard20K（MovieNet、LSMDC）：18,207 train；905／1,198 test","existing_method":"先用检测、姿态估计和人脸识别自动标注，再由7名标注者校准人物身份、框、关键点及开放类别场景物体，并进行多轮复核。将连续坐标量化到m个整数区间，与文本共同tokenize；大于96×96像素面积的人物用98个全身关键点，否则用17个关键点或边界框。预训练时随机交换script、objects、main characters的顺序，以最大似然学习联合分布；随后把概括性标注作为指令、剧本和空间标注作为回答进行微调。训练/testA/testB分别为18,207/905/1,198个分镜，testA使用概括描述，testB使用逐镜头描述。对照商业模型获得2个上下文示例。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ae98d5b379e4559e87c03a38fff8e526a151496ce91b01fd44f164cf25b22143","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0212","title":"ARCap: Collecting High-quality Human Demonstrations for Robot Learning with Augmented Reality Feedback","year":"2024","venue":"未核实","arxiv_id":"2410.08464","aliases":["library:P251"],"technical_key":"arxiv-2410.08464","catalog_id":"X212","library_id":"P251","topic_code":"X0","topic":"教程、技能反馈与空间指导的应用近邻","task":"其他（说明）","directions":["XR-L"],"source_url":"https://arxiv.org/abs/2410.08464","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 3090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ARCap 下游 autonomous policy 训练/测试工作站；不是预训练模型 fine-tuning 成本","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2410.08464","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.5（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"ARCap 数据采集与机器人模仿学习系统；policy 版本待核","training_datasets":"ARCap 采集的机器人操作示范；任务清单与规模待核","existing_method":"ARCap 通过增强现实视觉反馈和触觉警告，帮助用户避免碰撞并匹配机器人运动学。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"8a79ecb30f0de92d15d2544aaff9f96468708f26a61b7e308ad0f982f71a89d6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0213","title":"Theory of Space: Can Foundation Models Construct Spatial Beliefs through Active Exploration?","year":2026,"venue":"ICLR 2026","arxiv_id":"2602.07055","aliases":["library:P342","recent:SPACE"],"technical_key":"arxiv-2602.07055","catalog_id":"X213","library_id":"P342","topic_code":"B1","topic":"能力诊断与评测基准","task":"其他（说明）","directions":["AIGC-B","XR-S"],"source_url":"https://arxiv.org/abs/2602.07055","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"GPT-5.2、Gemini-3-Pro、Claude-4.5-Sonnet、GLM-4.6V、Qwen3-VL 等被测模型","training_datasets":"无新增权重训练；Theory of Space 合成房间场景用于评价","existing_method":"Theory of Space 用逐步空间信念探针分析探索效率、知识退化及错误信念的修正。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d0a98826f396a0052e7c6f25235cb8faced07f8c4aba6ab0a6f21e2a12251d63","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0214","title":"ESI-Bench: Towards Embodied Spatial Intelligence that Closes the Perception-Action Loop","year":"2026","venue":"未核实","arxiv_id":"2605.18746","aliases":["library:P372"],"technical_key":"arxiv-2605.18746","catalog_id":"X214","library_id":"P372","topic_code":"B1","topic":"能力诊断与评测基准","task":"其他（说明）","directions":["XR-L"],"source_url":"https://arxiv.org/abs/2605.18746","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"多种被测 foundation models；GPT-4o 辅助出题","training_datasets":"无新增权重训练；BEHAVIOR-1K／OmniGibson 场景构造 ESI-Bench","existing_method":"ESI-Bench 在 OmniGibson 中让观察者执行动作取证，比较主动探索、被动观察与三维辅助。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"07b402e40bf635a97cf4d04cdf1640649dab3144d1d97667f6d4332a30e651a6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0215","title":"Direct-a-Video: Customized Video Generation with User-Directed Camera Movement and Object Motion","year":"2024","venue":"未核实","arxiv_id":"2402.03162","aliases":["library:P214"],"technical_key":"arxiv-2402.03162","catalog_id":"X215","library_id":"P214","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"条件控制生成","directions":["AIGC-C"],"source_url":"https://arxiv.org/abs/2402.03162","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"V100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Direct-a-Video 相机控制模块训练；batch 8；未定位耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2402.03162","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Direct-a-Video；底座 checkpoint 待核","training_datasets":"静止相机视频与模拟相机运动增强；原始来源待核","existing_method":"Direct-a-Video 分别建模相机参数和对象运动，以新增时间交叉注意力和数据增强训练相机控制。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"9cf98a70f4e2857888e09fd2a552be6ed96083c83b1c095073e03027aaa4f168","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0216","title":"MovieBench: A Hierarchical Movie Level Dataset for Long Video Generation","year":"2024","venue":"未核实","arxiv_id":"2411.15262","aliases":["library:P259"],"technical_key":"arxiv-2411.15262","catalog_id":"X216","library_id":"P259","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"长视频与多镜头叙事","directions":["AIGC-C"],"source_url":"https://arxiv.org/abs/2411.15262","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"多种被测生成器；GPT-4o 等辅助标注","training_datasets":"MovieBench；AutoStory 对照为各角色训练 LoRA，不能归为统一新生成器","existing_method":"MovieBench 以电影和镜头层级组织视频、角色、音频与描述。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ac2d904f796cd233b64aa4d8394660ef24dc4824dd84eb0d1a1a6b0b0b7548ce","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0217","title":"Demo2Tutorial: From Human Experience to Multimodal Software Tutorials","year":"2026","venue":"未核实","arxiv_id":"2606.03951","aliases":["library:P379"],"technical_key":"arxiv-2606.03951","catalog_id":"X217","library_id":"P379","topic_code":"X0","topic":"教程、技能反馈与空间指导的应用近邻","task":"其他（说明）","directions":["XR-E"],"source_url":"https://arxiv.org/abs/2606.03951","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"GPT-4o；另测 Qwen-VL-32B-Instruct","training_datasets":"无新增权重训练；TutorialBench 110 教程／7 软件用于评价","existing_method":"GPT-4o 解析状态变化和意图，Step Planner 组成层级图；不是训练视频扩散模型来预测下一操作后果。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"30c9289d1a81042ed195671d5de8026414e972a68b02169687842c203ad4c39b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0218","title":"ReGRPO: Reflection-Augmented Policy Optimization for Tool-Using Agents","year":"2026","venue":"未核实","arxiv_id":"2606.31392","aliases":["library:P390"],"technical_key":"arxiv-2606.31392","catalog_id":"X218","library_id":"P390","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"其他（说明）","directions":["XR-E"],"source_url":"https://arxiv.org/abs/2606.31392","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen2-VL-7B；GPT-4o 离线生成反思","training_datasets":"MM-Traj 派生错误／修复轨迹；GTA、GAIA 是评测","existing_method":"ReGRPO 用错误类型、证据和修复计划构造训练数据，再联合优化反思与纠正动作并计入反思成本。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0a014d007c2c97b19cabcea4797f094ec2582fba52b80a7076619328b5c45cdc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0219","title":"AdaJEPA: An Adaptive Latent World Model","year":"2026","venue":"未核实","arxiv_id":"2606.32026","aliases":["library:P391"],"technical_key":"arxiv-2606.32026","catalog_id":"X219","library_id":"P391","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"动作条件的预测与交互世界模型","directions":["XR-S"],"source_url":"https://arxiv.org/abs/2606.32026","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"AdaJEPA；JEPA encoder／predictor","training_datasets":"环境实际转移；近期 5 个样本回放","existing_method":"测试时每次 MPC 后用自监督转移损失更新编码器、预测器的后层；近期 5 个样本回放，每次一步梯度更新。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"6f46f32606149d6f620b6df292c41cbe66cc7402f64181d5ea0e29fbb0cce066","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0220","title":"Vision-G1: Towards General Reasoning Vision-Language Models via Reinforcement Learning","year":"2026","venue":"未核实","arxiv_id":"","aliases":["library:P112"],"technical_key":"library-P112","catalog_id":"X220","library_id":"P112","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"其他（说明）","directions":["XR-L"],"source_url":"https://doi.org/10.1609/aaai.v40i33.40039","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":18.0,"wall_hours_max":18.0,"time_bound":"approximate","gpu_hours_min":144.0,"gpu_hours_max":144.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Vision-G1 Qwen2.5-VL RL 两轮；batch 128，32 rollouts / prompt","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://doi.org/10.1609/aaai.v40i33.40039","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.4（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Vision-G1-7B","training_datasets":"46 个数据集、5 域 13 维度；完整清单待核","existing_method":"统一题目/奖励格式，影响函数过滤低价值样本，再按中等难度及类别平衡进行多轮GRPO。","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"03b3a8e716c37de07a137c73082b68b76f2eaaaaeb108530f8a1b313bfaa6e64","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0221","title":"Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective","year":"2025","venue":"未核实","arxiv_id":"2506.14965","aliases":["library:P117"],"technical_key":"arxiv-2506.14965","catalog_id":"X221","library_id":"P117","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"其他（说明）","directions":["XR-L"],"source_url":"https://arxiv.org/abs/2506.14965","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"Hopper（具体型号未定位）","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":60.0,"wall_hours_max":60.0,"time_bound":"approximate","gpu_hours_min":1920.0,"gpu_hours_max":1920.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Cross-Domain RL 7B：4 nodes × 8 Hopper；2 epochs，2.5 天","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2506.14965","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"Hopper（具体型号未定位）","count_min":128.0,"count_max":128.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":60.0,"wall_hours_max":60.0,"time_bound":"approximate","gpu_hours_min":7680.0,"gpu_hours_max":7680.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Cross-Domain RL 32B：16 nodes × 8 Hopper；2 epochs，2.5 天；不是把总集群 24 nodes 全算入单个实验","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2506.14965","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen2.5-Base","training_datasets":"数学、代码、科学、逻辑、仿真、表格语料；17 任务评测另计","existing_method":"来源清洗去重、领域/难度过滤，Qwen2.5-Base上GRPO训练，控制域混合和难度研究迁移。","training_signal":"不确定：本轮尚未核实","hardware_bin":"large","max_reported_ft_gpus":128.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"916d81e4f0d2bda09e69aecd210315fab3f4c756763ca1423c66332c9c2e803b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0222","title":"Causally Steered Diffusion for Automated Video Counterfactual Generation","year":"2025","venue":"未核实","arxiv_id":"2506.14404","aliases":["library:P125"],"technical_key":"arxiv-2506.14404","catalog_id":"X222","library_id":"P125","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"编辑与修改","directions":["XR-C"],"source_url":"https://arxiv.org/abs/2506.14404v2","status":"已核对资源阶段","resource_note":"CSVC 通过 LLM 优化文本 prompt，视频 diffusion backbone 不进行 fine-tuning。","stages":[{"role":"inf","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"CSVC 的 prompt / textual gradient 优化与视频生成；不更新视频模型权重；另有 GPT-4 API","comparable":false,"is_gpu":true,"role_label":"推理／评测／rollout","source_url":"https://arxiv.org/abs/2506.14404v2","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"TokenFlow／FLATTEN／Tune-A-Video 等编辑工具；VLM 评价","training_datasets":"框架不新增编辑器权重训练；工具自身适配另计","existing_method":"因果目标/解耦提示→TokenFlow/FLATTEN/Tune-A-Video等编辑器→VLM评估损失→语言提示迭代，核心框架不需改编辑器内部。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ff5da7ff6badf2423b5d5b051b770371d26a85f65614a1340f31afc72c9eb583","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0223","title":"OneStory: Coherent Multi-Shot Video Generation with Adaptive Memory","year":2026,"venue":"CVPR 2026","arxiv_id":"2512.07802","aliases":["library:P133","lineage:ONE","recent:ONE"],"technical_key":"arxiv-2512.07802","catalog_id":"X223","library_id":"P133","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"长视频与多镜头叙事","directions":["AIGC-C"],"source_url":"https://arxiv.org/html/2512.07802v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":128.0,"count_max":128.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"OneStory 整个 Wan2.1 I2V 模型 fine-tuning；1 epoch，480×832；未定位耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2512.07802v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.1 I2V","training_datasets":"自建 60K multi-shot dataset","existing_method":"按当前镜头字幕检索历史信息帧，重要性引导patchification压缩上下文，直接条件注入下一镜头视频生成器。","training_signal":"不确定：本轮尚未核实","hardware_bin":"large","max_reported_ft_gpus":128.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"16c64eec4eb258fd0dd9331f88024e17ff2ef69fc949ad74975992677f51bc4a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0224","title":"Qwen-Image-Agent: Bridging the Context Gap in Real-World Image Generation","year":2026,"venue":"arXiv 2026","arxiv_id":"2606.26907","aliases":["library:VG17","lineage:QWEN","recent:QWEN"],"technical_key":"arxiv-2606.26907","catalog_id":"X224","library_id":"VG17","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"其他（说明）","directions":["AIGC-D","XR-L"],"source_url":"https://arxiv.org/html/2606.26907v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Qwen-Image-2.0；MLLM planner（版本待核）","training_datasets":"无新增权重训练；IA-Bench、Mindbench、WISE-Verified 为评测","existing_method":"相较直接渲染提示，它把缺什么信息、去哪里补以及如何整合显式化，并用 IA-Bench 评测。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"7adedfacbdf3c408520217763b24d04a7697245320ac4569de5b3d003f3de9f5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0225","title":"Video-as-Answer: Predict and Generate Next Video Event with Joint-GRPO（VANS）","year":2026,"venue":"CVPR 2026","arxiv_id":"2511.16669","aliases":["library:VG31","lineage:VANS","reasoning:CVPR-2026-3642"],"technical_key":"arxiv-2511.16669","catalog_id":"X225","library_id":"VG31","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"视频作为推理或回答","directions":["AIGC-A","AIGC-D","XR-C"],"source_url":"https://arxiv.org/html/2511.16669v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Wan2.1-1.3B；VLM planner（具体版本见原模型记录）","training_datasets":"VANS-Data-100K","existing_method":"训练有参考文字／视频答案。Joint-GRPO 分阶段：先固定 VDM 更新 VLM，再固定 VLM 优化 VDM；第二阶段筛选与参考语义相符的锚点。使用格式、文本及视频语义一致性等奖励。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c4240cffd222d9984b709f550b54a34270de6b290c8f36f74bce092e1cc814bb","reasoning_id":"CVPR-2026-3642","reasoning_role":"R2","reasoning_topic":"通用生成框架 / 联合音视频"},{"id":"R0226","title":"VOID: Video Object and Interaction Deletion","year":2026,"venue":"ECCV 2026","arxiv_id":"2604.02296","aliases":["library:VG13","lineage:VOID","recent:VOID","reasoning:ECCV-2026-0955"],"technical_key":"arxiv-2604.02296","catalog_id":"X226","library_id":"VG13","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"编辑与修改","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2604.02296v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"CogVideoX DiT，初始化 Generative Omnimatte 权重后微调","training_datasets":"Kubric + HUMOTO 构建成对 counterfactual object-removal 视频；含删除后 collision / support 等后果变化","existing_method":"配对模拟在保持其余场景设定的情况下移除对象；quadmask 表达删除、受影响与保持等区域语义。用户确定目标，VLM 辅助判断影响范围，视频生成器执行连续运动与视觉补全。","training_signal":"相同场景删除对象前后的成对视频监督，让模型同时学习删除目标与改写后续互动","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"27e18f30c1d542a4fca0a61ab9501026471c18708d23f7738a9997f5551c2cd7","reasoning_id":"ECCV-2026-0955","reasoning_role":"R1","reasoning_topic":"Editing / 修改已有视频"},{"id":"R0227","title":"Geometry-Instructed Video Editing（GIVE）","year":2026,"venue":"arXiv 2026","arxiv_id":"2606.24225","aliases":["library:VG46","lineage:GIVE","recent:GIVE"],"technical_key":"arxiv-2606.24225","catalog_id":"X227","library_id":"VG46","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"编辑与修改","directions":["AIGC-A","AIGC-D","XR-C"],"source_url":"https://arxiv.org/html/2606.24225v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"约 1B 预训练 latent video Transformer；另测 Wan2.1","training_datasets":"Unreal Engine 构建 GIVE-Data；GIVE-Bench 为评测","existing_method":"深度、包围盒、朝向等几何流与 RGB 联合条件化；训练用引擎配对数据。推理以用户输入、视频分割和单目几何估计得到控制，不要求测试时完整 3D 代理、引擎特权缓冲或逐视频优化。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"afd331554620a37c19732bfde4ccb9a371f1d8bad6a99a000ae76056ec3842c3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0228","title":"What-If World: A Causal Benchmark for General World Models in Embodied Scenarios","year":"2026","venue":"未核实","arxiv_id":"2605.27589","aliases":["library:VG16"],"technical_key":"arxiv-2605.27589","catalog_id":"X228","library_id":"VG16","topic_code":"B1","topic":"能力诊断与评测基准","task":"动作条件的预测与交互世界模型","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2605.27589","status":"已核对资源阶段","resource_note":"What-If World 明确说明所有被测模型均不 fine-tune、distill 或 adapt。","stages":[{"role":"inf","device":"RTX A6000","count_min":4.0,"count_max":4.0,"card_memory_gb":48.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"What-If World 评测生成总计约 1,300 GPU-hours；明确不 fine-tune / distill / adapt；不可计作训练成本","comparable":false,"is_gpu":true,"role_label":"推理／评测／rollout","source_url":"https://arxiv.org/html/2605.27589","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"多个被测 world／video models；Gemini 3.1 Pro judge","training_datasets":"无模型微调；nuScenes／DROID 派生 319 配对任务为评测","existing_method":"319 个配对任务来自 nuScenes／DROID，覆盖六类物理变量；APEO 评价含条件遵循、物理、环境与配对结果，模型评判经人工核查。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d34db53cd77c1d8d8caa94827e60243369fbbfb702a79f42d03c8c529f63c3ca","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0229","title":"The Past Frames the Future: Memory for Autoregressive Video Generation — A Survey","year":"2026","venue":"未核实","arxiv_id":"2609.28466","aliases":["library:VG30"],"technical_key":"arxiv-2609.28466","catalog_id":"X229","library_id":"VG30","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"长视频与多镜头叙事","directions":["AIGC-B","XR-S"],"source_url":"https://arxiv.org/html/2609.28466","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"综述；不提出单一模型","training_datasets":"不适用：综述无自身训练集","existing_method":"相较把长上下文等同于记忆，它强调历史信息需在离开窗口后仍准确、可访问并影响生成。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b31074a51273b8959dcaea3399a6345aa22545dae6b96513a83ecc5a2fd12a5b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0230","title":"VideoSSM: Autoregressive Long Video Generation with Hybrid State-Space Memory","year":"2025","venue":"未核实","arxiv_id":"2512.04519","aliases":["library:VG63"],"technical_key":"arxiv-2512.04519","catalog_id":"X230","library_id":"VG63","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"长视频与多镜头叙事","directions":["AIGC-B","XR-S"],"source_url":"https://arxiv.org/html/2512.04519","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"VideoSSM；具体初始化 checkpoint 待核","training_datasets":"不确定：本轮尚未核实","existing_method":"相较滑窗遗忘或固定首帧锚导致重复，它协调动态全局记忆与局部运动信息。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0666c5ecbd2b20aeea064ffb90dbfcb953e56435217662a871c64ca81557aea0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0231","title":"A²RD: Agentic Autoregressive Diffusion for Long Video Consistency","year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","arxiv_id":"2605.06924","aliases":["library:VG15","lineage:A2RD"],"technical_key":"arxiv-2605.06924","catalog_id":"X231","library_id":"VG15","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"长视频与多镜头叙事","directions":["AIGC-B","XR-S"],"source_url":"https://arxiv.org/html/2605.06924v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"多种现成视频生成工具；具体配置待核","training_datasets":"无新增权重训练；LVBench-C、VBench-Long 为评测","existing_method":"相较开放环路拼接，它在片段进入后续生成前检查并改善实体、环境与叙事的一致性。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4855365bcf05db581d504e7d40afa62ca3927b4e0583e8c880d1392059440580","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0232","title":"See, Remember, Explore: A Benchmark and Baselines for Streaming Spatial Reasoning（S³-Bench / AMF-VLM）","year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","arxiv_id":"2603.23864","aliases":["library:VG60","lineage:S3"],"technical_key":"arxiv-2603.23864","catalog_id":"X232","library_id":"VG60","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"其他（说明）","directions":["AIGC-D","XR-L"],"source_url":"https://arxiv.org/html/2603.23864v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"B200","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"S³-Bench baseline training 使用 B200；数量和耗时未定位；latency 测量另用 H800","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.23864v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"AMF-VLM／Qwen3-VL-8B","training_datasets":"streaming spatial／active exploration／memory folding 自建数据＋VSI590K","existing_method":"相较看完整视频再答题，它要求仅用当时可得观察，在有限计算下主动弥补视野不足。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"6da3c16e5eab095c740e738fe87458e17081de6aedbc8dd8f7c3a4a1df5c996e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0233","title":"ArtiMo: Agent-Driven Articulated Mesh Animation","year":"2026","venue":"未核实","arxiv_id":"2608.20699","aliases":["library:VG43"],"technical_key":"arxiv-2608.20699","catalog_id":"X233","library_id":"VG43","topic_code":"G3","topic":"三维资产、重建与动画","task":"其他（说明）","directions":["XR-C"],"source_url":"https://arxiv.org/html/2608.20699","status":"已核对资源阶段","resource_note":"ArtiMo 用 VLM/LLM APIs 规划与修正 Blender 动作程序，本地渲染，不训练模型。","stages":[{"role":"inf","device":"RTX 4090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ArtiMo 本地 Blender / PyTorch3D 几何与渲染；VLM / LLM 通过 API；不属于权重训练","comparable":false,"is_gpu":true,"role_label":"推理／评测／rollout","source_url":"https://arxiv.org/html/2608.20699","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Gemini 3.1 Pro；ChatGPT 5.4；Blender／PyTorch3D","training_datasets":"无新增权重训练；关节物体动画任务用于评测","existing_method":"依据 URDF 中的关节结构与约束进行动作规划，用渲染关键帧和视觉反馈修订关节物体动画。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e50d1bf6f9a4dc31fb6ae65f9c180f79c741f257e9858d02686dd563e7d95553","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0234","title":"VLMs are Good Teachers for Video Reasoning via Adaptive Test-Time Optimization","year":"2026","venue":"未核实","arxiv_id":"2606.02564","aliases":["library:VG37"],"technical_key":"arxiv-2606.02564","catalog_id":"X234","library_id":"VG37","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"视频作为推理或回答","directions":["XR-E","XR-S"],"source_url":"https://arxiv.org/html/2606.02564","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"4-step distilled Wan2.2-5B；Qwen3-VL-4B teacher；LightX2V decoder","training_datasets":"VBVR-Bench 30K 训练实例作域适应；测试时 VLM 规则奖励","existing_method":"相较 VLM 只写提示或反复抽样选优，它让规则验证反馈直接改变视频模型的执行行为。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"38804d643877451ddc5182652ee745097e221482bb8109bf1b33db8ac4c6ddee","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0235","title":"JoyAI-Video-Edit: Real-Time Open-Ended Video Editing with Autoregressive Diffusion","year":"2026","venue":"未核实","arxiv_id":"2608.03974","aliases":["library:VG50"],"technical_key":"arxiv-2608.03974","catalog_id":"X235","library_id":"VG50","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"编辑与修改","directions":["AIGC-B","XR-S"],"source_url":"https://arxiv.org/html/2608.03974","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"JoyAI-Video-Edit；MLLM＋causal video VAE＋MM-DiT","training_datasets":"不确定：本轮尚未核实","existing_method":"相较离线整段编辑或逐段独立处理，它联合解决有界计算、源内容保持和少步生成的长期漂移。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2d92ef387bc929b16066b50b761236d36e802f22cca5c4ae0c6ab44ea482edcc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0236","title":"VOYAGER: An Open-Ended Embodied Agent with Large Language Models","year":2024,"venue":"TMLR 2024","arxiv_id":"2305.16291","aliases":["library:VG02","lineage:VOY"],"technical_key":"arxiv-2305.16291","catalog_id":"X236","library_id":"VG02","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"其他（说明）","directions":["XR-E","XR-L"],"source_url":"https://arxiv.org/html/2305.16291v2","status":"无需本文权重训练","resource_note":"Voyager 通过自动课程、技能代码库与交互反馈改善行为，不更新 LLM 权重。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"GPT-4；Mineflayer 可执行技能","training_datasets":"无权重训练；Minecraft 执行经验写入 skill library","existing_method":"相较一次性任务规划，它把成功程序存成可检索技能，在不更新模型参数的情况下支持持续积累。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3f2d3dd30c69b3616258eae52d62862972439aacfd250780759bbb5b7887b139","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0237","title":"What Changed and What Could Have Changed? State-Change Counterfactuals for Procedure-Aware Video Representation Learning","year":2025,"venue":"ICCV 2025","arxiv_id":"2503.21055","aliases":["library:VG04","lineage:STATE"],"technical_key":"arxiv-2503.21055","catalog_id":"X237","library_id":"VG04","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"其他（说明）","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2503.21055v6","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"L40S","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"State-Change Counterfactuals representation pretraining；Ego4D，7 epochs；每 GPU batch 18","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2503.21055v6","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"视频编码器＋语言状态描述；checkpoint 待核","training_datasets":"Ego4D 预训练；GTEA、EgoPER、EPIC-KITCHENS、Charades-Ego 等评测","existing_method":"从带叙述的 Ego4D 视频构造状态与反事实文本，使用帧／片段／视频层次对比学习；文本编码器冻结，另有缺步／换序等视频负例。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2c6e0104472c5be0833bdaed77b250a277ec57bb2cb24aeb8872ba01729a73d9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0238","title":"CounterVid: Counterfactual Video Generation for Mitigating Action and Temporal Hallucinations in Video-Language Models","year":"2026","venue":"未核实","arxiv_id":"2601.04778","aliases":["library:VG09"],"technical_key":"arxiv-2601.04778","catalog_id":"X238","library_id":"VG09","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"其他（说明）","directions":["XR-C","XR-L"],"source_url":"https://arxiv.org/html/2601.04778","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"CounterVid 下游 VLM LoRA-DPO；rank 64，3 epochs；数据合成另需约 3,000 GPU-hours 和 API，不混入 fine-tuning","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2601.04778","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen2.5-VL-3B／7B、InternVL3-9B（被训练理解模型）；生成器另计","training_datasets":"CounterVid preference pairs；VideoHallucer／VidHalluc／EventHallusion／VideoHallu 为评测","existing_method":"生成与原场景接近但动作／时序不同的视频负例，以文本和视觉偏好对及 MixDPO 训练视频语言理解模型。","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"1e70f6839115336be0653d4d6bfde6aca05c7ca4b999cb5ec2446d1c665a99c3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0239","title":"VQQA: An Agentic Approach for Video Evaluation and Quality Improvement","year":"2026","venue":"未核实","arxiv_id":"2603.12310","aliases":["library:VG11"],"technical_key":"arxiv-2603.12310","catalog_id":"X239","library_id":"VG11","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"文本或图像生成视频","directions":["AIGC-A","AIGC-C","XR-S"],"source_url":"https://arxiv.org/html/2603.12310","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"VQQA；多种 generator；展示配置 CogVideoX-5B＋Gemini-3-Pro","training_datasets":"无新增权重训练；视频质量 benchmark 用于评测","existing_method":"需可重复调用生成器与 VLM 评判器；已有针对原始提示的 Global Selection 和动态停止，以控制语义漂移，不应描述为完全没有全局保持。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f2a0bae721544f110c8d9bb54db2a1c2ca8cdf9d3be5c328bd37e09fd84f8994","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0240","title":"CineForge: Self-Improving Agents for Long-Horizon Video Generation","year":2026,"venue":"arXiv 2026","arxiv_id":"2608.29621","aliases":["library:VG19","lineage:CINE","recent:CINE"],"technical_key":"arxiv-2608.29621","catalog_id":"X240","library_id":"VG19","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"长视频与多镜头叙事","directions":["XR-E"],"source_url":"https://arxiv.org/html/2608.29621v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CineForge-Produce／Evolve；具体工具版本待核","training_datasets":"CineScope-Data 100 scripts 用于策略演化；AnimeShooter、ViStoryBench 评测","existing_method":"相较只修当前视频或存自由文本反思，它把错误追溯到生产阶段并学习可版本化的持续改进。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0d790d960c6074c4b2bfabe7837ca139bdd0a00c71dd2a8b7ca2a9a18ce61c50","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0241","title":"VideoGen-Agent: Reinforcing Video Generation Agents","year":2026,"venue":"arXiv 2026","arxiv_id":"2609.24997","aliases":["library:VG20","lineage:VGEN","recent:VGEN"],"technical_key":"arxiv-2609.24997","catalog_id":"X241","library_id":"VG20","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"条件控制生成","directions":["AIGC-D","XR-L"],"source_url":"https://arxiv.org/html/2609.24997v3","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":141.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VideoGen-Agent Qwen3-VL-8B SFT：16K trajectories，2 epochs；RL 8K prompts 1 epoch；外部生成工具开销另计","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2609.24997v3","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen3-VL-8B-Instruct；可替换视频／验证工具","training_datasets":"16K teacher trajectories SFT；8K prompts RL；VABench 600 prompts 评测","existing_method":"相较固定工作流，它学习根据提示和中间结果选择工具，覆盖知识、身份、物理与多镜头需求。","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"16588f9723b655f0c48fce14349984672fd0df2cb2960c6315a5a5973798616b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0242","title":"MaLiang-Harness: A Programmable Path to Image and Video Generation","year":"2026","venue":"未核实","arxiv_id":"2609.34309","aliases":["library:VG22"],"technical_key":"arxiv-2609.34309","catalog_id":"X242","library_id":"VG22","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"其他（说明）","directions":["AIGC-B","XR-S"],"source_url":"https://arxiv.org/html/2609.34309","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"多种 MLLM 代码生成器；含 GPT-6-Astra 等；不是新 diffusion 模型","training_datasets":"无新增生成器权重训练；MaLiang-IBench／VBench 用于评测","existing_method":"相较代码能运行就算成功，它专门检查可执行程序是否真的产生符合要求的图像或视频。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f910062aed2fd0f857d4f8e5b9176613bacb04f7923a06ff249aef7de738a382","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0243","title":"ParaVT: Taming the Tool Prior Paradox for Parallel Tool Use in Agentic Video Reinforcement Learning","year":"2026","venue":"未核实","arxiv_id":"2605.20342","aliases":["library:VG58"],"technical_key":"arxiv-2605.20342","catalog_id":"X243","library_id":"VG58","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"其他（说明）","directions":["XR-L"],"source_url":"https://arxiv.org/html/2605.20342","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":7.0,"count_max":7.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ParaVT RL 的参数更新用 7 GPUs，另 1 GPU rollout；附录总机器数与单轮分配需区分；80GB 为原文下限","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.20342","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen3-VL-8B-Instruct","training_datasets":"97K multi-task SFT；独立 4,406 样本 RL；各来源待核","existing_method":"相较串行裁剪的错误累积和成本，它研究工具先验如何同时帮助探索与诱发格式崩溃或跳工具捷径。","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":7.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"c065e45a18b8bc9abc8a673c069c005e3419ab7e8c3051865097783d707b07b6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0244","title":"Understanding Data Influence in Reinforcement Finetuning（RFT-Inf）","year":"2025","venue":"未核实","arxiv_id":"","aliases":["library:VG61"],"technical_key":"library-VG61","catalog_id":"X244","library_id":"VG61","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"其他（说明）","directions":["XR-L"],"source_url":"https://proceedings.neurips.cc/paper_files/paper/2025/file/3755a02b1035fbadd5f93a022170e46f-Paper-Conference.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"RFT-Inf 实验集群 2 servers × 8；具体各训练 run 是否占满集群不明确，不纳入 GPU 数分档","comparable":false,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://proceedings.neurips.cc/paper_files/paper/2025/file/3755a02b1035fbadd5f93a022170e46f-Paper-Conference.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"DeepSeek-R1-Distill-Qwen-1.5B／7B；Llama-3.2-3B-Instruct","training_datasets":"DeepScaleR 约 40K 数学 QA（来源含 Omni-MATH／STILL）；AIME24、MATH-500、AMC23、Minerva、OlympiadBench 评测","existing_method":"相较只按难度或奖励筛数据，它以样本梯度与总体更新方向的关系估计训练价值。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"8b41e4ca78ba0a2dd743175ccbafa3e42e861c5a69ca45b157d2a71200d446ff","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0245","title":"Aero Realtime: Fully Aligned Input-Output Streams for Low-Latency Streaming Multimodal Generation","year":"2026","venue":"未核实","arxiv_id":"2608.08469","aliases":["library:VG68"],"technical_key":"arxiv-2608.08469","catalog_id":"X245","library_id":"VG68","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"其他（说明）","directions":["XR-S"],"source_url":"https://arxiv.org/html/2608.08469","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Aero Realtime 训练 GPU A100-40G；只定位到吞吐，没有设备总数和完整耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2608.08469","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Aero Realtime 4B","training_datasets":"不确定：本轮尚未核实","existing_method":"相较轮流预填—解码或微轮询，它以双工架构联合学习回应内容和时机。","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"78cf9c0d51d1da904bc7b66ba71d6aab3ab41976985851eb39ec48f75bc0e58a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0246","title":"Procedure Planning in Instructional Videos","year":2020,"venue":"ECCV 2020","arxiv_id":"1907.01172","aliases":["lineage:PROC"],"technical_key":"arxiv-1907.01172","catalog_id":"X246","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/1907.01172v3","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"GTX 1080 Ti","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Procedure Planning 新规划模型训练 200 epochs；batch 256；不是视频生成器 fine-tuning","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/1907.01172v3","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Dual Dynamics Network","training_datasets":"CrossTask","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2e684609b151e6187a435692bfab903ad267387504a29de981b6c363aed262c9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0247","title":"SCHEMA: State CHangEs MAtter for Procedure Planning in Instructional Videos","year":2024,"venue":"ICLR 2024","arxiv_id":"2403.01599","aliases":["lineage:SCHEMA","external:XR-REF-SCHEMA"],"technical_key":"arxiv-2403.01599","catalog_id":"X247","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2403.01599v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"V100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":1.0,"wall_hours_max":1.0,"time_bound":"approximate","gpu_hours_min":1.0,"gpu_hours_max":1.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"SCHEMA CrossTask 500 epochs；训练规划头；冻结预训练表征；不是像素视频模型训练","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2403.01599v1","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"pre","device":"V100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":5.5,"wall_hours_max":5.5,"time_bound":"approximate","gpu_hours_min":5.5,"gpu_hours_max":5.5,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"SCHEMA COIN 400 epochs；与 CrossTask 为不同实验","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2403.01599v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"双 decoder procedure planner；GPT-3.5 状态描述","training_datasets":"CrossTask、COIN、NIV","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a3b5c6a989a4610f23a64cea77903c814c009748c390587a6436aca5a6691b59","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0248","title":"Deep Structural Causal Models for Tractable Counterfactual Inference","year":2020,"venue":"NeurIPS 2020","arxiv_id":"2006.06485","aliases":["lineage:DSCM","external:XR-REF-DSCM"],"technical_key":"arxiv-2006.06485","catalog_id":"X248","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2006.06485v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Deep Structural Causal Models","training_datasets":"Morpho-MNIST；UK Biobank brain MRI","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"34360acf61efcac9e05e3762ffe851e3917e0a3d1522453d102a6bb979e48a9e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0249","title":"Generative Omnimatte: Learning to Decompose Video into Layers","year":2025,"venue":"CVPR 2025","arxiv_id":"2411.16683","aliases":["lineage:OMNI"],"technical_key":"arxiv-2411.16683","catalog_id":"X249","library_id":"","topic_code":"","topic":"","task":"编辑与修改","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2411.16683v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Casper；CogVideoX video-inpainting prior","training_datasets":"合成分层／对象去除训练数据；完整来源待核","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"98ef8c332f400e09c0874c5f881f18b5720e65f1557dffebe69f38f78ffd6239","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0250","title":"VideoPhysEdit: Physical Counterfactual Video Editing via Rigid-Body Physical Scene Reconstruction","year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","arxiv_id":"2609.35134","aliases":["lineage:PHYS","external:EXT-PHYS"],"technical_key":"arxiv-2609.35134","catalog_id":"X250","library_id":"","topic_code":"","topic":"","task":"编辑与修改","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2609.35134v1","status":"无需本文权重训练","resource_note":"VideoPhysEdit 明确提出 training-free 的物理场景重建、模拟与视频编辑 pipeline。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"物理场景重建＋刚体模拟＋轨迹引导生成器（版本待核）","training_datasets":"无新增权重训练；PCVE-RigidBench 20 场景／129 任务为评测","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"7df3732651b9a5735b2c80d66a4cf89ddcbccc940da1cb18ac0384e12367e093","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0251","title":"Robots That Ask For Help: Uncertainty Alignment for Large Language Model Planners","year":2023,"venue":"CoRL 2023","arxiv_id":"2307.01928","aliases":["lineage:KNOW"],"technical_key":"arxiv-2307.01928","catalog_id":"X251","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-D","XR-L"],"source_url":"https://arxiv.org/html/2307.01928v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"KnowNo；LLM planner","training_datasets":"无 LLM 微调；使用 calibration set（不是零数据）","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"17bcede566e866b72379b8f5c8c3928723cc16b6ad187ef4f685874996ee3552","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0252","title":"Generative World Explorer","year":2025,"venue":"ICLR 2025","arxiv_id":"2411.11844","aliases":["lineage:GENEX","external:EXT-GENEX"],"technical_key":"arxiv-2411.11844","catalog_id":"X252","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":["AIGC-B","XR-S"],"source_url":"https://arxiv.org/html/2411.11844v3","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"GenEx-Diffuser 训练总 384 A100 GPU-hours；2 A100 / batch 是 model-parallel 配置，不等于整次 run 只有 2 GPUs","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2411.11844v3","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":384,"gpu_hours_max":384,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"GenEx-Diffuser 全部训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2411.11844v3","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"GenEx world explorer；具体 video checkpoint 待核","training_datasets":"GenEx-DB 合成城市场景；GenEx-EQA 为评测","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"a4fb5719f5e75f5f2268bfbce59fa6d8e9585208e377f55726ab75005d939b3d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0253","title":"StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text","year":2025,"venue":"CVPR 2025","arxiv_id":"2403.14773","aliases":["lineage:STREAM"],"technical_key":"arxiv-2403.14773","catalog_id":"X253","library_id":"","topic_code":"","topic":"","task":"长视频与多镜头叙事","directions":["AIGC-B","XR-S"],"source_url":"https://arxiv.org/html/2403.14773v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"StreamingT2V；CAM＋APM＋video enhancer","training_datasets":"视频训练来源待核；50 条 prompts 用于评价","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f7a422eb59db6e64cbf92f459e6b5a8421bc6facd86246338ea1baef88ba4193","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0254","title":"Weave Forcing: Compositional Memory Routing for Interactive Long Video Generation","year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","arxiv_id":"2610.03510","aliases":["lineage:WEAVE"],"technical_key":"arxiv-2610.03510","catalog_id":"X254","library_id":"","topic_code":"","topic":"","task":"长视频与多镜头叙事","directions":["AIGC-B","XR-S"],"source_url":"https://arxiv.org/html/2610.03510v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"inf","device":"RTX PRO 6000","count_min":1.0,"count_max":1.0,"card_memory_gb":96.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Weave Forcing memory routing 实验；不要把部署显存当成训练显存","comparable":false,"is_gpu":true,"role_label":"推理／评测／rollout","source_url":"https://arxiv.org/html/2610.03510v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Self Forcing／Wan2.1-T2V-1.3B；Gemini 2.5 Pro","training_datasets":"无新增权重训练","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e3393b50447f2759407ba6a9be96e75e26a71015c3832be39452362193e16c45","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0255","title":"VideoAgent: Self-Improving Video Generation","year":2024,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","arxiv_id":"2410.10076","aliases":["lineage:VIDEOAG"],"technical_key":"arxiv-2410.10076","catalog_id":"X255","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":["XR-E"],"source_url":"https://arxiv.org/html/2410.10076v3","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"VideoAgent；AVDC 类视觉规划器","training_datasets":"MetaWorld、iTHOR 训练／交互；各 split 需复核","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"086719e2af07ee984f2abf6bb9a9bd2c45875a2aa557d1de3701b5b73e772731","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0256","title":"VideoDirectorGPT: Consistent Multi-scene Video Generation via LLM-Guided Planning","year":2024,"venue":"COLM 2024","arxiv_id":"2309.15091","aliases":["lineage:DIRECTOR"],"technical_key":"arxiv-2309.15091","catalog_id":"X256","library_id":"","topic_code":"","topic":"","task":"长视频与多镜头叙事","directions":["AIGC-C"],"source_url":"https://arxiv.org/html/2309.15091v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A6000","count_min":8.0,"count_max":8.0,"card_memory_gb":48.0,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":384.0,"gpu_hours_max":384.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"VideoDirectorGPT 的 Layout2Vid grounding MLP + gated attention；0.64M images，50K steps","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2309.15091v2","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"GPT-4 Video Planner；Layout2Vid","training_datasets":"image-level layout annotations；具体数据源待核","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"8910666330c50eba8add68be205c48d878fdf16704db76b332614f1ce8c296b8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0257","title":"VIS-Ground: Video Interactive Storytelling with Contextual Grounding","year":2026,"venue":"arXiv 2026","arxiv_id":"2610.09326","aliases":["lineage:VIS","recent:VIS","external:EXT-VIS"],"technical_key":"arxiv-2610.09326","catalog_id":"X257","library_id":"","topic_code":"","topic":"","task":"长视频与多镜头叙事","directions":["AIGC-B","AIGC-C","XR-S"],"source_url":"https://arxiv.org/html/2610.09326v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"上下文编译／校验 agent；三个 video backbones（配置待核）","training_datasets":"是否有新增权重训练待核；narrative／knowledge grounded continuation 评测","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b9e5e39969c5d78b2eed2e7532337cdee8bd12f406a920e457689d3104300c8a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0258","title":"Multi-Grid Post-Training for Long-Form Multi-Shot Video Generation","year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","arxiv_id":"2609.06373","aliases":["lineage:GRID"],"technical_key":"arxiv-2609.06373","catalog_id":"X258","library_id":"","topic_code":"","topic":"","task":"长视频与多镜头叙事","directions":["AIGC-C"],"source_url":"https://arxiv.org/html/2609.06373v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"B200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"MovieGrid LoRA rank 32 + Grid Embedding；63.1M trainable params；10 epochs；2560×1536 grid video","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2609.06373v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-5B；Qwen3-VL-8B 辅助标注","training_datasets":"MGLV：1,000 长视频构造约 54K grid videos；89 stories 为评测","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"3e3677fe6c3bd6889c648952a8ad7ed692c9e71e7a8924b658a4a2c3b2a4c93f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0259","title":"Dramatis: A Computational Model of Suspense","year":2014,"venue":"AAAI 2014","arxiv_id":"","aliases":["lineage:DRAM"],"technical_key":"lineage-DRAM","catalog_id":"X259","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-C"],"source_url":"https://ojs.aaai.org/index.php/AAAI/article/view/8836/8695","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Dramatis；STRIPS／MEI-P","training_datasets":"无神经网络训练；三对文本故事与人评","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"dd01d5b5f48976e3c4b2ba4904f60c2e51a9efb7ec7a80ed0c338f5b5c3c2a54","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0260","title":"Creating Suspenseful Stories: Iterative Planning with Large Language Models","year":2024,"venue":"EACL 2024","arxiv_id":"2402.17119","aliases":["lineage:SUSP","recent:SUSP"],"technical_key":"arxiv-2402.17119","catalog_id":"X260","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-C"],"source_url":"https://arxiv.org/html/2402.17119v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"gpt-3.5-turbo-0613","training_datasets":"无额外训练语料；90 人评价 30 对故事","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3d8f1c0038f25f3d3fa0f2ef9fd6f62181ddff240bcbeffcd11331c55b980160","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0261","title":"Adding Conditional Control to Text-to-Image Diffusion Models","year":2023,"venue":"ICCV 2023","arxiv_id":"2302.05543","aliases":["lineage:CONTROL"],"technical_key":"arxiv-2302.05543","catalog_id":"X261","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-D"],"source_url":"https://arxiv.org/html/2302.05543v3","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 3090 Ti","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":120.0,"wall_hours_max":120.0,"time_bound":"approximate","gpu_hours_min":120.0,"gpu_hours_max":120.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"ControlNet depth-to-image 小数据配置；200K images，5 天；是图像条件控制，不能直接外推长视频","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2302.05543v3","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Stable Diffusion＋ControlNet","training_datasets":"条件图／图像训练对；具体分支数据集待核","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"d7bae580fad1db9fa57cf5386a185a32f36f0db7e8d2c0e56328e0545d287c0a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0262","title":"SketchDynamics: Exploring Free-Form Sketches for Dynamic Intent Expression in Animation Generation","year":2026,"venue":"CHI 2026","arxiv_id":"2601.20622","aliases":["lineage:SKETCH"],"technical_key":"arxiv-2601.20622","catalog_id":"X262","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":["AIGC-D"],"source_url":"https://arxiv.org/html/2601.20622v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"VLM＋Manim；VLM 版本待核","training_datasets":"无新增权重训练；三阶段共 24 人研究","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e16569b07fa051c74b7b9b575efeda136bbde5e4fc9ffe259c69147d6b479b28","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0263","title":"Plug-and-Play Clarifier: A Zero-Shot Multimodal Framework for Egocentric Intent Disambiguation","year":2026,"venue":"AAAI 2026","arxiv_id":"2511.08971","aliases":["recent:CLARIFIER"],"technical_key":"arxiv-2511.08971","catalog_id":"X263","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-D","XR-L"],"source_url":"https://arxiv.org/html/2511.08971","status":"已核对资源阶段","resource_note":"Plug-and-Play Clarifier 为 zero-shot 框架，采用预训练模块与规则处理歧义。","stages":[{"role":"inf","device":"RTX 4090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Plug-and-Play Clarifier zero-shot 几何与多模态澄清；另有 LLM/VLM 推理延迟","comparable":false,"is_gpu":true,"role_label":"推理／评测／rollout","source_url":"https://arxiv.org/html/2511.08971","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen／Llama／GPT-4o／Gemini 等；视觉指向辅助模块","training_datasets":"无新增权重训练；IN3、CLAMBER、VRA-Ego 1,000 样本用于评测","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"87db8d106ae04f72252f027b07f3d733adb14b958138bf51d5ba4841f3f2e885","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0264","title":"Asking What Matters: Reward-Driven Clarification for Software Engineering Tasks","year":2026,"venue":"ICML 2026","arxiv_id":"","aliases":["recent:CLARITI"],"technical_key":"recent-CLARITI","catalog_id":"X264","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-D","XR-L"],"source_url":"https://raw.githubusercontent.com/mlresearch/v306/main/assets/vijayvargiya26a/vijayvargiya26a.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen3-8B clarifier；Qwen3-32B reward；Seed OSS 36B executor","training_datasets":"SWE-bench 派生缺信息任务；训练划分完整数量待核","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"21c78a87bb077599898ec7767f58b72568cae145f34e46b09d5316189b07765a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0265","title":"Uncertain Pointer: Situated Feedforward Visualizations for Ambiguity-Aware AR Target Selection","year":2026,"venue":"CHI 2026","arxiv_id":"2602.13433","aliases":["recent:POINTER"],"technical_key":"arxiv-2602.13433","catalog_id":"X265","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["XR-L"],"source_url":"https://arxiv.org/html/2602.13433","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Unity／Cinemachine；Gaussian splat 场景","training_datasets":"无学习模型训练；两项在线研究 N=60、N=40","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a2fb16442463b0a6390c16835d7f412e8f412b42bc33597449c6c2c5a39a0e1c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0266","title":"Uncertainty-Aware Clarification in LLM Agents with Information Gain","year":2026,"venue":"ICML 2026","arxiv_id":"","aliases":["recent:IGCLAR"],"technical_key":"recent-IGCLAR","catalog_id":"X266","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["XR-L"],"source_url":"https://raw.githubusercontent.com/mlresearch/v306/main/assets/deng26l/deng26l.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX A6000","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":7.0,"wall_hours_max":12.0,"time_bound":"approximate","gpu_hours_min":28.0,"gpu_hours_max":48.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Information-Gain Clarifier 每次 DAPO/PPO 训练；四卡总 28–48 GPU-hours；不是 XR 视频生成器训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://raw.githubusercontent.com/mlresearch/v306/main/assets/deng26l/deng26l.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.19（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen3-1.7B clarifier；另测 Qwen2.5-1.5B／Qwen3-4B","training_datasets":"扩展 τ-Bench；strict user simulator 产生训练交互","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"e7f4aba4335296d6ff29d35a7717c938616d0d8f660e549f70c211b9d6bb16dd","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0267","title":"SC²-WM: A Self-Correcting World Model with Closed-Loop Feedback for Vision-and-Language Navigation in Continuous Environments","year":2026,"venue":"ICML 2026","arxiv_id":"","aliases":["recent:SC2"],"technical_key":"recent-SC2","catalog_id":"X267","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":["AIGC-B","XR-S"],"source_url":"https://raw.githubusercontent.com/mlresearch/v306/main/assets/yao26u/yao26u.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 3090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":40.0,"wall_hours_max":40.0,"time_bound":"upper","gpu_hours_min":40.0,"gpu_hours_max":40.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"SC²-WM 闭环导航模型训练 <40h，为上界；不是开放域像素视频世界模型","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://raw.githubusercontent.com/mlresearch/v306/main/assets/yao26u/yao26u.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.7（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"VLN-3DFF／3D Feature Fields；panoramic 版本另见原文","training_datasets":"R2R-CE、RxR-CE；具体训练／测试 split 见正文","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"85142f0d19f786ee1dac6e246471c462e974b23c6bc3718dd576cc43c24d92a3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0268","title":"Spatia: Video Generation with Updatable Spatial Memory","year":2026,"venue":"CVPR 2026","arxiv_id":"2512.15716","aliases":["recent:SPATIA"],"technical_key":"arxiv-2512.15716","catalog_id":"X268","library_id":"","topic_code":"","topic":"","task":"3D/4D 一致生成","directions":["AIGC-B","XR-S"],"source_url":"https://arxiv.org/html/2512.15716","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"AMD MI250","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Spatia ControlNet + rank-64 LoRA 两阶段；LoRA 5K iterations；batch 64；不换算 NVIDIA 等效时长","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2512.15716","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-5B；ControlNet blocks；SLAM","training_datasets":"RealEstate 40K train videos＋SpatialVID HD 10K videos","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"3eeae0f4362cc3229b88e42bb215c473d06d4772d02d90c8ca9529c2a9725846","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0269","title":"Composition of Memory Experts for Diffusion World Models","year":2026,"venue":"ICLR 2026","arxiv_id":"","aliases":["recent:COME"],"technical_key":"recent-COME","catalog_id":"X269","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":["AIGC-B","XR-S"],"source_url":"https://proceedings.iclr.cc/paper_files/paper/2026/file/19eae75beed66321d62272e794a9c2ac-Paper-Conference.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Memory Maze DiT；DMLab UViT；Memory-Cards U-Net；NavigationWM","training_datasets":"Memory Maze 30K×1,000 帧轨迹；DMLab、Memory-Cards、RECON／NavigationWM 各配置","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e2a1b37301ce299b239198d5e4d824bd46a7c8255c0b64237cac0a661716f0e7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0270","title":"Causal-Adapter: Taming Text-to-Image Diffusion for Faithful Counterfactual Generation","year":2026,"venue":"ICML 2026","arxiv_id":"2509.24798","aliases":["recent:CAUSAL"],"technical_key":"arxiv-2509.24798","catalog_id":"X270","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2509.24798","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":24.0,"wall_hours_max":24.0,"time_bound":"upper","gpu_hours_min":24.0,"gpu_hours_max":24.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Causal-Adapter 训练一天以内，为上界；图像 counterfactual adapter，不是长视频生成器","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2509.24798","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Stable Diffusion v1.5＋Causal-Adapter","training_datasets":"Pendulum、CelebA、ADNI、CelebA-HQ","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"432d8ec98034baca1d3d6c047839a44d4f5d054b89a5c2ca13c1538edaecce89","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0271","title":"Multiverse Mechanica: A Testbed for Learning Game Mechanics via Counterfactual Worlds","year":2026,"venue":"ICLR 2026","arxiv_id":"","aliases":["recent:MULTI"],"technical_key":"recent-MULTI","catalog_id":"X271","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-A","XR-C"],"source_url":"https://proceedings.iclr.cc/paper_files/paper/2026/file/f9bc7653cf176075e88203d1e83e8c65-Paper-Conference.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Stable Diffusion 变体；具体版本未在所读段落命名","training_datasets":"Multiverse Mechanica 游戏机制数据；peak-action snapshots 微调","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f1bc0d2ac103e8c5165aaafe39bc0a44037a0b9ed728d440503aa0fcdfddd848","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0272","title":"ARify: Leveraging Narrated Instructional Videos to Create Augmented Reality Tutorials for Procedural Tasks","year":2026,"venue":"CHI 2026","arxiv_id":"","aliases":["recent:ARIFY"],"technical_key":"recent-ARIFY","catalog_id":"X272","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["XR-E"],"source_url":"https://engineering.purdue.edu/cdesign/wp/arify/","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0273","title":"AgentCoach: LLM-Based Adaptive Coaching Feedback for Motor Skill Learning","year":2026,"venue":"CHI 2026","arxiv_id":"","aliases":["recent:COACH"],"technical_key":"recent-COACH","catalog_id":"X273","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["XR-E"],"source_url":"https://engineering.purdue.edu/cdesign/wp/agentcoach/","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"不确定：本轮尚未核实","training_datasets":"不确定：本轮尚未核实","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0274","title":"Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models","year":2026,"venue":"ICLR 2026","arxiv_id":"2510.04618","aliases":["recent:ACE"],"technical_key":"arxiv-2510.04618","catalog_id":"X274","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["XR-E"],"source_url":"https://arxiv.org/html/2510.04618","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"DeepSeek-V3.1（Generator／Reflector／Curator）","training_datasets":"AppWorld、FiNER、Formula 等任务的适应数据；test split 独立","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"82067011a2d555dbf63185f41f2265e4adee4660abd5b97f2d6c752830afbfeb","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0275","title":"Meta Context Engineering via Agentic Skill Evolution","year":2026,"venue":"ICML 2026","arxiv_id":"2601.21557","aliases":["recent:MCE"],"technical_key":"arxiv-2601.21557","catalog_id":"X275","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["XR-E"],"source_url":"https://raw.githubusercontent.com/mlresearch/v306/main/assets/ye26g/ye26g.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"DeepSeek-V3.1；MiniMax-M2.1 agent；AEGIS2 用 Qwen3-8B","training_datasets":"五域任务的 train／val 子集：含化学、Symptom2Disease、LawBench、AEGIS2；完整清单待核","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ef0713442395ececea6e209462acdf8146f2719557c9416d1acec2d10bea5eef","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0276","title":"Darwinian Memory: A Training-Free Self-Regulating Memory System for GUI Agent Evolution","year":2026,"venue":"ICML 2026","arxiv_id":"2601.22528","aliases":["recent:DARWIN"],"technical_key":"arxiv-2601.22528","catalog_id":"X276","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["XR-E"],"source_url":"https://raw.githubusercontent.com/mlresearch/v306/main/assets/mi26b/mi26b.pdf","status":"无需本文权重训练","resource_note":"Darwinian Memory 明确为 training-free 的外部记忆演化；8 A100 是部署服务器。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen2.5-VL-72B、Qwen3-VL-30B-A3B、GLM-4.5V、Seed1.6-VL","training_datasets":"AndroidWorld 在线任务经验；非离线参数训练集","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f5bd87e196a9c23e434090de58103e720cd9d4efffd3dfd9fcdec322026bdba0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0277","title":"SILVR: Self-Improving Loops for Visual Robotic Planning","year":2026,"venue":"ICLR 2026","arxiv_id":"2506.06658","aliases":["recent:SILVR"],"technical_key":"arxiv-2506.06658","catalog_id":"X277","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":["XR-E"],"source_url":"https://arxiv.org/html/2506.06658","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A6000 / RTX 3090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"SILVR in-domain 小 UNet；Panda Cup 8K steps / Drawer 10K；单卡配置仅对应 in-domain，不能泛化 internet-domain 模型","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2506.06658","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"AVDC 视频模型＋inverse dynamics model","training_datasets":"MetaWorld；真实 Panda cup-pushing 120 初始 demonstrations、drawer-opening 144；后续成功 rollouts","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"2cc0d1b3925674cbe98e38ceb84912e6d887f0177608850f1e0b55af00b44b93","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0278","title":"JarvisEvo: Towards a Self-Evolving Photo Editing Agent with Synergistic Editor-Evaluator Optimization","year":2026,"venue":"CVPR 2026","arxiv_id":"2511.23002","aliases":["recent:JARVIS"],"technical_key":"arxiv-2511.23002","catalog_id":"X278","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["XR-E"],"source_url":"https://arxiv.org/html/2511.23002","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"JarvisEvo editor/evaluator 与 reflection fine-tuning；反思数据 5K、1 epoch；总耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2511.23002","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Qwen3-VL-8B-Instruct；Lightroom 工具","training_datasets":"ArtEdit 170K：ArtEdit-Lr 120K／Eval 50K；SFT 用 110K＋40K，RL 子集另列原文","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"24ee03515b714a51457a627d064c4791b084f15e9df40d90fa6a9adf40cde8dc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0279","title":"VISTA: A Test-Time Self-Improving Video Generation Agent","year":2026,"venue":"CVPR 2026","arxiv_id":"2510.15831","aliases":["recent:VISTA","reasoning:CVPR-2026-3841"],"technical_key":"arxiv-2510.15831","catalog_id":"X279","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":["XR-E"],"source_url":"https://arxiv.org/html/2510.15831","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Veo-3.0-generate-preview；Veo2 对照","training_datasets":"无新增权重训练；benchmark 为评价","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ecd07fe9a90486d8cd55edbb5856a121b9e891ac21ba67c70647360dc9002a46","reasoning_id":"CVPR-2026-3841","reasoning_role":"R1","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0280","title":"Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents","year":2026,"venue":"ICLR 2026","arxiv_id":"2505.22954","aliases":["recent:DGM"],"technical_key":"arxiv-2505.22954","catalog_id":"X280","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["XR-E"],"source_url":"https://arxiv.org/html/2505.22954","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"coding agent；具体 FM 版本待核","training_datasets":"SWE-bench Verified／Polyglot 子集用于选择候选；有独立迁移评价","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c3d489188b5dc57fbda313f554c8057b9e98f3d7193dd8c79201920e3774b8c6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0281","title":"Generative Tutorial: Towards Live Contextualized Visual Instructions for Physical Tasks","year":2026,"venue":"arXiv 2026","arxiv_id":"2609.24955","aliases":["recent:TUTOR"],"technical_key":"arxiv-2609.24955","catalog_id":"X281","library_id":"","topic_code":"","topic":"","task":"条件控制生成","directions":["XR-E"],"source_url":"https://arxiv.org/html/2609.24955","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"图像／视频生成 API；具体型号待核","training_datasets":"无新增权重训练；45 unit actions／15 physical tasks；N=24 用户研究","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4b734b0745f47222289e1d9d4a2fbf1c0d22fe5706933fd7969cbb934ad13e5a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0282","title":"Twin Rollouts: Noise-Coupled Counterfactual Branching in Interactive Video World Models","year":"不确定","venue":"未核实","arxiv_id":"2608.08982","aliases":["external:EXT-TWIN"],"technical_key":"arxiv-2608.08982","catalog_id":"X282","library_id":"","topic_code":"","topic":"","task":"动作条件的预测与交互世界模型","directions":["AIGC-A"],"source_url":"https://arxiv.org/pdf/2608.08982","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"形式化框架；未给已完成实证模型","training_datasets":"无已完成训练实验；原文写 experiments forthcoming","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2a08810c77cf1aee77557cea0c8a78477b67f083bef8dbd048ea7d8d08abb22c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0283","title":"LayerRecall: A State-Conditioned Memory Router for Long-Horizon Consistency in Video Generation","year":"不确定","venue":"未核实","arxiv_id":"2608.28460","aliases":["external:EXT-LAYER"],"technical_key":"arxiv-2608.28460","catalog_id":"X283","library_id":"","topic_code":"","topic":"","task":"长视频与多镜头叙事","directions":["AIGC-B"],"source_url":"https://arxiv.org/html/2608.28460","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":8.0,"wall_hours_max":8.0,"time_bound":"approximate","gpu_hours_min":128.0,"gpu_hours_max":128.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"LayerRecall CHPM；1,600 cases，200 steps，1 epoch；原文 128 H100 GPU-hours，8h 为相除换算","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2608.28460","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"LongLive-2.0；Wan2.2-TI2V-5B","training_datasets":"384 latent-frame 长序列；原始视频来源待核","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"32e827db831aa67b2f975488093286155fb9d752a69ad8f37adaf47f7aa76c8a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0284","title":"ViMax: Agentic Video Generation","year":"不确定","venue":"未核实","arxiv_id":"2606.07649","aliases":["external:EXT-VIMAX"],"technical_key":"arxiv-2606.07649","catalog_id":"X284","library_id":"","topic_code":"","topic":"","task":"长视频与多镜头叙事","directions":["AIGC-C"],"source_url":"https://arxiv.org/html/2606.07649","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Gemini-3-Flash；视频工具配置待核","training_datasets":"是否新增训练待核；ViMax-Bench 35 stories 为评价","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"fbbd838d6753f41df928a271549917a1a4d4716bc3076c13cbfc76193cee4997","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0285","title":"SARA: Semantically Adaptive Relational Alignment for Video Diffusion Models","year":"不确定","venue":"未核实","arxiv_id":"2605.07800","aliases":["external:XR-REF-SARA"],"technical_key":"arxiv-2605.07800","catalog_id":"X285","library_id":"","topic_code":"","topic":"","task":"文本或图像生成视频","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2605.07800","status":"已核对资源阶段","resource_note":"","stages":[{"role":"aux","device":"GPU（型号未定位）","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"SARA stage 1 对齐模块训练表：3K steps，batch 64；stage 2 diffusion 训练需另列","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2605.07800","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2；V-JEPA、SAM 3.1、Qwen3-VL-Embedding","training_datasets":"实体 mask／caption 训练 aligner；视频 continual-training 数据来源待核","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3d93f8f09bbef442c701520492c4c387c70598633223f36b7b0363e968d0b6c4","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0286","title":"CITRIS: Causal Identifiability from Temporal Intervened Sequences","year":"不确定","venue":"未核实","arxiv_id":"2202.03169","aliases":["external:XR-REF-CITRIS"],"technical_key":"arxiv-2202.03169","catalog_id":"X286","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2202.03169","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"Titan RTX","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":48.0,"gpu_hours_max":48.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"CITRIS-VAE Temporal Causal3DIdent；约 2 天；因果表征模型从零训练","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2202.03169","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"pre","device":"Titan RTX","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":12.0,"wall_hours_max":12.0,"time_bound":"approximate","gpu_hours_min":12.0,"gpu_hours_max":12.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"CITRIS-VAE Interventional Pong；12h","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2202.03169","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"pre","device":"Titan RTX","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":6.0,"wall_hours_max":10.0,"time_bound":"approximate","gpu_hours_min":6.0,"gpu_hours_max":10.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"CITRIS-NF；6–10h；不同变体不相加","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2202.03169","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CITRIS-VAE／CITRIS-NF","training_datasets":"Temporal Causal3DIdent；其他环境清单待核","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a49f6b0d3778251735cdab435022730e099a2881beb5b886c4ca0e9d1f73944a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0287","title":"Causal Abstractions of Neural Networks","year":"不确定","venue":"未核实","arxiv_id":"2106.02997","aliases":["external:XR-REF-ABSTRACTION"],"technical_key":"arxiv-2106.02997","catalog_id":"X287","library_id":"","topic_code":"","topic":"","task":"其他（说明）","directions":["AIGC-A","XR-C"],"source_url":"https://arxiv.org/html/2106.02997","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 2080 Ti / GTX Titan X / Titan XP / Titan V","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":5.5,"wall_hours_max":5.5,"time_bound":"approximate","gpu_hours_min":5.5,"gpu_hours_max":5.5,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Causal Abstractions 的模型每个 grid-search run 平均 5.5h；11–12GB 卡；四 seeds 与搜索全预算另计","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2106.02997","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"BERT 12-layer；BiLSTM","training_datasets":"MQNLI：500K train／60K dev／10K test","existing_method":"","training_signal":"不确定：本轮尚未核实","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"dfcf9e23217642d852af0eab021c1c5f1b4e60f04a9d56abe07bc69cca273983","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0288","title":"HG-DAgger: Interactive Imitation Learning with Human Experts","year":"2019","venue":"","arxiv_id":"1810.02890","aliases":["library:P001"],"technical_key":"","catalog_id":"","library_id":"P001","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/1810.02890v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"迭代执行新手策略；专家接管时独占控制并收集示范，数据聚合后重训网络集成；用集成输出分歧及接管前的不确定性学习阈值。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0690aa3c4dece223b5ba95fcb44b7c77f787ba4af38b42b3ef4a0161b633583c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0289","title":"Abductive Commonsense Reasoning","year":"2020","venue":"","arxiv_id":"1908.05739","aliases":["library:P002"],"technical_key":"","catalog_id":"","library_id":"P002","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/1908.05739v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"众包构造观察及正负解释，使用对抗筛选降低捷径；比较 GPT/BERT 微调和不同观察依赖结构，进行人工与自动评价。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b9973d4135070d90abcee798880b7e047c3e60cb687e79128cb6c4460782d971","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0290","title":"WebGPT: Browser-assisted question-answering with human feedback","year":"2021 / 2022","venue":"","arxiv_id":"2112.09332","aliases":["library:P004"],"technical_key":"","catalog_id":"","library_id":"P004","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2112.09332v3","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"约 6,000 条人类示范做行为克隆，约 21,500 组偏好比较训练奖励模型；比较 PPO 与拒绝采样，主要最优配置采用行为克隆加奖励筛选。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"828efe0286a9031617aad3da6f86ca2ac3ce0386ec93319560f030ac51cf544e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0291","title":"There’s a Time and Place for Reasoning Beyond the Image","year":"2022","venue":"","arxiv_id":"","aliases":["library:P005"],"technical_key":"","catalog_id":"","library_id":"P005","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://github.com/zeyofu/TARA","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"自动抽取新闻时空标签，众包验证可推理子集；比较 CLIP、微调、WIT 监督及 segment-wise 表征，使用 Accuracy 和层级 Example-F1。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"af6495578d2a77755b1e2d07c44b9d7a9ca61db60ef7a075650a9695f4d54779","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0292","title":"Learning Iterative Reasoning through Energy Minimization","year":"2022","venue":"","arxiv_id":"2206.15448","aliases":["library:P006"],"technical_key":"","catalog_id":"","library_id":"P006","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2206.15448v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"Titan X","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":2.0,"wall_hours_max":2.0,"time_bound":"approximate","gpu_hours_min":2.0,"gpu_hours_max":2.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"IREM graphical algorithmic computation；10K iterations；batch 64；小型 energy model 从零训练","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2206.15448v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"学习输入与候选输出之间的能量景观；通过迭代梯度更新寻找低能量解，以回放缓冲和末步截断反传控制训练开销。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"146c26222a84d5ac5bcaa6a3feccdb008b11b4d321d79848e7af0605576278c4","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0293","title":"The Abduction of Sherlock Holmes: A Dataset for Visual Abductive Reasoning","year":"2022","venue":"","arxiv_id":"","aliases":["library:P007"],"technical_key":"","catalog_id":"","library_id":"P007","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://www.ecva.net/papers/eccv_2022/papers_ECCV/html/5263_ECCV_2022_paper.php","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":8.0,"count_max":8.0,"card_memory_gb":48.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Sherlock visual abductive reasoning 的 CLIP 适配实验；不同 backbone batch 64 / 200 / 512","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://www.ecva.net/papers/eccv_2022/papers_ECCV/html/5263_ECCV_2022_paper.php","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"自由观察式众包标注线索框和推断；训练图像、区域、文本兼容性评分器，比较 LXMERT、UNITER 和 CLIP，并加入多任务训练。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"ee62537dc7d4dca60a23c106ff5a6399cad5986c45c8124893aefc17b2dfcd98","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0294","title":"Visual Abductive Reasoning","year":"2022","venue":"","arxiv_id":"2203.14040","aliases":["library:P008"],"technical_key":"","catalog_id":"","library_id":"P008","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2203.14040v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 2080 Ti","count_min":2.0,"count_max":2.0,"card_memory_gb":11.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Visual Abductive Reasoning 描述生成模型；耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2203.14040v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"使用预提取动作/外观特征，contextualized directional position embedding 区分关系；三个级联 Transformer 解码器通过句子置信度交换信息。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":2.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"2ca333ede9b87ba4e15395c431d04e0536f4fce653cf5ac083dc476c38d2fd4d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0295","title":"Breaking Common Sense: WHOOPS! A Vision-and-Language Benchmark of Synthetic and Compositional Images","year":"2023","venue":"","arxiv_id":"2303.07274","aliases":["library:P009"],"technical_key":"","catalog_id":"","library_id":"P009","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2303.07274v4","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A6000","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":576.0,"gpu_hours_max":576.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"WHOOPS 全部 5 splits × 2 learning rates × 3 models 的实验合计；不是单模型单次训练","comparable":false,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2303.07274v4","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"设计师用文生图工具制作异常组合，众包标注解释；比较端到端 BLIP2 和图像描述→LLM 管线，并用人类及 GPT-4 评价解释。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"93f7190bf287550116c1645d3b6c58f955a941d42f349a2f2e805f572a44b4c0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0296","title":"Visual Writing Prompts: Character-Grounded Story Generation with Curated Image Sequences","year":"2023","venue":"","arxiv_id":"2301.08571","aliases":["library:P010"],"technical_key":"","catalog_id":"","library_id":"P010","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2301.08571v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"过滤低信息和重复镜头，提供角色名称/裁剪图引导写作；冻结 Swin 特征提取，结合角色特征与角色网格建模跨句连贯性。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"09bb1475d017db2221566b7135765fdd324cb8a8f560ebb3e5633cbdac119934","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0297","title":"Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks","year":"2024","venue":"","arxiv_id":"2307.02477","aliases":["library:P011"],"technical_key":"","catalog_id":"","library_id":"P011","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2307.02477v3","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"保持任务结构、改动进制/索引/坐标/游戏规则等条件，比较 GPT-4、GPT-3.5、Claude、PaLM-2，控制 CoT 与 few-shot 提示。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4c0d9c525af7c5ae7d8655c7bc3aff708e22c3c1caf91882ee57554eb7487df2","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0298","title":"Sequential Modeling Enables Scalable Learning for Large Vision Models","year":"2024 / 2023","venue":"","arxiv_id":"2312.00785","aliases":["library:P013"],"technical_key":"","catalog_id":"","library_id":"P013","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.00785v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"TPU v3-512","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"LVM-3B 最大模型约 14 天；原文为 v3-512 TPU pod，保留拓扑名，不换算 GPU","comparable":false,"is_gpu":false,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2312.00785v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"VQGAN 将图像离散化，再用自回归 Transformer 学习跨图像序列；在 420B 视觉 token 上训练 300M、600M、1B、3B 模型。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1e6071a1342d4f5e1d71908815fee2f28a9e77bee920f89acee87a32702d7d06","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0299","title":"Inductive or Deductive? Rethinking the Fundamental Reasoning Abilities of LLMs","year":"2024","venue":"","arxiv_id":"2408.00114","aliases":["library:P015"],"technical_key":"","catalog_id":"","library_id":"P015","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2408.00114v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"比较直接输入输出提示、显式给规则、由样例生成 Python 规则三种设置，测试算术、空间等默认/反事实任务。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ebec98d988e1c8bca39760604a9ca9a97ca358bbb6f2e2d1a99d7b8502cceb79","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0300","title":"DIVE: Towards Descriptive and Diverse Visual Commonsense Generation","year":"2023 / 2024","venue":"","arxiv_id":"2408.08021","aliases":["library:P016"],"technical_key":"","catalog_id":"","library_id":"P016","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2408.08021v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX A6000","count_min":6.0,"count_max":6.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":24.0,"wall_hours_max":24.0,"time_bound":"upper","gpu_hours_min":144.0,"gpu_hours_max":144.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"DIVE 全部训练流程一天内，为上界；不是 full pretrained backbone 的从零训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2408.08021v1","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.6（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"利用频次及图像语义集中度过滤通用推断，再用图结构构造对比检索监督；在原始、unique、novel 子集及人评中比较。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":6.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"b408a926f6cc3948d2bdb619c522a05caadfda54fc70c039de6a916dd09318dd","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0301","title":"Exploring Defeasibility in Causal Reasoning","year":"2024","venue":"","arxiv_id":"2401.03183","aliases":["library:P017"],"technical_key":"","catalog_id":"","library_id":"P017","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2401.03183v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"构造原因—结果及支持/反驳论据；以 token 级因果关联和注意力构建 CESAR，另测论据生成。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"bfd0498aadee7d46454d3ebacb75b16a99f22133cec6c3f51bf125e447b81a12","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0302","title":"Learning Iterative Reasoning through Energy Diffusion","year":"2024","venue":"","arxiv_id":"2406.11179","aliases":["library:P018"],"technical_key":"","catalog_id":"","library_id":"P018","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2406.11179v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 2080","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":2.0,"wall_hours_max":2.0,"time_bound":"approximate","gpu_hours_min":2.0,"gpu_hours_max":2.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Energy Diffusion 的 continuous tasks；50K iterations；discrete tasks 另有配置","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2406.11179v1","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.12（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"学习多噪声级能量函数，用对比监督塑造低能量正确解；推理时沿退火景观迭代梯度优化。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d0f9d018827fd05b2a0599d6f6315283648f7d0cf9addd9a253be7bdab7794c9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0303","title":"NL-EYE: Abductive NLI for Images","year":"2024","venue":"","arxiv_id":"2410.02613","aliases":["library:P019"],"technical_key":"","catalog_id":"","library_id":"P019","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2410.02613v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"人工撰写场景并迭代文生图；比较 triplet/pair 设置、合理性预测及解释，分析纯视觉和文本替代条件。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"60fef9172587cf57297497059fe65b11e54375882ee41c07ce4cee4a9a847f72","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0304","title":"ViCor: Bridging Visual Understanding and Commonsense Reasoning with Large Language Models","year":"2024","venue":"","arxiv_id":"2310.05872","aliases":["library:P020"],"technical_key":"","catalog_id":"","library_id":"P020","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2310.05872v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"先从图像描述评估候选；不确定时由 LLM 分类为 VCU/VCI，调用 BLIP2 直接匹配或询问关键视觉因素，再重新推理。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b9852e7226c846f4f81feb86a33a1de97fc43698527d0aae0a9441aaf19f5fb1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0305","title":"Generalizable Reasoning through Compositional Energy Minimization","year":"2025","venue":"","arxiv_id":"2510.20607","aliases":["library:P023"],"technical_key":"","catalog_id":"","library_id":"P023","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2510.20607v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A10","count_min":1.0,"count_max":1.0,"card_memory_gb":24.0,"peak_memory_gb":null,"wall_hours_min":5.0,"wall_hours_max":5.0,"time_bound":"approximate","gpu_hours_min":5.0,"gpu_hours_max":5.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Compositional Energy Minimization 的 N-queens / MIS 配置；20K epochs；batch 2048","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2510.20607v1","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.15（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"扩散与对比损失训练局部能量函数；求和形成全局目标，用多个并行候选进行能量最小化和筛选。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b716fc82ec73b5a619031c03ea15880e3585d8adf26c0f77bda08869c94ae764","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0306","title":"VCD: A Dataset for Visual Commonsense Discovery in Images","year":"2025","venue":"","arxiv_id":"2402.17213","aliases":["library:P027"],"technical_key":"","catalog_id":"","library_id":"P027","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2402.17213v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX A6000","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VCD 的 VCM baseline；3 epochs；未定位耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2402.17213v2","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.13（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"结合 Visual Genome 与 ConceptNet，映射/扩展三元组并进行质量控制，再用指令微调视觉语言模型生成常识。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"415ecb328511a2c8f91ff03afd54a5c2d3df61260836d838845bcd49a4d0b65d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0307","title":"Video models are zero-shot learners and reasoners","year":"2025","venue":"","arxiv_id":"2509.20328","aliases":["library:P028"],"technical_key":"","catalog_id":"","library_id":"P028","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2509.20328v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"输入图像及任务提示生成 8 秒视频，再按任务读取最好/最后一帧；比较 Veo 2、Veo 3 和图像/语言基线。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"412055f0054d109ec4514d436362f8ad462f562e1e42fc5357f607e615a8ce93","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0308","title":"VideoDiff: Human-AI Video Co-Creation with Alternatives","year":"2025","venue":"","arxiv_id":"","aliases":["library:P029"],"technical_key":"","catalog_id":"","library_id":"P029","topic_code":"H3","topic":"人机共创与可编辑工作流","task":"未归入视频任务轴","directions":[],"source_url":"https://doi.org/10.1145/3706598.3713417","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"8 名专业创作者形成性研究；实现 React/D3/Remotion 界面，Whisper 转录与 GPT-4o 编辑建议；12 人被试内对照和 3 个真实素材案例。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"63429576a022e7487d7a7de20abf4344ec42514b3ca2ea4f0cdbd03255aa3dc7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0309","title":"AbductiveMLLM: Boosting Visual Abductive Reasoning Within MLLMs","year":"2026","venue":"","arxiv_id":"2601.02771","aliases":["library:P030"],"technical_key":"","catalog_id":"","library_id":"P030","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2601.02771v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A800","count_min":4.0,"count_max":4.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"AbductiveMLLM stage I 10 epochs、stage II 1 epoch；Reasoner/Imaginer 适配","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2601.02771v1","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.5（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"LLM 根据视频描述提出候选，经因果相关对比学习筛选；条件扩散模块通过适配器融合视频和推理表示，联合优化语言与图像分支。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"b537c6d6ba4c65cb6154851f94f705be6dd8bcb0f94e81e6682c6171e75cd4c1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0310","title":"Agentic Reasoning for Large Language Models","year":"2026","venue":"","arxiv_id":"2601.12538","aliases":["library:P031"],"technical_key":"","catalog_id":"","library_id":"P031","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2601.12538v1","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"分别梳理规划/工具/搜索、反馈/记忆/适应、角色/分工/协作演化，区分上下文编排与 SFT/RL 后训练，并整理跨领域基准。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"425c0db72e31cbcf7536521b4a112eff20070e3458f3d39d1c8b546d64e950c1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0311","title":"Autoregressive, Yet Revisable: In Decoding Revision for Secure Code Generation","year":"2026","venue":"","arxiv_id":"2602.01187","aliases":["library:P032"],"technical_key":"","catalog_id":"","library_id":"P032","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.01187v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"In-Decoding Revision 所有 SFT runs；2 epochs；batch 2，accumulation 2","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2602.01187v2","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.16（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"从真实 CVE 修复对构造训练数据，学习编辑动作及普通代码；用安全检查通过率、HumanEval 功能测试和 AST 解析率分别验证。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"2f62f4c0e0fe1916cbf8f10995568804ecb0933d50a0e81a750fb056bea58080","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0312","title":"Lost in Transmission: When and Why LLMs Fail to Reason Globally","year":"2025 / 2026","venue":"","arxiv_id":"2505.08140","aliases":["library:P034"],"technical_key":"","catalog_id":"","library_id":"P034","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2505.08140v5","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"用 prefix/attention oracle 抽象信息流，区分 BAPO-easy/hard；在受控和现实任务上比较模型与 CoT 条件。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1ed0815beab98ee8fb04f7a2b3d60731a83dd799e5ff188d86748c3072a10adc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0313","title":"Mind-Brush: Integrating Agentic Cognitive Search and Reasoning into Image Generation","year":"2026","venue":"","arxiv_id":"2602.01756","aliases":["library:P035"],"technical_key":"","catalog_id":"","library_id":"P035","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.01756v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"检测认知缺口，搜索多模态证据、补充约束后生成；用 checklist-based strict accuracy（全部清单项通过）和 WISE/RISEBench 评价。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c9950f9e7b9c6927488e384a76704c50572bba3b20cc120fd7da770c529b70e0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0314","title":"SLM-MUX: Orchestrating Small Language Models for Reasoning","year":"2026","venue":"","arxiv_id":"2510.05077","aliases":["library:P037"],"technical_key":"","catalog_id":"","library_id":"P037","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2510.05077","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"每个模型多次采样，取自身众数并比较一致性；按 union accuracy 与冲突惩罚搜索组合，平局用验证准确率处理。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2a12155c518d720ffc54182580bed461c8356d77a30bc0f39e9e5b980c4de149","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0315","title":"“I’m Not Sure, But...”: Examining the Impact of Large Language Models’ Uncertainty Expression on User Reliance and Trust","year":"2024","venue":"","arxiv_id":"","aliases":["library:P038"],"technical_key":"","catalog_id":"","library_id":"P038","topic_code":"H2","topic":"人工核查、推理交互与依赖校准","task":"未归入视频任务轴","directions":[],"source_url":"https://doi.org/10.1145/3630106.3658941","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"随机分组，对照有/无 AI 与不同措辞；用行为、答案正确性、信心和自报信任评价，并区分确认性/探索性分析。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"96ed117228e01eec556ccb002278d33d375f2ac7f467034d1cf4d54922fe7f22","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0316","title":"Make-a-Story: Visual Memory Conditioned Consistent Story Generation","year":"2023","venue":"","arxiv_id":"","aliases":["library:P039"],"technical_key":"","catalog_id":"","library_id":"P039","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://github.com/ubc-vision/Make-A-Story","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"潜在扩散逐帧生成，当前句子对过去图像和文本记忆做软注意力，以解析 he/she/they 等角色或背景指代。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c15d06e20e900e011cf8d880b636b9d54c9fb658a201aaf9ee931847ee107b38","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0317","title":"Collaposer: Transforming Photo Collections into Visual Assets for Storytelling with Collages","year":"2026","venue":"","arxiv_id":"2601.18428","aliases":["library:P040"],"technical_key":"","catalog_id":"","library_id":"P040","topic_code":"H3","topic":"人机共创与可编辑工作流","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2601.18428v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"6 名专家形成性访谈；流水线预处理照片，LLM 选择中心及相关标签，再按语义层次布局；12 人与两种消融基线被试内比较。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3ece9fad7a355a3349ae51a34dc2a506b059dbd82a39b4e09ee09cfdb5fffee0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0318","title":"Learning to Complement Humans","year":"2020","venue":"","arxiv_id":"2005.00582","aliases":["library:P041"],"technical_key":"","catalog_id":"","library_id":"P041","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2005.00582v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"建模标签、人类响应及条件预测；同时优化分类和查询决策，将求助成本/非对称错误代价纳入期望损失。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5881c20eaab5e07dc4ad203375e235ec70a66023a1bf4b7408f0b625b0986580","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0319","title":"LLMs Struggle to Balance Reasoning and World Knowledge in Causal Narrative Understanding","year":"2026","venue":"","arxiv_id":"","aliases":["library:P042"],"technical_key":"","catalog_id":"","library_id":"P042","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://openreview.net/pdf/171146600ae3d44c55031ab5bcc202e937ad640e.pdf","status":"无需本文权重训练","resource_note":"该因果 narrative 分析使用预训练模型 API；正文说明没有预训练和本地 GPU 需求。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"从链、fork、collider 和 CauseNet 构造叙事，操控叙述顺序/知识一致性/长度；比较直接回答与先抽因果图。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"33dc9ae62b988717a592818d7cf941d73b529628a95e75cbd83d69775c69732b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0320","title":"LLMs Struggle to Perform Counterfactual Reasoning with Parametric Knowledge","year":"2025","venue":"","arxiv_id":"2506.15732","aliases":["library:P043"],"technical_key":"","catalog_id":"","library_id":"P043","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2506.15732","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A6000","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Counterfactual Parametric Knowledge toy experiment；LoRA rank 8，5 epochs；总 72 GPU-hours 是实验总预算","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2506.15732","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A6000","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":72,"gpu_hours_max":72,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"toy experiments 合计；不是单个 run","comparable":false,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2506.15732","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"在 GPT-4o/Llama3.1 与玩具模型上比较直接提示、CoT、160 条反事实示例微调及预训练干预。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"5838b05e3e342ac92e9b5e5fdfe32e8a25a377db3d4060685622ce09713fc94a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0321","title":"Orthogonal Causal Calibration","year":"2025","venue":"","arxiv_id":"2406.01933","aliases":["library:P044"],"technical_key":"","catalog_id":"","library_id":"P044","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2406.01933v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"构造广义伪结果或条件损失，在独立样本上校准；把误差分为 nuisance 估计误差与假想已知 nuisance 下的校准误差。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9e5544043d093b5dded395cdd822ae11a0c6305e0c9d4e5c55b1dd0d3bed92a7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0322","title":"Explaining Concept Shift with Interpretable Feature Attribution","year":"2025","venue":"","arxiv_id":"2505.20634","aliases":["library:P046"],"technical_key":"","catalog_id":"","library_id":"P046","topic_code":"M5","topic":"不确定性、决策损失与鲁棒优化","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2505.20634v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"在原模型输出上拟合带 L1 惩罚的加性修正，结合两域共享误设项；用 knockoffs 筛选发生变化的特征。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"27116734fef1210a64c1f65ed12ce9607e2e5faeba26810a4bf0a9cfed4c92f7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0323","title":"Can LLMs Propose Instrumental Variables for Causal Reasoning?","year":"2025","venue":"","arxiv_id":"","aliases":["library:P047"],"technical_key":"","catalog_id":"","library_id":"P047","topic_code":"M5","topic":"不确定性、决策损失与鲁棒优化","task":"未归入视频任务轴","directions":[],"source_url":"https://cispa.de/en/research/publications/104502-can-llms-propose-instrumental-variables-for-causal-reasoning-","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"先测试恢复文献工具和拒绝已被推翻的工具；代理提议处理—结果—工具组合并批判假设，再比较第一阶段强度与估计一致性。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"95f7a95e83e8dd8da02553e389cf6378f6052979d98c1ca40b63882104c19284","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0324","title":"Utility-Directed Conformal Prediction: A Decision-Aware Framework for Actionable Uncertainty Quantification","year":"2025","venue":"","arxiv_id":"","aliases":["library:P048"],"technical_key":"","catalog_id":"","library_id":"P048","topic_code":"M5","topic":"不确定性、决策损失与鲁棒优化","task":"未归入视频任务轴","directions":[],"source_url":"https://proceedings.iclr.cc/paper_files/paper/2025/hash/0c6b452f1bbfb6905f6bac957d73b321-Abstract-Conference.html","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"利用标签或集合成本构造惩罚比率/保形分数，再做校准；在分类层级中偏好临床或语义更一致的集合。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"cc79ee6bc7f8559759758b69bd639a586c4dcb30cde118233e7283722241f742","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0325","title":"Decision-Focused Learning without Differentiable Optimization: Learning Locally Optimized Decision Losses","year":"2022","venue":"","arxiv_id":"2203.16067","aliases":["library:P049"],"technical_key":"","catalog_id":"","library_id":"P049","topic_code":"M5","topic":"不确定性、决策损失与鲁棒优化","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2203.16067","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"在真实标签附近扰动预测，用决策质量作为目标，拟合加权 MSE/方向二次型等局部损失，再用于训练预测网络。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0bdfb5450404378ec5292f82f4e381928ef24de6c73caaeb4f2fd3e2e113c211","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0326","title":"Distributionally Robust Feature Selection","year":"2025","venue":"","arxiv_id":"","aliases":["library:P050"],"technical_key":"","catalog_id":"","library_id":"P050","topic_code":"M5","topic":"不确定性、决策损失与鲁棒优化","task":"未归入视频任务轴","directions":[],"source_url":"https://github.com/maitreyiswaroop/dro-feature-selection","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"对特征加可控噪声，估计各组目标并最小化最差组损失，用重参数化和核估计做梯度优化。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1dd085188ed1cc423e8507a1ef48f8ec9530d7249f770262775ce2a446169999","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0327","title":"Fostering the Ecosystem of AI for Social Impact Requires Expanding and Strengthening Evaluation Standards","year":"2025","venue":"","arxiv_id":"2510.18238","aliases":["library:P051"],"technical_key":"","catalog_id":"","library_id":"P051","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2510.18238","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"分析既有评审标准和应用研究案例，分别讨论试点、随机与非随机部署、反方意见。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"27fdc2a82043d8d1f15fc3d2d729b3319fde0110304f979814f74505f53fe3ee","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0328","title":"Valid Inference with Imperfect Synthetic Data","year":"2025","venue":"","arxiv_id":"","aliases":["library:P052"],"technical_key":"","catalog_id":"","library_id":"P052","topic_code":"M5","topic":"不确定性、决策损失与鲁棒优化","task":"未归入视频任务轴","directions":[],"source_url":"https://papers.nips.cc/paper_files/paper/2025/hash/ede6376c80b38796c688a77dce9fb227-Abstract-Conference.html","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"联合真实与合成矩残差，利用它们的预测相关性，估计权重和协方差构造参数与置信区间。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"837a0f89b17a54f100919a85d85b45bc8ed9936b29d9a397bb73f693ce93e16d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0329","title":"Superhuman performance of a large language model on the reasoning tasks of a physician","year":"2026","venue":"","arxiv_id":"","aliases":["library:P053"],"technical_key":"","catalog_id":"","library_id":"P053","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://doi.org/10.1126/science.adz4433","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"五类病例推理实验加急诊研究，专家使用 Bond/R-IDEA 等量表评分，与历史和医生基线比较；分别考察 o1-preview 与 o1。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f34b639572548a80f16ad690d9c05f93cf5ebe87f0d3edc396aef740618ac0e7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0330","title":"Melding the Data-Decisions Pipeline: Decision-Focused Learning for Combinatorial Optimization","year":"2019","venue":"","arxiv_id":"1809.05504","aliases":["library:P054"],"technical_key":"","catalog_id":"","library_id":"P054","topic_code":"M5","topic":"不确定性、决策损失与鲁棒优化","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/1809.05504","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"将离散优化连续化，计算决策目标对预测参数的梯度；以优化效用训练模型并与两阶段 MSE 训练比较。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9fb351afae98d036bd90200d2fab39bc9be65955ef60fc7d6140073a0215bee2","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0331","title":"Principles of Mixed-Initiative User Interfaces","year":"1999","venue":"","arxiv_id":"","aliases":["library:P055"],"technical_key":"","catalog_id":"","library_id":"P055","topic_code":"H1","topic":"主动交互、协作与用户需求","task":"未归入视频任务轴","directions":[],"source_url":"https://www.microsoft.com/en-us/research/publication/principles-mixed-initiative-user-interfaces/","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"SVM文本分类并校准意图概率；按收益、误操作及打扰成本设置决策阈值；结合阅读停留时间选择介入时机；允许手动调用、取消、降级到日历范围及编辑建议。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ff2e0c039b6dd3685012d8f763eb59e6b19af4b213eedb4e48c306c5a547e953","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0332","title":"Vibe Coding XR: Accelerating AI + XR Prototyping with XR Blocks and Gemini","year":"待核实","venue":"","arxiv_id":"","aliases":["library:P056"],"technical_key":"","catalog_id":"","library_id":"P056","topic_code":"H4","topic":"XR 开发、协作与无障碍","task":"未归入视频任务轴","directions":[],"source_url":"https://xrblocks.github.io/gem","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Reality Model封装用户、物理场景、感知/交互组件；Gemini生成脚本；XR Blocks v0.11.0通过Playwright无头Chromium监控错误，四种配置各运行5轮。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"00386d1eda45e3dfca86f53706d05a1da0dcf713377c95a227f6cbf76a5ab765","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0333","title":"Thing2Reality: Enabling Spontaneous Creation of 3D Objects from 2D Content using Generative AI in XR Meetings","year":"2025","venue":"","arxiv_id":"","aliases":["library:P057"],"technical_key":"","catalog_id":"","library_id":"P057","topic_code":"H4","topic":"XR 开发、协作与无障碍","task":"未归入视频任务轴","directions":[],"source_url":"https://doi.org/10.1145/3746059.3747621","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"用户圈选对象→MobileSAM分割→条件多视图扩散与LGM Gaussian重建→Unity多人XR；支持缩放、移动、正交视图和白板投影。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"65c443b04fe192201beee9555da57dd6beb247da0ac0109cb92bd6600c4dcb20","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0334","title":"How Well Can 3D Accessibility Guidelines Support XR Development? An Interview Study with XR Practitioners in Industry","year":"2026","venue":"","arxiv_id":"","aliases":["library:P058"],"technical_key":"","catalog_id":"","library_id":"P058","topic_code":"H4","topic":"XR 开发、协作与无障碍","task":"未归入视频任务轴","directions":[],"source_url":"https://doi.org/10.1145/3772318.3790520","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"半结构访谈、指南排序/讨论、主题分析，将优先级、开发阶段和责任组织成矩阵。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"385e35eb3ef535974ac62fae25a3368be55d1a051a299441b67d465aaf621bc1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0335","title":"XR Blocks: Accelerating Human-centered AI + XR Innovation","year":"2025","venue":"","arxiv_id":"2509.25504","aliases":["library:P060"],"technical_key":"","catalog_id":"","library_id":"P060","topic_code":"H4","topic":"XR 开发、协作与无障碍","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2509.25504v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"WebXR、three.js、TensorFlow及Gemini之上封装相机/深度/音频、物理、手势与智能模块，同一高层脚本在模拟器和设备运行。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"97cd982d9d48f4cbb8b12e7693e9408173c86be6fcddff3a47f76596fe89829e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0336","title":"InstructPipe: Generating Visual Blocks Pipelines with Human Instructions and LLMs","year":"2025","venue":"","arxiv_id":"","aliases":["library:P062"],"technical_key":"","catalog_id":"","library_id":"P062","topic_code":"H3","topic":"人机共创与可编辑工作流","task":"未归入视频任务轴","directions":[],"source_url":"https://doi.org/10.1145/3706598.3713905","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"LLM按指令选择节点→依据节点文档生成伪代码→解释器修复并编译JSON节点图→Visual Blocks交互修改。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"cad9e7f81e7c226fd544bd33d6eafded8f1fb5afd791f505462b6690d9c454d1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0337","title":"τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains","year":"2024","venue":"","arxiv_id":"2406.12045","aliases":["library:P067"],"technical_key":"","catalog_id":"","library_id":"P067","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2406.12045v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"给代理政策与工具、模拟用户给需求，通过交互完成状态改变；匹配最终状态并重复运行估计一致成功率。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"398fca396822d58564a3df93c764c066f5ced7e6c1026f9b612dae6ac2b4e5ca","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0338","title":"Thinking in 360°: Humanoid Visual Search in the Wild","year":"2025","venue":"","arxiv_id":"2511.20351","aliases":["library:P079"],"technical_key":"","catalog_id":"","library_id":"P079","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2511.20351v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Thinking in 360° 两阶段 SFT / RL；1280×720；耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2511.20351v2","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.12（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"从全景渲染当前视角，根据图像/历史选择头部动作；Qwen2.5-VL-3B全参数SFT后RL优化搜索成功。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"9eabc1074cc00ed00d8242545fe3c98347afd3a5ba30f08ebb1e5a06f4f3a4e5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0339","title":"SaaS-Bench: Can Computer-Use Agents Leverage Real-World SaaS to Solve Professional Workflows?","year":"2026","venue":"","arxiv_id":"2605.15777","aliases":["library:P094"],"technical_key":"","catalog_id":"","library_id":"P094","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.15777v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"提供任务、应用入口和凭据，代理操作浏览器/GUI；检查各中间状态和最终目标，分析阶段衰减与错误类型。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4acfd285390f335e9d6b80b97bec515de0e4d06e01b151ba3d9b98feea38d838","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0340","title":"Large Language Models and Causal Inference in Collaboration: A Comprehensive Survey","year":"2025","venue":"","arxiv_id":"","aliases":["library:P063"],"technical_key":"","catalog_id":"","library_id":"P063","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://aclanthology.org/2025.findings-naacl.427/","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"按“因果推断→LLM”和“LLM→因果推断”组织研究，人工筛选相关论文/预印本并归纳挑战。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"6c8c92b856e343fdfee68219f509fa5f2af027daffc6617146615ced4a7007cc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0341","title":"Watch-And-Help: A Challenge for Social Perception and Human-AI Collaboration","year":"2021 / 2020","venue":"","arxiv_id":"2010.09890","aliases":["library:P141"],"technical_key":"","catalog_id":"","library_id":"P141","topic_code":"H1","topic":"主动交互、协作与用户需求","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2010.09890v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Transformer编码示范中的物体状态，再经LSTM与时间聚合预测目标谓词；对比层次规划HP、层次RL及学习高层策略与规划低层动作的Hybrid。规划器考虑Alice的未来计划，降低重复劳动与动作冲突。训练1,011个任务；两个测试集各100个任务，保留未见公寓及谓词组合，单轮最多250步。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3046ec1c6b04f13bf1909519dd21d3abcf1511c0ed371e3dd5c00e9fa67c82e2","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0342","title":"Alchemy: A Benchmark and Analysis Toolkit for Meta-Reinforcement Learning Agents","year":"2021","venue":"","arxiv_id":"2102.02926","aliases":["library:P142"],"technical_key":"","catalog_id":"","library_id":"P142","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2102.02926v3","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Unity第一人称3D环境与共享机制的符号环境；药水在三维属性立方体上改变石头状态，目标是提高投入坩埚的石头价值。训练IMPALA/LSTM与VMPO/Transformer-XL基线，另实现维护可能机制集合并搜索动作的理想观察者。通过额外提供真值机制或Bayesian belief、增加预测辅助损失，诊断感知与推断瓶颈。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"559efd53b1b8fe120d2b2ed36799c235fcb488130942b7ad9539c917501ab396","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0343","title":"Causality Model for Semantic Understanding on Videos","year":"2025","venue":"","arxiv_id":"2503.12447","aliases":["library:P064"],"technical_key":"","catalog_id":"","library_id":"P064","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.12447v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"V100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"IGV 训练 MSRVTT-QA 约 2.6h；GPU 数量未明确，不能换算 GPU-hours","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2503.12447v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"IVRD以对象轨迹和关系原型降低对象—关系偏置；IGV置换环境保持答案不变；EIGV再要求因果场景/问题语义变化引发答案等变；TranSTR可微选择时刻/对象并改进候选答案解码。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"ccbcdbddcfc1a9178657e2e0dde9cd7c730ab20fb0f26283d96a51237dc5defa","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0344","title":"Human-in-the-Loop through Chain-of-Thought","year":"2023","venue":"","arxiv_id":"2306.07932","aliases":["library:P065"],"technical_key":"","catalog_id":"","library_id":"P065","topic_code":"H2","topic":"人工核查、推理交互与依赖校准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2306.07932","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"人工检查并修正推理子步骤，继续生成；可结合self-consistency；以Cobb–Douglas式效用与预算分析纠错策略。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"dfd3afde30daeddbac23f7569b84477a6e8fd481c60a35744cc35375b92bcade","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0345","title":"Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation","year":"2023","venue":"","arxiv_id":"2302.09664","aliases":["library:P144"],"technical_key":"","catalog_id":"","library_id":"P144","topic_code":"M5","topic":"不确定性、决策损失与鲁棒优化","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2302.09664v3","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"对同一问题采样M个答案；将问题与答案一起输入NLI模型，在两个方向均为entailment时聚为同一含义；聚合簇内序列概率并估计语义熵。对照普通/长度归一化预测熵、词面相似度和p(True)，用错误检测AUROC评价。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"de81564e8da91783b726f7b41c2d7eaf356d6dbea560a2b41ebfd438839ef77c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0346","title":"Let’s Verify Step by Step","year":"2024","venue":"","arxiv_id":"","aliases":["library:P066"],"technical_key":"","catalog_id":"","library_id":"P066","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://proceedings.iclr.cc/paper_files/paper/2024/file/aca97732e30bcf1303bc22ac3924fd16-Paper-Conference.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"人类逐步判断正确/错误/中性，训练步骤奖励模型；与自动最终答案标签训练的ORM和多数投票比较；主动选择高价值错误样本。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"bf5188de69a8c43780c1ba557a9012a0ddc335cbcb81698aabb95097602a6d61","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0347","title":"Generative Agent-Based Modeling with Actions Grounded in Physical, Social, or Digital Space Using Concordia","year":"2023","venue":"","arxiv_id":"2312.03664","aliases":["library:P145"],"technical_key":"","catalog_id":"","library_id":"P145","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.03664v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"每个智能体用长期关联记忆与工作记忆组件回答身份、情境、适当行动等问题；GM解析尝试动作、更新状态并向不同角色发送其可见观察。数字案例通过嵌套PhoneGameMaster和PhoneUniverse选择app、函数和参数，再更新应用状态及通知。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4211d73303e28ff1e6c77870f44a1f411aff6d362152e494a83648da27e631c6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0348","title":"Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model","year":"2025","venue":"","arxiv_id":"2411.19108","aliases":["library:P162"],"technical_key":"","catalog_id":"","library_id":"P162","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2411.19108v2","status":"无需本文权重训练","resource_note":"TeaCache 是推理时缓存方法，表中 A800 数量对应 inference scaling。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"计算相邻时间步的调制输入相对L1距离，以多项式拟合输入差异到输出差异的尺度关系；累计估计差异未过阈值时复用残差，超过阈值则重算并清零累计值。用T2V-CompBench的70个提示、覆盖7类属性作拟合；slow/fast阈值分别为0.1/0.2。在Latte、Open-Sora 1.2、Open-Sora-Plan及A800 80GB GPU上评估速度和VBench、LPIPS、SSIM、PSNR。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"7655822e0615d448451219f51b167682da6c2e32f47b68fbbb6f95405a155802","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0349","title":"Automating the Search for Artificial Life with Foundation Models","year":"2025","venue":"","arxiv_id":"2412.17799","aliases":["library:P163"],"technical_key":"","catalog_id":"","library_id":"P163","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2412.17799v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"将初始状态、动力学和渲染器表示为参数theta。单目标搜索最大化渲染图像与文本的CLIP相似度，使用Sep-CMA-ES；NCA的时序目标用截断时间反向传播和Adam。开放性目标降低当前帧与历史最近邻的相似度，对2^18=262,144种Life-like CA穷举。多样性搜索用遗传算法维护8,192个解，通过最近邻表示距离淘汰最不新颖的解。以DINOv2和像素表示做表示空间消融。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3130bdf6d55493f27c843f9ffafed3ed099e1dd3a61897f715eb3ea13ad4c67d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0350","title":"DreamGarden: A Designer Assistant for Growing Games from a Single Prompt","year":"2025 / 2024","venue":"","arxiv_id":"","aliases":["library:P147"],"technical_key":"","catalog_id":"","library_id":"P147","topic_code":"H3","topic":"人机共创与可编辑工作流","task":"未归入视频任务轴","directions":[],"source_url":"https://www.microsoft.com/en-us/research/publication/dreamgarden-a-designer-assistant-for-growing-games-from-a-single-prompt/?lang=ko-kr","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"GPT-4o将seed prompt递归展开为计划树，叶节点分派给代码/素材模块；代码经历编译、布局加载、运行、截图评价的反复修正。GUI允许变更叶节点状态、直接改代码或反馈，再继续生成。N=10、每次最多45分钟的形成性可用性研究记录访谈、操作与系统日志。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2855c78af350348f2c6a22312a340ad241fba4cba6f789723f228e4366f97def","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0351","title":"Detecting Hallucinations in Large Language Models Using Semantic Entropy","year":"2024","venue":"","arxiv_id":"","aliases":["library:P148"],"technical_key":"","catalog_id":"","library_id":"P148","topic_code":"M5","topic":"不确定性、决策损失与鲁棒优化","task":"未归入视频任务轴","directions":[],"source_url":"https://doi.org/10.1038/s41586-024-07421-0","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"对同一问题多次采样，用双向蕴含判断将回答按含义聚类；有概率时聚合语义概率，无概率时用簇的样本频率估计离散语义熵。长文本先拆分事实陈述，再围绕陈述生成问题、重采样回答并汇总不确定性；以AUROC和拒答—准确率曲线等评估。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"47dd7837e5f7c6d36390d7670b8e94e822f62b38ad17bae648c4f42f9ea8cbe0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0352","title":"WorldCoder, a Model-Based LLM Agent: Building World Models by Writing Code and Interacting with the Environment","year":"2024","venue":"","arxiv_id":"","aliases":["library:P150"],"technical_key":"","catalog_id":"","library_id":"P150","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://proceedings.neurips.cc/paper_files/paper/2024/hash/820c61a0cd419163ccbd2c33b268816e-Abstract-Conference.html","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"分离Python转移函数与目标条件奖励函数。模型需满足对所有已收集transition的拟合约束，以及存在可达正奖励轨迹的乐观约束。GPT-4根据不一致样本修补代码，REx决定优先改哪个候选；配合有限深度值迭代/MCTS规划、经验回放和少量随机探索。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ad0ce8a558f3edccb5298ede1e7005ff37f52b608cc3eb40f3fa212617ab3c34","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0353","title":"COMBO: Compositional World Models for Embodied Multi-Agent Cooperation","year":"2025 / 2024","venue":"","arxiv_id":"2404.10775","aliases":["library:P151","reasoning:ICLR-2025-2989"],"technical_key":"","catalog_id":"","library_id":"P151","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2404.10775v3","status":"已核对资源阶段","resource_note":"","stages":[{"role":"aux","device":"V100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":0.333333,"wall_hours_max":0.333333,"time_bound":"approximate","gpu_hours_min":0.3333,"gpu_hours_max":0.3333,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"COMBO 的 supervised component baseline；10 epochs，约 20min；不是完整生成 world model","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/abs/2404.10775v3","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.22（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"融合历次局部RGBD为俯视点云投影并用扩散补全；通过组合各智能体动作条件的扩散预测，并按其可达区域加权学习。VLM负责动作提议、意图跟踪和结果评分，再用树搜索选择计划。对异常高预测收益设置阈值以抑制利用世界模型幻觉。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9f6c5094bc9da9e2417a4789b586e1efc776b73c67a61706c9e5bf40e908f5eb","reasoning_id":"ICLR-2025-2989","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0354","title":"DiscoveryWorld: A Virtual Environment for Developing and Evaluating Automated Scientific Discovery Agents","year":"2024","venue":"","arxiv_id":"","aliases":["library:P152"],"technical_key":"","catalog_id":"","library_id":"P152","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://proceedings.neurips.cc/paper_files/paper/2024/hash/13836f251823945316ae067350a5c366-Abstract-Datasets_and_Benchmarks_Track.html","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"提供物体、仪器、移动/读写/操作等动作及结构化记分卡；GPT-4o实现ReAct、Plan+Execute、维护实验假设与测量记录的Hypothesizer。简单任务最多100步，普通/挑战任务1,000步；知识按预设解释要求评价。另招募11名自然科学硕士/博士背景参与者。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"8488e239e87b6aa0891db6e028fe7a20f724612e2efda78ddabc7b5cd20e26e9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0355","title":"Internal Consistency and Self-Feedback in Large Language Models: A Survey","year":"2024","venue":"","arxiv_id":"2407.14507","aliases":["library:P153"],"technical_key":"","catalog_id":"","library_id":"P153","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2407.14507v3","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"按黑箱/灰箱/白箱访问条件，梳理不确定性、置信度、幻觉、语言批评及对比等反馈；归纳推理拓扑、迭代修改、多智能体、偏好学习、蒸馏和数据增广。以Llama3-8B-Instruct的标点计数展示三个一致性层次的区别，并讨论hourglass假说。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c66fb1b57701dd1cc3b36d5f629acab96b516c4504606ccc8550cc82f68df647","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0356","title":"Diffusion as Shader: 3D-aware Video Diffusion for Versatile Video Generation Control","year":"2025","venue":"","arxiv_id":"2501.03847","aliases":["library:P164"],"technical_key":"","catalog_id":"","library_id":"P164","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2501.03847v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H800","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":576.0,"gpu_hours_max":576.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Diffusion as Shader；少于 10K videos；3 天 fine-tuning","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2501.03847v2","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.1（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"从CogVideoX的42个DiT块复制前18块作为条件分支，VAE编码3D跟踪视频，逐块注入控制特征。使用MiraData真实视频和Mixamo渲染视频；真实视频用SpatialTracker跟踪4,900个点，合成视频用真值几何。训练视频为49帧720×480，2,000步、有效batch size 64；论文报告少于1万视频，在8张H800上训练3天。推理50步DDIM，生成49帧约2.5分钟/张H800。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"231038b643eb5d3e4148b3c1904ceccc1971a3b1b98fc5ad3b3486845368dc71","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0357","title":"TPDiff: Temporal Pyramid Video Diffusion Model","year":"2025","venue":"","arxiv_id":"2503.09566","aliases":["library:P068"],"technical_key":"","catalog_id":"","library_id":"P068","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.09566v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"TPDiff 在 100K OpenVid pairs 上适配；GPU 数和耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2503.09566v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"沿去噪过程分段求解概率流ODE，阶段间时序插值并对齐数据/噪声，仅末段使用完整帧率。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"943b4c6b9a7984f4ad0f79a134f27f9fa81918f3c126fb110445c339e0f4a978","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0358","title":"VideoMind: A Chain-of-LoRA Agent for Temporal-Grounded Video Reasoning","year":"2026","venue":"","arxiv_id":"2503.13444","aliases":["library:P069"],"technical_key":"","catalog_id":"","library_id":"P069","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.13444v3","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"planner决定需要何种角色，grounder提出时间段，verifier筛选，answerer基于片段回答；同一Qwen2-VL骨干动态加载角色LoRA。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"56cd747f8a5dba9089dec417aaffd3228963d2e724c96ebb7e336a4a34cb3464","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0359","title":"Impossible Videos","year":"2025","venue":"","arxiv_id":"2503.14378","aliases":["library:P070"],"technical_key":"","catalog_id":"","library_id":"P070","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.14378v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"按分类设计提示、收集筛选生成视频，对生成质量/提示跟随及Video-LLM异常理解做评测。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"67b4a88e13de3ced666c8be00d0594eee296083c14f536bc80e3b48c2f688466","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0360","title":"Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance","year":"2025","venue":"","arxiv_id":"2512.08765","aliases":["library:P072"],"technical_key":"","catalog_id":"","library_id":"P072","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.08765v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Wan-Move；sequence parallel size 4；未定位耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2512.08765v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"将点轨迹投影到VAE潜空间，沿轨迹复制首帧特征，形成时空条件图，对基础I2V模型微调。","training_signal":"不确定","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"456416958be917841520560bbc4f1657047c23a0e86a718229e0aba0d6f5e4fc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0361","title":"Text2World: Benchmarking Large Language Models for Symbolic World Model Generation","year":"2025","venue":"","arxiv_id":"","aliases":["library:P154"],"technical_key":"","catalog_id":"","library_id":"P154","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://aclanthology.org/2025.findings-acl.1337/","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"从已有PDDL域构造自然语言描述，多阶段自动过滤后由6位计算机背景标注者及2位复核者做质量控制。评价9个系列的16个LLM；比较零样本CoT、不做修正EC0与最多3次解析报错反馈EC3、两样本提示、601个合成域上的SFT及智能体轨迹训练。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9515b61e1f05370debca09ab72f734046c5d544bd2e0f05712a5ca490510f1d8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0362","title":"The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity","year":"2025","venue":"","arxiv_id":"","aliases":["library:P073"],"technical_key":"","catalog_id":"","library_id":"P073","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://machinelearning.apple.com/research/illusion-of-thinking","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"调节任务规模、以模拟器验证逐步动作，每实例25次采样；匹配推理预算比较pass@k，分析解答出现位置及token使用。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"81b4476c977788b7d3ea3cba5553a845e5256eb6e06d7d8acc775e73b781b999","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0363","title":"Beyond Single-Sentence Prompts: Upgrading Value Alignment Benchmarks with Dialogues and Stories","year":"2025","venue":"","arxiv_id":"2503.22115","aliases":["library:P074"],"technical_key":"","catalog_id":"","library_id":"P074","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.22115v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"自动/人工清洗原始问题，生成对话陷阱与故事，组合人工和自动评分，比较不同模型在对话/叙事场景的表现。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2908770ad7f21944d18c4f188d541a099d35476f2ccf416e57eb2c8ddf492a0a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0364","title":"Chain-of-Evidence Multimodal Reasoning for Few-shot Temporal Action Localization","year":"2025","venue":"","arxiv_id":"2504.13460","aliases":["library:P075"],"technical_key":"","catalog_id":"","library_id":"P075","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2504.13460v5","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A6000","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Chain-of-Evidence temporal localization；ActivityNet 200 epochs；小型任务模块，非视频生成器","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2504.13460v5","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"从视频产生事件证据描述，逐层精化时间/动作关系；对齐query/support视觉及语义特征，联合建模动作共同性与差异。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"63b6594cc6f4dfa9ca7e891f25b309956705fedcaad2323f8981ebb8a758f9e3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0365","title":"Interactive Reasoning: Visualizing and Controlling Chain-of-Thought Reasoning in Large Language Models","year":"2025","venue":"","arxiv_id":"2506.23678","aliases":["library:P076"],"technical_key":"","catalog_id":"","library_id":"P076","topic_code":"H2","topic":"人工核查、推理交互与依赖校准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2506.23678v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"DeepSeek-R1生成推理，GPT-4o负责整理树、澄清和连接节点；反馈节点/直接编辑后续写；16人被试内比较。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"78c3f5dfa697a9db99d3193f678abb1e991ff689646f330f520bf73669cc0949","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0366","title":"Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens","year":"2026","venue":"","arxiv_id":"2508.01191","aliases":["library:P077"],"technical_key":"","catalog_id":"","library_id":"P077","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2508.01191v6","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"CoT Mirage 实验资源；另有 4 H200；预训练/适配/推理分配不明确，不纳入并行数统计","comparable":false,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2508.01191v6","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.37（同文多阶段请结合上方范围）"},{"role":"ft","device":"H200","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"CoT Mirage 另一设备池；不同阶段资源归属未拆分，不纳入并行数统计","comparable":false,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2508.01191v6","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.37（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"控制训练变换组合、链长度和表达格式，训练/微调多架构规模模型并对比ID、组合迁移和OOD；分别评估链与答案exact match。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c2fdd4d8f6026b08b117a35ec4ea0c0a28614ade7b946ddc26e5be5fc805894f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0367","title":"EgoPlan-Bench2: A Benchmark for Multimodal Large Language Model Planning in Real-World Scenarios","year":"2026 / 2025 / 2024","venue":"","arxiv_id":"2412.04447","aliases":["library:P155"],"technical_key":"","catalog_id":"","library_id":"P155","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2412.04447v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"GPT-4辅助清理动作叙述、推断目标及生成问答；InternVL1.5与GPT-4帮助自适应选择时间点，再做循环评测与人工核查。训练自由的增强流程结合动作历史摘要、GroundingDINO对象边界框、CoT及自一致性投票。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"55d1f4d4d6e312a65ee45cecf68aa37d02e0be89499a84826c0a890a71c2890b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0368","title":"PoE-World: Compositional World Modeling with Products of Programmatic Experts","year":"2025","venue":"","arxiv_id":"2505.10819","aliases":["library:P156"],"technical_key":"","catalog_id":"","library_id":"P156","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2505.10819v4","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"LLM合成短Python程序，其确定性属性预测转换为带噪专家分布，未规定属性采用均匀分布；按加权乘积组合。L-BFGS最大似然拟合权重并删去弱专家，新轨迹触发补充/修正。使用对象接触关系的高层图和低层动作搜索做分层规划，也用世界模型训练RL策略。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"de94ad39af6787de033663faf57f8d29f7c4f65100842c046706fca6a6382138","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0369","title":"Seeing but Not Believing: Probing the Disconnect Between Visual Attention and Answer Correctness in VLMs","year":"2025","venue":"","arxiv_id":"2510.17771","aliases":["library:P078"],"technical_key":"","catalog_id":"","library_id":"P078","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2510.17771v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"定位视觉grounding层，提取证据分数，选择性遮罩/突出图像区域后重推理；跨4个系列8个VLM比较。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3964cb39f172431f490395350878905f6645db4e766816c11c613798ed3992c9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0370","title":"Vis-CoT: A Human-in-the-Loop Framework for Interactive Visualization and Intervention in LLM Chain-of-Thought Reasoning","year":"2025","venue":"","arxiv_id":"2509.01412","aliases":["library:P080"],"technical_key":"","catalog_id":"","library_id":"P080","topic_code":"H2","topic":"人工核查、推理交互与依赖校准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2509.01412v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Llama2-70B-Chat、React/D3界面与FastAPI，24名相关专业参与者；GSM8K/StrategyQA及100项自建规划任务，人工评价部分任务的答案与有效推理链。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1202be03fa79d5a060ab3a915be880962b4602fdc4a7cfb6b80ea48a91372627","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0371","title":"RULER-Bench: Probing Rule-based Reasoning Abilities of Next-level Video Generation Models for Vision Foundation Intelligence","year":"2025","venue":"","arxiv_id":"2512.02622","aliases":["library:P158"],"technical_key":"","catalog_id":"","library_id":"P158","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.02622v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"人工与GPT-5构造任务、隐含解释和核查表并复核；评测10个视频模型（6闭源、4开源），由o3逐项评分。另比较加入规则/预期结果解释的增强提示，并以80个视频、813个问题的人类标注核对自动评审。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5b25796cac77e7f705bdcde2a9ea3b0c8aa9b3925dceca21c0a54bd5aa1321fa","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0372","title":"Causal-JEPA: Learning World Models through Object-Level Latent Masking","year":"2026","venue":"","arxiv_id":"2602.11389","aliases":["library:P083"],"technical_key":"","catalog_id":"","library_id":"P083","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.11389v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX A6000","count_min":2.0,"count_max":2.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Causal-JEPA；450K steps；batch 32/GPU；初始化自 pretrained SAVi decoder，完整耗时未知","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2602.11389v2","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.18（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"对象编码器产生slots，预测被遮对象的历史/未来潜状态，避免像素重建；基于紧凑潜模型规划控制。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":2.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"29c36d557599af454ee38ec78b993675dd5d0a78ec1361856feec8494028b6f2","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0373","title":"CRONOS: Benchmarking Counterfactual Physical Consistency in Video Models","year":"2026","venue":"","arxiv_id":"2605.23699","aliases":["library:P084"],"technical_key":"","catalog_id":"","library_id":"P084","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.23699v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"生成碰撞、跌落、遮挡等受控视频；比较Cosmos2.5、CogVideoX1.5、MAGI1、Wan2.2；部分支持5帧V2V；结合规则、视觉模型和VLM评判。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"76697444e47ca3eb3220867e0f8276dec8f83b795b5a08eb56706bd03b706a0a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0374","title":"SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models","year":"2026","venue":"","arxiv_id":"2609.02886","aliases":["library:P085"],"technical_key":"","catalog_id":"","library_id":"P085","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2609.02886v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"对齐帧、度量相机几何、字幕和来源；双向适配→teacher-forced自回归初始化/AnyFlow→分布匹配蒸馏；保留各骨干原生表征。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3e12e824b36c53dd79afe109de6df9e17df9af15ff45fb9852c4e21dd97a3ec7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0375","title":"See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs","year":"2026","venue":"","arxiv_id":"2602.21497","aliases":["library:P086"],"technical_key":"","catalog_id":"","library_id":"P086","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.21497v2","status":"无需本文权重训练","resource_note":"See It, Say It, Sorted 明确为 training-free 的视觉推理流程。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"改进VDGD的候选词和KL证据偏好，融合基础置信度，按当前上下文决定是否追加图像证据并继续解码。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"45e63892ea0cdfed5b5e0a9aa067d520d8685e1e032641af801b185585084dec","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0376","title":"Don't Blink: Evidence Collapse during Multimodal Reasoning","year":"2026","venue":"","arxiv_id":"2604.04207","aliases":["library:P087"],"technical_key":"","catalog_id":"","library_id":"P087","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2604.04207v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"用标注证据区域选视觉grounding层，分析推理前后注意力、输出熵和正确性，拟合跨任务交互并在固定覆盖率筛除高风险回答。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"fccfa3f205730f41e445c0bb3d17427e0eb2138aaa00f14463caf6efcd7ab5aa","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0377","title":"V-Retrver: Evidence-Driven Agentic Reasoning for Universal Multimodal Retrieval","year":"2026","venue":"","arxiv_id":"2602.06034","aliases":["library:P088"],"technical_key":"","catalog_id":"","library_id":"P088","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.06034v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A800","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"V-Retrver Qwen2.5-VL-7B SFT / rejection fine-tuning；2 epochs，batch 64；RL 成本另缺","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2602.06034v2","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.6（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"监督激活推理→拒绝采样精炼→证据对齐RL，学习针对模糊候选进行选择性图像检查。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"62b45007ab83930c62f04e18aecdbb3aaf123983f2f6d34bae24e6288a48fb76","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0378","title":"Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models","year":"2026","venue":"","arxiv_id":"2605.12374","aliases":["library:P089"],"technical_key":"","catalog_id":"","library_id":"P089","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.12374v4","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"B200","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":3.916667,"wall_hours_max":3.916667,"time_bound":"approximate","gpu_hours_min":125.3333,"gpu_hours_max":125.3333,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"GAP student fine-tuning 3h55min；不含 offline image generation、teacher 和训练数据构造","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2605.12374v4","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.15（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"构造49,309个图像—问题—辅助图像样本，学习映射回视觉输入子空间，选择基础模型困难题应用潜监督，并控制latent token预算。","training_signal":"不确定","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"a353c19517533dd98866eaf193086bfea639ed2dd167d68d7b86f2d64601c4ec","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0379","title":"From <Answer> to <Think>: Multi-Dimensional Supervision of Reasoning Process for LLM Optimization","year":"未注明（匿名稿）","venue":"","arxiv_id":"","aliases":["library:P090"],"technical_key":"","catalog_id":"","library_id":"P090","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://openreview.net/pdf?id=MjJrhfAmRG","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"置信度校准、Qwen3-8B-Reranker语义相关性及Nemotron70B奖励模型连贯性组成分数，构造偏好对或在线奖励，可与可验证答案奖励组合。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3b134cbc0f5318c2a0c5a77a33b79b7ee9c137196322d4a3150f7decf4c836b1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0380","title":"Improving Human Verification of LLM Reasoning through Interactive Explanation Interfaces","year":"2026","venue":"","arxiv_id":"","aliases":["library:P091"],"technical_key":"","catalog_id":"","library_id":"P091","topic_code":"H2","topic":"人工核查、推理交互与依赖校准","task":"未归入视频任务轴","directions":[],"source_url":"https://doi.org/10.1145/3742413.3789134","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"保持解释底层内容，改变结构、渐进展示、变量高亮和依赖关系；被试组间比较正确性核验、错误定位、时间及体验。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9c063b20133876c79d7af03bd605bc991a764cff322b38f84cc58da0d7e5cb4c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0381","title":"Improving Visual Reasoning with Iterative Evidence Refinement","year":"2026","venue":"","arxiv_id":"2603.14117","aliases":["library:P092"],"technical_key":"","catalog_id":"","library_id":"P092","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2603.14117v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":2.0,"count_max":2.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"SIEVE Qwen3-VL-4B / 8B GRPO；60 rollout steps，16 prompts × 8 rollouts","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2603.14117v1","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"依据文本锚点和视觉显著性抽取区域表示，在生成中触发重访并注入embedding，按回答奖励学习调用策略。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":2.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"8c3cf425eaafbd00e83a9575b8877ef59bc3279fcf25774388d27b8ae6362a80","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0382","title":"Instruction-Evidence Contrastive Dual-Stream Decoding for Grounded Vision-Language Reasoning","year":"2026","venue":"","arxiv_id":"2604.25809","aliases":["library:P093"],"technical_key":"","catalog_id":"","library_id":"P093","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2604.25809v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"每步计算instruction/evidence两分布；一致时保留表达性，分歧大时抑制仅语言流偏好的token；跨caption/VQA/对象幻觉评估。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"98f33ba3012f78380939658ba05c519f6e407b26cb5ad8bc4208c6c3e521353c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0383","title":"Fictional Worldbuilding: Multi-Agent LLM Collaboration with Hierarchical Context Compression and Iterative Review","year":"2026","venue":"","arxiv_id":"2607.09403","aliases":["library:P160"],"technical_key":"","catalog_id":"","library_id":"P160","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2607.09403v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"将输入拆成有依赖的任务，拓扑排序后按语义与批大小分组；从已完成批次检索相关概念，按Essential/Relevant/Summary/Collaboration四层分配上下文。21种专门角色按技能配置加载，8种Auditor参与评分、修订和筛选；主要测试中文输入。20个案例覆盖5类世界，每种后端各运行20次，共40次。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"684b789b5233a1d6e48e6a249c267468bff6d0cd3ef7c56f12401e12ad2f64df","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0384","title":"TIR-Flow: Active Video Search and Reasoning with Frozen VLMs","year":"2026","venue":"","arxiv_id":"2601.06176","aliases":["library:P095"],"technical_key":"","catalog_id":"","library_id":"P095","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2601.06176v1","status":"无需本文权重训练","resource_note":"TIR-Flow 使用冻结 VLM 的 active video search，明确为 training-free。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"把问题拆成可验证子任务，针对假设搜索/放大视频证据，累积修订线索再回答。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"be48a0f4cf27cf2c3df3c58c080c9d2f9a30adb602d8b2310a6c569ddaf3c699","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0385","title":"TSRBench: A Comprehensive Multi-task Multi-modal Time Series Reasoning Benchmark for Generalist Models","year":"2026","venue":"","arxiv_id":"2601.18744","aliases":["library:P096"],"technical_key":"","catalog_id":"","library_id":"P096","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2601.18744v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"统一时间序列多模态输入和任务定义，比较文本/视觉/合并输入、模型规模及子任务关联。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"00cd41be3e257efab518058dc368d7d5290e671d3401cb5e239d2a5a9e889453","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0386","title":"First Logit Boosting: Visual Grounding Method to Mitigate Object Hallucination in Large Vision-Language Models","year":"2026","venue":"","arxiv_id":"2604.00455","aliases":["library:P098"],"technical_key":"","catalog_id":"","library_id":"P098","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2604.00455v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"缓存l₀，在每一步对当前logits加入加权初始分布后采样；跨对象幻觉任务、模型及基准比较。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f409975c5f5cfde14f17f590e56ffc75cc74186ac29e6a4ee7077f812dc8a133","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0387","title":"VGS-Decoding: Visual Grounding Score Guided Decoding for Hallucination Mitigation in Medical VLMs","year":"2026","venue":"","arxiv_id":"2603.20314","aliases":["library:P099"],"technical_key":"","catalog_id":"","library_id":"P099","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2603.20314v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"向原图加Gaussian/Poisson噪声，双路得到token分布，以视觉退化后的概率变化定义VGS并重加权候选。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"123cabd480a7097986ab10ea53a6cedd390c1f2857bc20e88ed5b4b7280c8262","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0388","title":"When Thinking Drifts: Evidential Grounding for Robust Video Reasoning","year":"待核实","venue":"","arxiv_id":"2510.06077","aliases":["library:P100"],"technical_key":"","catalog_id":"","library_id":"P100","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2510.06077","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Evidential Grounding；GRPO group 8；2K iterations","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2510.06077","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"将答案正确、视觉证据、格式和长度奖励结合GRPO（证据权重0.3），训练模型产生与问题相关视频内容一致的推理。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"24e028d0c7acb59804cf2a037aa3323e5de373b9a834029b8b9f267fb2ade9c7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0389","title":"Computational Understanding of Narratives: A Survey","year":"2022","venue":"","arxiv_id":"","aliases":["library:P101"],"technical_key":"","catalog_id":"","library_id":"P101","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://doi.org/10.1109/ACCESS.2022.3205314","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"围绕实体/事件、主题、意图和关系组织IR/NLU工作，比较自包含故事与跨来源演化叙事。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"778fe4d13f42b0f1097039655fab479896cff008cd33a94a6aad2cf344133fd6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0390","title":"Training Large Language Models to Reason in a Continuous Latent Space","year":"2025","venue":"","arxiv_id":"2412.06769","aliases":["library:P103"],"technical_key":"","catalog_id":"","library_id":"P103","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2412.06769v3","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"逐阶段用连续thought替代更多文本步骤，回馈隐状态，再生成文字答案；比较无CoT、iCoT、pause token等。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"eebe49ae115218a68fe272340b7bb2eb875c74399cfb8a42d3bc63dc11345aef","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0391","title":"VIMA: General Robot Manipulation with Multimodal Prompts","year":"2023","venue":"","arxiv_id":"2210.03094","aliases":["library:P165"],"technical_key":"","catalog_id":"","library_id":"P165","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2210.03094","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"VIMA 通过多模态提示、自回归动作预测和程序化桌面任务研究分层泛化。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b7351d797e8d209434a6428251fcbd139ec8e1a70ffc8c0eaf0f59585ecd3173","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0392","title":"MimicPlay: Long-Horizon Imitation Learning by Watching Human Play","year":"2023","venue":"","arxiv_id":"2302.12422","aliases":["library:P168"],"technical_key":"","catalog_id":"","library_id":"P168","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2302.12422","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"GPU（型号未定位）","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":12.0,"wall_hours_max":12.0,"time_bound":"approximate","gpu_hours_min":12.0,"gpu_hours_max":12.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"MimicPlay latent planner；100K iterations","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2302.12422","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"pre","device":"GPU（型号未定位）","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":12.0,"wall_hours_max":12.0,"time_bound":"approximate","gpu_hours_min":12.0,"gpu_hours_max":12.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"MimicPlay low-level policy；100K iterations；设备型号未定位","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2302.12422","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MimicPlay 从人类视频学习潜计划，指导少量遥操作数据训练的低层控制。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a7e399799dafda725fa629946da93427ea63eabd5e51503dc6cd04029624e993","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0393","title":"Modeling Dynamic Environments with Scene Graph Memory","year":"2023","venue":"","arxiv_id":"2305.17537","aliases":["library:P175"],"technical_key":"","catalog_id":"","library_id":"P175","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2305.17537","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Scene Graph Memory 累积对象和房间关系，以节点边预测指导动态家庭环境中的搜索。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"df60242fba9c10016a9a33d0cd6bd64a47c3ed744738fa9fde53a8133f9b71a5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0394","title":"Sonicverse: A Multisensory Simulation Platform for Embodied Household Agents that See and Hear","year":"2023","venue":"","arxiv_id":"2306.00923","aliases":["library:P176"],"technical_key":"","catalog_id":"","library_id":"P176","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2306.00923","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Sonicverse 将音视频模拟结合，研究多感官导航与仿真到真实迁移。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3b94420c75cd69e57f001a77684717f8e97d102d68d929eb39d44f694d6de37b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0395","title":"The ObjectFolder Benchmark: Multisensory Learning with Neural and Real Objects","year":"2023","venue":"","arxiv_id":"2306.00956","aliases":["library:P177"],"technical_key":"","catalog_id":"","library_id":"P177","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2306.00956","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"ObjectFolder Benchmark 结合神经对象及真实多感官对象数据，比较识别、重建和操作任务。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f3c84cba42ba1d36c6870374f6823f11920d901bdc8eca7786505dbf9b6266c8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0396","title":"Dynamic-Resolution Model Learning for Object Pile Manipulation","year":"2023","venue":"","arxiv_id":"2306.16700","aliases":["library:P179"],"technical_key":"","catalog_id":"","library_id":"P179","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2306.16700","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者以动态粒子分辨率和图神经网络，在效率与操作精度间选择抽象层级。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9d84835f2305eebdaf5b89834307d13525760e4cfd443c4867ec7b93b90d1683","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0397","title":"VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models","year":"2023","venue":"","arxiv_id":"2307.05973","aliases":["library:P182"],"technical_key":"","catalog_id":"","library_id":"P182","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2307.05973","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"VoxPoser 组合三维价值图进行闭环轨迹规划，并可从在线接触经验学习动力学。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"fd5f244fc3aa6af3ced2da27026a28d1d3ec0ff427feb8cc73f1dfa84aaf2d1a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0398","title":"Primitive Skill-based Robot Learning from Human Evaluative Feedback","year":"2023","venue":"","arxiv_id":"2307.15801","aliases":["library:P184"],"technical_key":"","catalog_id":"","library_id":"P184","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2307.15801","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"SEED 将参数化基础技能与人类评价强化学习结合，研究样本效率与人类负担。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2fac6229642822f146e63fc5778ee257c35086feaf4c295ddc9dd5a88a8c3fdc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0399","title":"Sequential Dexterity: Chaining Dexterous Policies for Long-Horizon Manipulation","year":"2023","venue":"","arxiv_id":"2309.00987","aliases":["library:P185"],"technical_key":"","catalog_id":"","library_id":"P185","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2309.00987","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 3090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Sequential Dexterity policy 训练仿真：1,024 parallel agents，5K FPS；未据仿真吞吐推算训练时长","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2309.00987","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.18（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Sequential Dexterity 通过强化学习衔接多个策略，并研究新物体与真实机器人的迁移。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a7a8112a8ad096af52dd820b46023bcf77f557ac8c7e1cea2e353c2d1e8f8a25","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0400","title":"MindAgent: Emergent Gaming Interaction","year":"2023","venue":"","arxiv_id":"2309.09971","aliases":["library:P186"],"technical_key":"","catalog_id":"","library_id":"P186","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2309.09971","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MindAgent 提供 CUISINEWORLD 和协作效率评价，通过反馈进行上下文协调。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3129f9a2651d66d916dce4eaa94a616f60a0854d23508d43c7b8bf3ab9065b76","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0401","title":"D$^3$Fields: Dynamic 3D Descriptor Fields for Zero-Shot Generalizable Rearrangement","year":"2023","venue":"","arxiv_id":"2309.16118","aliases":["library:P188"],"technical_key":"","catalog_id":"","library_id":"P188","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2309.16118","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"D3Fields 将多视角视觉特征投影到三维描述场，用于零样本重排。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4b12b747cafccc1c508e7828bff0c8615b3aaa6d194c4c800a710a2a4555ad2b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0402","title":"Mini-BEHAVIOR: A Procedurally Generated Benchmark for Long-horizon Decision-Making in Embodied AI","year":"2023","venue":"","arxiv_id":"2310.01824","aliases":["library:P189"],"technical_key":"","catalog_id":"","library_id":"P189","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2310.01824","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Mini-BEHAVIOR 以程序化环境提供日常活动式规划问题，便于扩大任务变化。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"6be9ab5ca1ad254ca70e709cfdfe34325976025508fecb38672c09d69a55c831","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0403","title":"ZeroNVS: Zero-Shot 360-Degree View Synthesis from a Single Image","year":"2023","venue":"","arxiv_id":"2310.17994","aliases":["library:P192"],"technical_key":"","catalog_id":"","library_id":"P192","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2310.17994","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"ZeroNVS 训练三维感知扩散先验，处理场景数据混合与尺度模糊，并改善完整视角生成。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5f6b955ba52f7b24fbfb8c9766b2cf35d424d904b35115cdabe4a26710baab1a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0404","title":"NOIR: Neural Signal Operated Intelligent Robots for Everyday Activities","year":"2023","venue":"","arxiv_id":"2311.01454","aliases":["library:P193"],"technical_key":"","catalog_id":"","library_id":"P193","topic_code":"H1","topic":"主动交互、协作与用户需求","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2311.01454","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"NOIR 将脑信号接口与智能机器人技能组合，研究多类家庭活动控制。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"df055c5afbfd7ab1a6080d6577c3f34401e5ab2f15442fa42c5fb476c27e8dce","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0405","title":"Chain of Code: Reasoning with a Language Model-Augmented Code Emulator","year":"2023","venue":"","arxiv_id":"2312.04474","aliases":["library:P201"],"technical_key":"","catalog_id":"","library_id":"P201","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.04474","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Chain of Code 用代码结构组织推理，并让语言模型模拟不易直接执行的语义步骤。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5d3171a5b22b99e01e6957e6207c6e180ce9acefb7630ede52c5f26b1c98c073","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0406","title":"Photorealistic Video Generation with Diffusion Models","year":"2023","venue":"","arxiv_id":"2312.06662","aliases":["library:P202"],"technical_key":"","catalog_id":"","library_id":"P202","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.06662","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"W.A.L.T. 结合因果编码、窗口注意力和超分辨级联生成写实视频。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"aa3e7ee03466f6e8b5ef70bd1be451823461a5db6d6946f2bf8e97016c6b2c95","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0407","title":"Model-Based Control with Sparse Neural Dynamics","year":"2023","venue":"","arxiv_id":"2312.12791","aliases":["library:P205"],"technical_key":"","catalog_id":"","library_id":"P205","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.12791","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 2080 Ti / Titan RTX","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Sparse Neural Dynamics 模型训练；闭环优化另用 CPU / Gurobi","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2312.12791","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者剪除冗余 ReLU 神经元，再用混合整数规划和分支定界求解动作。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"fd53e8ac34fb07f6448f0a8e54ca3eb2924b8db2df93a756ac4b21e40d48bd50","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0408","title":"Agent AI: Surveying the Horizons of Multimodal Interaction","year":"2024","venue":"","arxiv_id":"2401.03568","aliases":["library:P210"],"technical_key":"","catalog_id":"","library_id":"P210","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2401.03568","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"V100","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":24.0,"wall_hours_max":24.0,"time_bound":"approximate","gpu_hours_min":384.0,"gpu_hours_max":384.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Agent AI 综述包含 250M Minecraft prototype 的实际训练；5K videos；不是纯无训练综述","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2401.03568","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Agent AI 综述梳理物理与虚拟环境中的交互系统和下一动作预测路线。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b3d0f83b0c928804504e09d73b4f32d000083bbe12ff14c5eff5389204394cf8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0409","title":"An Interactive Agent Foundation Model","year":"2024","venue":"","arxiv_id":"2402.05929","aliases":["library:P215"],"technical_key":"","catalog_id":"","library_id":"P215","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2402.05929","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"V100","count_min":192.0,"count_max":192.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":175.0,"wall_hours_max":175.0,"time_bound":"approximate","gpu_hours_min":33600.0,"gpu_hours_max":33600.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Interactive Agent Foundation Model；12 nodes × 16 V100；原文定义的 pretraining；不要混入下游 fine-tuning","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2402.05929","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.7（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Interactive Agent Foundation Model 结合视觉遮挡、语言建模和下一动作预测，研究机器人及游戏等任务。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"fd7a08182a563ad92d476d65769c544b80fcc5a39de28cf2ba3cefdccdedd615","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0410","title":"Position Paper: Agent AI Towards a Holistic Intelligence","year":"2024","venue":"","arxiv_id":"2403.00833","aliases":["library:P218"],"technical_key":"","catalog_id":"","library_id":"P218","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2403.00833","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Agent AI 立场论文讨论感知、基础模型与行动模型的整合路线。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a4694aa58cd7f1674aa19e8fa0fa9a50c84ca6a8f8f56ecf533dc75f8750dc9a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0411","title":"DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation","year":"2024","venue":"","arxiv_id":"2403.07788","aliases":["library:P220"],"technical_key":"","catalog_id":"","library_id":"P220","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2403.07788","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"DexCap 提供手部捕捉系统，DexIL 将人类动作数据用于机器人技能学习。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"08ec338188a4938ed7485e516a3908ce2caa372a8b4dffa78572d09b7d3e3fff","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0412","title":"BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation","year":"2024","venue":"","arxiv_id":"2403.09227","aliases":["library:P221"],"technical_key":"","catalog_id":"","library_id":"P221","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2403.09227","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 2080 Ti / RTX A6000","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":9.0,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"BEHAVIOR-1K 每次 baseline RL 训练；训练显存约 9GB；完整任务耗时范围另需定位","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2403.09227","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.30（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"BEHAVIOR-1K 提供大量人本活动与交互仿真，测试复杂操作及迁移。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d66ce6d43d5257ca82066b82af2f295eefe00430f22acf1442bd7567b58ab953","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0413","title":"BEHAVIOR Vision Suite: Customizable Dataset Generation via Simulation","year":"2024","venue":"","arxiv_id":"2405.09546","aliases":["library:P227"],"technical_key":"","catalog_id":"","library_id":"P227","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2405.09546","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"BEHAVIOR Vision Suite 支持改变场景、外观和物理相关参数，生成带定制标注的合成数据。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0c5f0f9859464ee3f96a635ffdf66e4e56f9726fca38cf5ec411681db33f7766","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0414","title":"TRANSIC: Sim-to-Real Policy Transfer by Learning from Online Correction","year":"2024","venue":"","arxiv_id":"2405.10315","aliases":["library:P229"],"technical_key":"","catalog_id":"","library_id":"P229","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2405.10315","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 3090","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"TRANSIC 约 3 天；GPU 数未明确，不计算 GPU-hours","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2405.10315","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.35（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"TRANSIC 将仿真策略与真实执行时的人类辅助结合，学习应对接触复杂任务中的迁移误差。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"6acef95ef7994ef211931a9c8b963f4367d1032185f0a3a4fe4baf13f497e877","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0415","title":"OccFusion: Rendering Occluded Humans with Generative Diffusion Priors","year":"2024","venue":"","arxiv_id":"2407.00316","aliases":["library:P238"],"technical_key":"","catalog_id":"","library_id":"P238","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2407.00316","status":"已核对资源阶段","resource_note":"","stages":[{"role":"opt","device":"Titan RTX","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":0.166667,"wall_hours_max":0.166667,"time_bound":"approximate","gpu_hours_min":0.1667,"gpu_hours_max":0.1667,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"OccFusion 每个人物的三阶段 3D appearance / geometry 优化约 10min；不是视频 diffusion backbone fine-tuning","comparable":false,"is_gpu":true,"role_label":"逐实例表示优化","source_url":"https://arxiv.org/abs/2407.00316","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.5（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"OccFusion 将三维 Gaussian 表示与二维扩散监督结合，改善遮挡条件的人体重建。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d767b6aafd464942e1b7234dd59b921b806e73f476d4a0e2c80f1eb8d17591b7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0416","title":"ReKep: Spatio-Temporal Reasoning of Relational Keypoint Constraints for Robotic Manipulation","year":"2024","venue":"","arxiv_id":"2409.01652","aliases":["library:P244"],"technical_key":"","catalog_id":"","library_id":"P244","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2409.01652","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"ReKep 自动生成视觉关键点约束，并通过分层优化求解末端位姿序列。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f1b46c93212b43458578038c718270352497f4eb6b1661866a6a12310f16d828","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0417","title":"Embodied Agent Interface: Benchmarking LLMs for Embodied Decision Making","year":"2024","venue":"","arxiv_id":"2410.07166","aliases":["library:P249"],"technical_key":"","catalog_id":"","library_id":"P249","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2410.07166","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Embodied Agent Interface 形式化任务和模块接口，超越单一终局成功率分析。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"6c0a7c93917aa790d7bce38794a00e7d91ec721454ac3f45399d18c075de845f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0418","title":"Automated Creation of Digital Cousins for Robust Policy Learning","year":"2024","venue":"","arxiv_id":"2410.07408","aliases":["library:P250"],"technical_key":"","catalog_id":"","library_id":"P250","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2410.07408","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Digital Cousins 自动生成保留几何与语义可供性的相近场景，组成 real-to-sim-to-real 训练。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b17c6b4b46ed5daaee598b29780e5d6da33e2ee77419dab071e344327748bfb5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0419","title":"HourVideo: 1-Hour Video-Language Understanding","year":"2024","venue":"","arxiv_id":"2411.04998","aliases":["library:P257"],"technical_key":"","catalog_id":"","library_id":"P257","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2411.04998","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"HourVideo 包含总结、回忆、追踪、推理和导航任务，提供长视频问答评价。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4b23bc346ba19b10c7a240de8888f61f854ea91f7161b553bf7ba5cb024398a0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0420","title":"The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion","year":"2024","venue":"","arxiv_id":"2412.10523","aliases":["library:P263"],"technical_key":"","catalog_id":"","library_id":"P263","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2412.10523","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"The Language of Motion 第一阶段 modality alignment full model；后训练阶段成本另缺","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2412.10523","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"The Language of Motion 用多模态语言模型连接语言、手势、情绪与动作编辑。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"1ddaa35c46f07a9620b0ea97e1865b864dd38db12b05dde380fc765040559f38","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0421","title":"Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces","year":"2024","venue":"","arxiv_id":"2412.14171","aliases":["library:P266"],"technical_key":"","catalog_id":"","library_id":"P266","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2412.14171","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Thinking in Space 构建 VSI-Bench 并用语言与视觉表达探查空间推理。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0652fca3ddfa1ba77f2b28b7adb0b29daded97aded3085feed8077fb3e84df5e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0422","title":"Why Automate This? Exploring Correlations Between Desire for Robotic Automation, Invested Time and Well-Being","year":"2025","venue":"","arxiv_id":"2501.06348","aliases":["library:P267"],"technical_key":"","catalog_id":"","library_id":"P267","topic_code":"H1","topic":"主动交互、协作与用户需求","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2501.06348","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者结合日常活动与福祉调查，分析自动化偏好和时间、情绪的关联。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"578bca1c7aa866d4ae4e62845824a8cceed0cc97f066c17649a64d896f3e534d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0423","title":"A Real-to-Sim-to-Real Approach to Robotic Manipulation with VLM-Generated Iterative Keypoint Rewards","year":"2025","venue":"","arxiv_id":"2502.08643","aliases":["library:P272"],"technical_key":"","catalog_id":"","library_id":"P272","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2502.08643","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"GPU（型号未定位）","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":0.083333,"wall_hours_max":0.083333,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Real-to-Sim-to-Real 每任务 policy 训练约 5min；硬件数量未定位，不能认定单卡 5min","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2502.08643","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"IKER 用 VLM 生成和修改 Python 奖励，在重建仿真中训练策略，再部署到真实环境。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"785e247843c1090db5e2ab7333f122106aa7f0b01bcc9e344303c4161083af70","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0424","title":"BEHAVIOR Robot Suite: Streamlining Real-World Whole-Body Manipulation for Everyday Household Activities","year":"2025","venue":"","arxiv_id":"2503.05652","aliases":["library:P280"],"technical_key":"","catalog_id":"","library_id":"P280","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.05652","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX A5000 / RTX 4090 / A40","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"BEHAVIOR Robot Suite WB-VIMA training；DDP；GPU 总数未定位","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2503.05652","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"BEHAVIOR Robot Suite 提供机器人系统、示范接口和学习方案，覆盖多种困难家庭活动。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ff587ebf142153b27b8e7d11393797a5b1f3a913cbbb38efef619033b7837058","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0425","title":"Repurposing 2D Diffusion Models with Gaussian Atlas for 3D Generation","year":"2025","venue":"","arxiv_id":"2503.15877","aliases":["library:P285"],"technical_key":"","catalog_id":"","library_id":"P285","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.15877","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Gaussian Atlas 将三维结构展开到稠密二维网格，构建 GaussianVerse 数据进行微调。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9d2af724ea71399f3cbdd17b0004e36c728dbad8c00b39c8a94b49e6dfc5f268","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0426","title":"Chain-of-Modality: Learning Manipulation Programs from Multimodal Human Videos with Vision-Language-Models","year":"2025","venue":"","arxiv_id":"2504.13351","aliases":["library:P289"],"technical_key":"","catalog_id":"","library_id":"P289","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2504.13351","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Chain-of-Modality 用多模态示范与 VLM 提取任务计划和控制参数。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4a48b8fbbce95e27983e119d0686e8177c2700c2b5a2e228d5a3309f5e14bd3a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0427","title":"RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learning","year":"2025","venue":"","arxiv_id":"2504.20073","aliases":["library:P292"],"technical_key":"","catalog_id":"","library_id":"P292","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2504.20073","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100 / A100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"RAGEN StarPO；100–200 rollout-update iterations；数量和总耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2504.20073","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.5（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"RAGEN 提供轨迹级训练框架并分析奖励、采样和思维监督的作用。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"198449adf1b79b4630859bd810c6b6089b2499a827f434d26667e2c89c5603ef","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0428","title":"MindCube: Spatial Mental Modeling from Limited Views","year":"2025","venue":"","arxiv_id":"2506.21458","aliases":["library:P302"],"technical_key":"","catalog_id":"","library_id":"P302","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2506.21458","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":2.0,"count_max":2.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"MindCube Qwen2.5-VL-3B full SFT；10K QA，3 epochs；batch 256","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2506.21458","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.59（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MindCube 设计空间问答，并通过先生成认知地图再推理的训练改善空间理解。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":2.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"79791cca82b99a76311960002da46e8d790b63b7c761e9a889e208e78ea357f9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0429","title":"VAGEN: Reinforcing World Model Reasoning for Multi-Turn VLM Agents","year":"2025","venue":"","arxiv_id":"2510.16907","aliases":["library:P308"],"technical_key":"","catalog_id":"","library_id":"P308","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2510.16907","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VAGEN 每 run 4–8h；服务器 8 H100 可同时两实验，未据此反推每 run 4 卡","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2510.16907","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.31（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"VAGEN 用强化学习约束视觉状态推理，并提供逐轮状态预测奖励。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"1e840f6c8536e03c0211945d271bd2c5f42af56213b90044d80b9be1e4d67efa","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0430","title":"MoMaGen: Generating Demonstrations under Soft and Hard Constraints for Multi-Step Bimanual Mobile Manipulation","year":"2025","venue":"","arxiv_id":"2510.18316","aliases":["library:P309"],"technical_key":"","catalog_id":"","library_id":"P309","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2510.18316","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"MoMaGen 下游 π0 3.3B；batch 64；设备型号与耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2510.18316","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MoMaGen 在软硬约束下扩展示范，综合安排底座、相机和多步操作。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"0eb10a937402b4f34311f4fe2517931b2cf4b3fdde34434744b9160576e58192","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0431","title":"ENACT: Evaluating Embodied Cognition with World Modeling of Egocentric Interaction","year":"2025","venue":"","arxiv_id":"2511.20937","aliases":["library:P316"],"technical_key":"","catalog_id":"","library_id":"P316","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2511.20937","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"ENACT 用 BEHAVIOR 合成的状态动作轨迹形成视觉问答，测试具身认知与世界建模。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4752299af48ab4d302b1546a83cbefa9f1ea828c09a03226cdc79e2fc55f5bd3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0432","title":"QuantiPhy: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models","year":"2025","venue":"","arxiv_id":"2512.19526","aliases":["library:P328"],"technical_key":"","catalog_id":"","library_id":"P328","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.19526","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"QuantiPhy 以数值真值评价尺寸、速度和加速度估计，并改变背景和反事实先验。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3f82a6791cc9985a96fb2c44adf0b632256db0e43b313ee174342204f17fd1bd","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0433","title":"Dream2Flow: Bridging Video Generation and Open-World Manipulation with 3D Object Flow","year":"2025","venue":"","arxiv_id":"2512.24766","aliases":["library:P330"],"technical_key":"","catalog_id":"","library_id":"P330","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.24766","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Dream2Flow 从生成视频恢复三维物体流，再以轨迹跟踪连接操作控制。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3c6da40ac1825814e4c85e77ac678450d6218c28fe3ada1114c776fc33819092","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0434","title":"PointWorld: Scaling 3D World Models for In-The-Wild Robotic Manipulation","year":"2026","venue":"","arxiv_id":"2601.03782","aliases":["library:P333"],"technical_key":"","catalog_id":"","library_id":"P333","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2601.03782","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H100","count_min":128.0,"count_max":128.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":480.0,"wall_hours_max":480.0,"time_bound":"approximate","gpu_hours_min":61440.0,"gpu_hours_max":61440.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"PointWorld-1B 从零世界模型训练；300 epochs，20 天；不计入 fine-tuning 排名","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2601.03782","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.28（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"PointWorld 从 RGB-D 与动作预测逐像素三维位移，并以真实、合成轨迹研究跨本体规模化训练。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4a90f0799551da74c366739c1ea5350b6009470e40e1532040ead2ae40b46a34","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0435","title":"Learning from Trials and Errors: Reflective Test-Time Planning for Embodied LLMs","year":"2026","venue":"","arxiv_id":"2602.21198","aliases":["library:P347"],"technical_key":"","catalog_id":"","library_id":"P347","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.21198","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Reflective Test-Time Planning 在行动前评分候选，行动后以外部反馈更新模型，并追溯早期决策。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"61e4cb23e3a253e28d92b63b1b8cd160646a7ee4834e40c74956e32a1909af9c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0436","title":"MIRAGE: The Illusion of Visual Understanding","year":"2026","venue":"","arxiv_id":"2603.21687","aliases":["library:P354"],"technical_key":"","catalog_id":"","library_id":"P354","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2603.21687","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"DGX Spark Blackwell","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"MIRAGE 的 fine-tuned model 实验；具体训练时长未定位；不把统一内存当 GPU 专属显存","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2603.21687","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MIRAGE 检查缺失图像时仍产生详细描述和高分的现象，并提出减少文本泄漏的评价方法。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"6736c270756c8b225ee67f5975291a05f8d770322837a6d40cbc4a3d139d210a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0437","title":"CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation","year":"2026","venue":"","arxiv_id":"2603.22435","aliases":["library:P356"],"technical_key":"","catalog_id":"","library_id":"P356","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2603.22435","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"CaP-X 统一代码策略评价，并研究多轮反馈、技能合成和强化学习改进。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"eb7e6ce37b4e71eefc0c273d112543659d78258b3b869b5c549a922f2a3eb101","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0438","title":"IMPASTO: Integrating Model-Based Planning with Learned Dynamics Models for Robotic Oil Painting Reproduction","year":"2026","venue":"","arxiv_id":"2603.29315","aliases":["library:P357"],"technical_key":"","catalog_id":"","library_id":"P357","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2603.29315","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX A5000 / RTX 3090 / A40","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"IMPASTO U-Net dynamics training；DDP，设备数量未定位；RTX 4090 为部署","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2603.29315","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.9（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"IMPASTO 将像素变化预测与模型规划结合进行机器人油画复现。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"dfe84256f2dfdacfcfa1ef58b9981e5aefb4c64999b4c8f5fe1e85580f058d8b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0439","title":"RAGEN-2: Reasoning Collapse in Agentic RL","year":"2026","venue":"","arxiv_id":"2604.06268","aliases":["library:P360"],"technical_key":"","catalog_id":"","library_id":"P360","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2604.06268","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"RAGEN-2 区分输出多样性与跨输入可辨别性，提出互信息代理诊断和信噪比筛选。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a9bb008bb1fa9251490042cd0d80b9c028824c634a0ffe9df5271b6ca7714058","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0440","title":"HumanScore: Benchmarking Human Motions in Generated Videos","year":"2026","venue":"","arxiv_id":"2604.20157","aliases":["library:P362"],"technical_key":"","catalog_id":"","library_id":"P362","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2604.20157","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"HumanScore 用多种运动强度提示及物理相关指标评估生成视频中的人体运动。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ff30b3253f1d5b55dbda4870455e443b87c017bbdb49fb728efed872307bbf54","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0441","title":"Planning with the Views","year":"2026","venue":"","arxiv_id":"2605.29563","aliases":["library:P376"],"technical_key":"","catalog_id":"","library_id":"P376","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.29563","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Planning with the Views；401 steps；FSDP / checkpointing；型号和时长未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2605.29563","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Planning with the Views 在 ViewSuite 中评估视角变换，通过自探索和视图图蒸馏改善规划。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"aece378fe262cb12b966787e00147977fc259bda8b50e4344036702e004638d6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0442","title":"SimFoundry: Modular and Automated Scene Generation for Policy Learning and Evaluation","year":"2026","venue":"","arxiv_id":"2606.28276","aliases":["library:P389"],"technical_key":"","catalog_id":"","library_id":"P389","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2606.28276","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"SimFoundry 以模块化 real-to-sim 管线生成场景，并比较仿真排序和真实执行结果。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ed74badff79def2faeb111ec10ed13692585a7c9b9eb0ee4e430ef57a533ab54","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0443","title":"RoboTTT: Context Scaling for Robot Policies","year":"2026","venue":"","arxiv_id":"2607.15275","aliases":["library:P394"],"technical_key":"","catalog_id":"","library_id":"P394","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2607.15275","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"GB200","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"RoboTTT 每个 pretraining run","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2607.15275","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.7（同文多阶段请结合上方范围）"},{"role":"ft","device":"GB200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"RoboTTT 每个 task-specific post-training run；与 pretraining 分开","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2607.15275","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.7（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"RoboTTT 用测试时训练机制扩展视觉运动上下文，研究示范学习、扰动恢复与多阶段任务。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"28742493cc54433769dd99eb52142a89d822ef9aaa9a5507b573fc1135795a08","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0444","title":"Masked Visual Actions for Unified World Modeling","year":"2026","venue":"","arxiv_id":"2607.19343","aliases":["library:P396","reasoning:NeurIPS-2026-4979"],"technical_key":"","catalog_id":"","library_id":"P396","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2607.19343","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":96.0,"wall_hours_max":96.0,"time_bound":"approximate","gpu_hours_min":768.0,"gpu_hours_max":768.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Masked Visual Actions；LoRA rank 256，10K steps，4 天","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2607.19343","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Wan-Fun-Control 2.2 14B；同一 VAE 编码条件，缺失区域置灰后拼接。LoRA rank 256，batch 4，8 张 H200，约 1 万步／4 天。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"dab5ef239da59df83eff17630c395fb6bfa530271dc7c95fd57b2bc204ccc94e","reasoning_id":"NeurIPS-2026-4979","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0445","title":"TrAct: Bridging Robot Control and Visual Prediction with Visual Tracks","year":"2026","venue":"","arxiv_id":"2608.24101","aliases":["library:P398"],"technical_key":"","catalog_id":"","library_id":"P398","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2608.24101","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"TrAct DROID+EgoDex 预训练 track 模型；30K steps，batch 64；下游适配另计","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2608.24101","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"π0.5 改为动作／轨迹联合 flow 训练；SVD 加轨迹 ControlNet。预训练用 DROID/EgoDex，轨迹来自校准投影和 CoTracker；动作模型报告 4 张 H100、3 万步。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"00691ce3b301531c80540cbf233f1522d49a52a9561f41b2ca81ca5b799b84de","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0446","title":"One Demonstration, Many Objects: Generalizing Manipulation via Local Contact Geometry","year":"2026","venue":"","arxiv_id":"2609.01938","aliases":["library:P400"],"technical_key":"","catalog_id":"","library_id":"P400","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2609.01938","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"DemoMimic 关注接触点附近的物体几何，学习更可泛化的灵巧操作策略。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"473d15293a705256bfd1ad3f26f7c74557acc68c29d5cce9341693595cc053d3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0447","title":"3D Video Loops from Asynchronous Input","year":"2023","venue":"","arxiv_id":"2303.05312","aliases":["library:P169"],"technical_key":"","catalog_id":"","library_id":"P169","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2303.05312","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Multi-Tile Video 以稀疏时空表示和两阶段优化构建视角一致的三维循环视频。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4bf37480fe2fd05fe38d6ca241688a2e1649a256fa39bd3f437da385bdb5ea09","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0448","title":"AvatarCraft: Transforming Text into Neural Human Avatars with Parameterized Shape and Pose Control","year":"2023","venue":"","arxiv_id":"2303.17606","aliases":["library:P171"],"technical_key":"","catalog_id":"","library_id":"P171","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2303.17606","status":"已核对资源阶段","resource_note":"","stages":[{"role":"opt","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":2.166667,"wall_hours_max":2.166667,"time_bound":"approximate","gpu_hours_min":2.1667,"gpu_hours_max":2.1667,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"AvatarCraft 每个 avatar：coarse 30min + fine 100min；优化 3D 表示，不训练 Stable Diffusion 底座","comparable":false,"is_gpu":true,"role_label":"逐实例表示优化","source_url":"https://arxiv.org/abs/2303.17606","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"AvatarCraft 用扩散先验优化人体几何和纹理，再通过参数化人体间的显式变形支持动画。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b9c7bd2f834e07adbba56e1df7bbf57fc307efe22fa22e5b0956d15fed41d613","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0449","title":"Learning Neural Duplex Radiance Fields for Real-Time View Synthesis","year":"2023","venue":"","arxiv_id":"2304.10537","aliases":["library:P172"],"technical_key":"","catalog_id":"","library_id":"P172","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2304.10537","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"将 NeRF 蒸馏到双层网格上的神经特征，结合屏幕空间卷积和多视图优化提升渲染效率。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0a40c23361bd47fdd4ea3aac0f2bbeda17251a6d9838025dbfa5c16ed1d3c74a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0450","title":"Text2NeRF: Text-Driven 3D Scene Generation with Neural Radiance Fields","year":"2023","venue":"","arxiv_id":"2305.11588","aliases":["library:P174"],"technical_key":"","catalog_id":"","library_id":"P174","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2305.11588","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Text2NeRF 结合文本到图像先验、单目深度与渐进场景补全更新，优化场景辐射场。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c0f69035c6dfaa56166dac4d2757151417b00479bacb5178a2b9e75bb7c85123","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0451","title":"VQ-NeRF: Neural Reflectance Decomposition and Editing with Vector Quantization","year":"2023","venue":"","arxiv_id":"2310.11864","aliases":["library:P191"],"technical_key":"","catalog_id":"","library_id":"P191","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2310.11864","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"VQ-NeRF 将连续分支与离散材料码结合，并提供材质编辑界面。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b2d939494fc0d521c8f7b7fab493d21754d5cd4586cf402786afde4197c6b1a0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0452","title":"HumanRef: Single Image to 3D Human Generation via Reference-Guided Diffusion","year":"2023","venue":"","arxiv_id":"2311.16961","aliases":["library:P196"],"technical_key":"","catalog_id":"","library_id":"P196","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2311.16961","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"HumanRef 采用参考引导的分数蒸馏和区域注意力，让生成的人体与输入外观对应。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d82c4b86cfde5dd5ae6e6fe8b628fab85f6583a93f2938c4ece48fae5286c154","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0453","title":"Mesh-Guided Neural Implicit Field Editing","year":"2023","venue":"","arxiv_id":"2312.02157","aliases":["library:P199"],"technical_key":"","catalog_id":"","library_id":"P199","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.02157","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"以可微网格提取、颜色映射和八叉树优化连接易编辑网格与辐射场，支持不同粒度的局部编辑。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"eab329f1fb0ae5ae59aef7e5e7248ab8bc0f0e687a7074ecf1278a585f0e03c7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0454","title":"CAD: Photorealistic 3D Generation via Adversarial Distillation","year":"2023","venue":"","arxiv_id":"2312.06663","aliases":["library:P203"],"technical_key":"","catalog_id":"","library_id":"P203","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.06663","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"V100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":288.0,"gpu_hours_max":288.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"CAD 图像条件配置：2D upsampler 1.5 天 + 联合 3D upsampler 1.5 天；256²","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2312.06663","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"CAD 比较多视图渲染与扩散先验的分布差异，以对抗学习支持单图条件三维合成与插值。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"97bc42bd45a34b68c941e9f0312e03dd9bee1200d740c039548c8cd7a43c8f8c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0455","title":"HeadArtist: Text-conditioned 3D Head Generation with Self Score Distillation","year":"2023","venue":"","arxiv_id":"2312.07539","aliases":["library:P204"],"technical_key":"","catalog_id":"","library_id":"P204","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.07539","status":"已核对资源阶段","resource_note":"","stages":[{"role":"opt","device":"RTX A6000","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":3.0,"wall_hours_max":3.0,"time_bound":"approximate","gpu_hours_min":3.0,"gpu_hours_max":3.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"HeadArtist 每个 3D head 的 geometry / texture 优化；15K / 20K iterations；不是视频模型适配","comparable":false,"is_gpu":true,"role_label":"逐实例表示优化","source_url":"https://arxiv.org/abs/2312.07539","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.6（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"HeadArtist 通过关键点引导的 ControlNet 先验与自分数蒸馏优化参数化头部。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1c7552a25bcfd5608930ed9161511f275c8250d147fbd2f908ac29e0f94fb9a7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0456","title":"Advances in 3D Generation: A Survey","year":"2024","venue":"","arxiv_id":"2401.17807","aliases":["library:P213"],"technical_key":"","catalog_id":"","library_id":"P213","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2401.17807","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"综述按前馈生成、优化生成、程序生成和生成式新视角合成组织三维文献。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"da4b32ffdc1a62a76bd21ef3b726c0f818fc2119c4bbe5ba35f1d4e6534ff7c8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0457","title":"RaFE: Generative Radiance Fields Restoration","year":"2024","venue":"","arxiv_id":"2404.03654","aliases":["library:P222"],"technical_key":"","catalog_id":"","library_id":"P222","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2404.03654","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"RaFE 先修复多视图图像，再通过生成式辐射场处理各视角修复结果的不一致。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"aa94b8d5b118e61302558e89d22de3eb68d6816f06dfcedd30c75d27c73444e5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0458","title":"AniClipart: Clipart Animation with Text-to-Video Priors","year":"2024","venue":"","arxiv_id":"2404.12347","aliases":["library:P225"],"technical_key":"","catalog_id":"","library_id":"P225","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2404.12347","status":"已核对资源阶段","resource_note":"","stages":[{"role":"opt","device":"RTX A6000","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":26.0,"wall_hours_min":0.416667,"wall_hours_max":0.416667,"time_bound":"approximate","gpu_hours_min":0.4167,"gpu_hours_max":0.4167,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"AniClipart 每个 clipart 的 Bézier motion 优化；500 steps，24 frames；视频 diffusion prior 冻结","comparable":false,"is_gpu":true,"role_label":"逐实例表示优化","source_url":"https://arxiv.org/abs/2404.12347","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"AniClipart 用贝塞尔曲线参数化关键点轨迹，并结合视频分数蒸馏与骨架保持损失优化动画。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9abedfc827f0c7d8701732b0cd765f2dde9edc836476221a714fd0c9f00c1526","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0459","title":"Learn2Talk: 3D Talking Face Learns from 2D Talking Face","year":"2024","venue":"","arxiv_id":"2404.12888","aliases":["library:P226"],"technical_key":"","catalog_id":"","library_id":"P226","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2404.12888","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 4090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":3.0,"wall_hours_max":5.0,"time_bound":"approximate","gpu_hours_min":3.0,"gpu_hours_max":5.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Learn2Talk 主 talking-face model；100 epochs；SyncNet3D 辅助模型另需约 12–15h，不能仅报主模型预算","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2404.12888","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Learn2Talk 从二维说话人方法获得唇音同步和教师监督，改进语音到三维面部运动的学习。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"7108b4654fce088ce20d6e021df3b9527dd7bd0c322a773926da787873930db3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0460","title":"EG4D: Explicit Generation of 4D Object without Score Distillation","year":"2024","venue":"","arxiv_id":"2405.18132","aliases":["library:P230"],"technical_key":"","catalog_id":"","library_id":"P230","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2405.18132","status":"已核对资源阶段","resource_note":"","stages":[{"role":"opt","device":"RTX 3090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"EG4D 每场景 4DGS 优化；5K steps；不是视频 diffusion 权重微调","comparable":false,"is_gpu":true,"role_label":"逐实例表示优化","source_url":"https://arxiv.org/abs/2405.18132","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"利用视频扩散生成多视图视频，再进行动态 Gaussian 重建与细节修整，减少对分数蒸馏的依赖。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e16c4f949feef760be33b43f9dbfa6c41d26a97bb70e6d40cb0e8caf82423fd9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0461","title":"Chat2Layout: Interactive 3D Furniture Layout with a Multimodal LLM","year":"2024","venue":"","arxiv_id":"2407.21333","aliases":["library:P239"],"technical_key":"","catalog_id":"","library_id":"P239","topic_code":"H3","topic":"人机共创与可编辑工作流","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2407.21333","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Chat2Layout 通过视觉文本提示和参考示例搜索引导 MLLM，利用反馈改善交互式家具布局，摘要声明无需修改模型权重。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"09056b0015bd5637ba4cca3c57a1157a5a6540d77075776cb30210eec8c09e49","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0462","title":"LVCD: Reference-based Lineart Video Colorization with Diffusion Models","year":"2024","venue":"","arxiv_id":"2409.12960","aliases":["library:P245"],"technical_key":"","catalog_id":"","library_id":"P245","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2409.12960","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX A6000","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":80.0,"wall_hours_max":80.0,"time_bound":"approximate","gpu_hours_min":320.0,"gpu_hours_max":320.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"LVCD 在 SVD 上适配；18K steps，576×320，batch 32","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2409.12960","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.5（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"LVCD 使用扩散模型和参考条件处理线稿视频上色，关注颜色与时序一致。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"8e684a51c4b8023d43db3867240673ee6cb47abcc14d07618f80351a9d42754b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0463","title":"Generative Object Insertion in Gaussian Splatting with a Multi-View Diffusion Model","year":"2024","venue":"","arxiv_id":"2409.16938","aliases":["library:P246"],"technical_key":"","catalog_id":"","library_id":"P246","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2409.16938","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A6000","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":30.0,"wall_hours_max":30.0,"time_bound":"approximate","gpu_hours_min":120.0,"gpu_hours_max":120.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"MVInpainter 从 SVD 初始化；50 epochs；每场景 Gaussian / SDS 优化成本另计","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2409.16938","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"通过多视图扩散先验实现 Gaussian Splatting 场景的生成式对象插入。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"87b40e46e6b2204741a6df81b5ac95b44f0049864d829ef3ac3889684cc4cc96","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0464","title":"SGEdit: Bridging LLM with Text2Image Generative Model for Scene Graph-based Image Editing","year":"2024","venue":"","arxiv_id":"2410.11815","aliases":["library:P253"],"technical_key":"","catalog_id":"","library_id":"P253","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2410.11815","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A6000","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":0.1,"wall_hours_max":0.166667,"time_bound":"approximate","gpu_hours_min":0.1,"gpu_hours_max":0.1667,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"SGEdit 每张图的 text embedding / model 适配约 6–10min；不是视频或通用编辑器完整训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2410.11815","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.7（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"SGEdit 将 LLM 与文本到图像模型通过场景图连接，支持结构化的对象及关系编辑。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"d42078721c31b6162868b9620b5633cd54ead3c65235035a2636a3efc8230a4c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0465","title":"Distractor-free Generalizable 3D Gaussian Splatting","year":"2024","venue":"","arxiv_id":"2411.17605","aliases":["library:P261"],"technical_key":"","catalog_id":"","library_id":"P261","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2411.17605","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 3090","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Distractor-free 3DGS re-training；10K iterations；数量与耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2411.17605","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.15（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"DGGS 在训练中预测并修整干扰掩码，在推理中重选参考图像并裁剪受干扰的 Gaussian 表示。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"b1e74bdb890461f73eb046b7f4a7ccfd47c69868c17ec1ef00de219a5f5ba189","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0466","title":"TalkingEyes: Pluralistic Speech-Driven 3D Eye Gaze Animation","year":"2025","venue":"","arxiv_id":"2501.09921","aliases":["library:P268"],"technical_key":"","catalog_id":"","library_id":"P268","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2501.09921","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 4090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":1.0,"wall_hours_max":1.0,"time_bound":"approximate","gpu_hours_min":1.0,"gpu_hours_max":1.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"TalkingEyes VQVAE pretraining；100 epochs；1h，已核对 HTML 数学标记","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2501.09921","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"pre","device":"RTX 4090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":5.0,"wall_hours_max":5.0,"time_bound":"approximate","gpu_hours_min":5.0,"gpu_hours_max":5.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"TalkingEyes VAE + autoregressive model；100 epochs，batch 1；TKED；不是视频 diffusion 微调","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2501.09921","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"TalkingEyes 使用音频与三维运动配对数据，在分开的潜空间中生成头部和视线，并与面部动画整合。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"193968a9d4a9867982b24bce9ca0dddf5fbd140009018a77bc304aeb153d1d17","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0467","title":"I2V3D: Controllable image-to-video generation with 3D guidance","year":"2025","venue":"","arxiv_id":"2503.09733","aliases":["library:P283"],"technical_key":"","catalog_id":"","library_id":"P283","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.09733","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"I2V3D 先用三维引导生成高质量关键帧，再采用双向引导的视频插值，支持相机、物体和角色动画控制。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"39f4a65eb267cf81cd6109aa66459a8132ed700a1c724d7e1bee9b2cf6609da0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0468","title":"MTV-Inpaint: Multi-Task Long Video Inpainting","year":"2025","venue":"","arxiv_id":"2503.11412","aliases":["library:P284"],"technical_key":"","catalog_id":"","library_id":"P284","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.11412","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MTV-Inpaint 以双分支空间注意力统一补全和插入，通过关键帧处理与中间帧传播扩展序列长度。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a72e244cd4eaf36df5bf8b75f9f0d6e696a77ef8e2ae26a89e99a3ea1fe16688","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0469","title":"AnimeGamer: Infinite Anime Life Simulation with Next Game State Prediction","year":"2025","venue":"","arxiv_id":"2504.01014","aliases":["library:P288"],"technical_key":"","catalog_id":"","library_id":"P288","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2504.01014","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"AnimeGamer 由 MLLM 根据历史预测含动作信息的多模态状态表示，再由视频扩散模型解码动画镜头。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b360f9e42c0d8ec42ca111ab55644dae66a07929038203e5f4158bb06e7732f7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0470","title":"CasaGPT: Cuboid Arrangement and Scene Assembly for Interior Design","year":"2025","venue":"","arxiv_id":"2504.19478","aliases":["library:P291"],"technical_key":"","catalog_id":"","library_id":"P291","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2504.19478","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A6000","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"CasaGPT scene layout model；3D-FRONT / 3DFRONT-NC；1,500 epochs","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2504.19478","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"CasaGPT 自回归排列长方体原语，使用拒绝采样降低对象碰撞，并整理场景布局数据。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c3590fa18b77fa22cb3f31121e1a5ed09361bda77fc8d8a2234d9b93f7771a8a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0471","title":"Animus3D: Text-driven 3D Animation via Motion Score Distillation","year":"2025","venue":"","arxiv_id":"2512.12534","aliases":["library:P325"],"technical_key":"","catalog_id":"","library_id":"P325","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.12534","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":1.0,"count_max":1.0,"card_memory_gb":24.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Animus3D ModelScopeT2V LoRA rank 4；3K steps；后续 motion distillation 5K + detailization 100 steps 另属于 3D 优化","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2512.12534","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Animus3D 以运动分数蒸馏优化运动场，利用 LoRA 静态分布、反演估噪及时空正则保持外观与运动细节。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"dd89f61bdc672dfa8598be706e811348a1f61b805d8789041ae976e60ae26fee","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0472","title":"Exploring a Multimodal Chatbot as a Facilitator in Therapeutic Art Activity","year":"2026","venue":"","arxiv_id":"2602.14183","aliases":["library:P344"],"technical_key":"","catalog_id":"","library_id":"P344","topic_code":"H3","topic":"人机共创与可编辑工作流","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.14183","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"系统分析创作画面并与创作者对话，报告五位相关专家的初步评价和后续设计问题。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"07651216dba3c0a4243767ea1f8ebbba6e2751d439690466bd7402fd1e9f26f1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0473","title":"Ani3DHuman: Photorealistic 3D Human Animation with Self-guided Stochastic Sampling","year":"2026","venue":"","arxiv_id":"2602.19089","aliases":["library:P345"],"technical_key":"","catalog_id":"","library_id":"P345","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.19089","status":"已核对资源阶段","resource_note":"","stages":[{"role":"opt","device":"A6000","count_min":1.0,"count_max":1.0,"card_memory_gb":48.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Ani3DHuman progressive 4D optimization 25K iterations；Wan2.1-1.3B 作为 pretrained prior","comparable":false,"is_gpu":true,"role_label":"逐实例表示优化","source_url":"https://arxiv.org/abs/2602.19089","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Ani3DHuman 将刚性骨架运动与残余非刚性运动分层，由粗渲染引导视频扩散，再通过自引导随机采样兼顾外观质量和身份保持。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e4c7977a9dc44f7dafab3aa200fc9582b9d9f0d16ae4fbbdf1589819ac28f67c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0474","title":"Cinematic Compositing Using Character-Environment-Harmonized Video Generation Models","year":"2026","venue":"","arxiv_id":"2606.20233","aliases":["library:P383"],"technical_key":"","catalog_id":"","library_id":"P383","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2606.20233","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"通过三类掩码和 RGB-D 联合去噪协调角色、交互道具与环境，并构造重光照训练配对。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"36efa6fb168c46db390b877d7a70fa07eac79da4d9346c05eeca72e13f77c9e8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0475","title":"BoxCtrl: 3D-Aware Visual Prompting for Geometric Image Editing","year":"2026","venue":"","arxiv_id":"2606.23270","aliases":["library:P385"],"technical_key":"","catalog_id":"","library_id":"P385","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2606.23270","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"BoxCtrl 以 RGB 三维框的二维投影作为视觉提示，先使用合成数据监督微调，再通过真实数据与几何、画质奖励进行强化学习。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1ec9b9cd3888519d296e3fd7cc412a116c21b5e7f38f84fc2fe511c4a91e924e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0476","title":"ThinkBLOX: 3D Indoor Scene Generation with Progressive Reasoning","year":"2026","venue":"","arxiv_id":"2607.13539","aliases":["library:P393"],"technical_key":"","catalog_id":"","library_id":"P393","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2607.13539","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX PRO 6000","count_min":8.0,"count_max":8.0,"card_memory_gb":96.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ThinkBLOX progressive scene reasoning RL；group 8；耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2607.13539","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.7（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"ThinkBLOX 用 VLM 逐步放置和修整场景，构建带多视图、推理说明和 JSON 布局的训练数据，并使用分层奖励优化。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"50a534cb26a6681afb5a91076a088072650aa4af99e726e311ade9301537dc89","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0477","title":"Return of Unconditional Generation: A Self-supervised Representation Generation Method","year":"2023","venue":"","arxiv_id":"2312.03701","aliases":["library:P200"],"technical_key":"","catalog_id":"","library_id":"P200","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.03701","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"V100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":24.0,"wall_hours_max":24.0,"time_bound":"upper","gpu_hours_min":1536.0,"gpu_hours_max":1536.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"RCG MAGE-B 从零生成训练少于 1 天，为上界；不是下游 fine-tuning","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2312.03701","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.7（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"RCG 先在自监督编码器的表征空间生成条件，再让图像生成器利用这些条件，降低无条件与有条件生成的差距。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2559bf658611f6264142beb55a50998a3a16a6e283a98572befc650138c47e31","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0478","title":"Deconstructing Denoising Diffusion Models for Self-Supervised Learning","year":"2024","venue":"","arxiv_id":"2401.14404","aliases":["library:P212"],"technical_key":"","catalog_id":"","library_id":"P212","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2401.14404","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"TPU v3 cores","count_min":256.0,"count_max":256.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":12.0,"wall_hours_max":12.0,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"DiT-L self-supervised representation learning；400 epochs；TPU core-hours 不与 GPU-hours 合并","comparable":false,"is_gpu":false,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2401.14404","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"逐项拆解去噪扩散模型并向去噪自编码器简化，对比其表征学习表现。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e2a7769b1c2f250dc835d2498e887a9a080fdea5ad088ef597b540b8f7f14a6b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0479","title":"TetSphere Splatting: Representing High-Quality Geometry with Lagrangian Volumetric Meshes","year":"2024","venue":"","arxiv_id":"2405.20283","aliases":["library:P232"],"technical_key":"","catalog_id":"","library_id":"P232","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2405.20283","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"TetSphere Splatting 通过四面体球的变形及几何正则构造高质量体网格，评价单视图和多视图重建。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f503720bcdaec2b0016e7c6f059051bc6a4229628e932a5510e9937d2ed5e430","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0480","title":"Physically Compatible 3D Object Modeling from a Single Image","year":"2024","venue":"","arxiv_id":"2405.20510","aliases":["library:P233"],"technical_key":"","catalog_id":"","library_id":"P233","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2405.20510","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"把静态平衡作为约束，将物理属性分解并连接到三维重建优化中，使模型更适合仿真和制造。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3040c3cf3a9344ca71b1338cac8a6c09e3108ec760beba6a10addf9647303a68","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0481","title":"Autoregressive Image Generation without Vector Quantization","year":"2024","venue":"","arxiv_id":"2406.11838","aliases":["library:P236"],"technical_key":"","catalog_id":"","library_id":"P236","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2406.11838","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"V100","count_min":128.0,"count_max":128.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":62.4,"wall_hours_max":62.4,"time_bound":"approximate","gpu_hours_min":7987.2,"gpu_hours_max":7987.2,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"MAR-L 400 epochs；16 servers × 8 V100；约 2.6 天","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2406.11838","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"通过扩散过程建模连续 token 的条件分布，将 Diffusion Loss 与自回归或掩码自回归架构结合。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ba0902147ab0ce247514038243f0fae99acc731fb75beba34321be0184697e16","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0482","title":"Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers","year":"2024","venue":"","arxiv_id":"2409.20537","aliases":["library:P248"],"technical_key":"","catalog_id":"","library_id":"P248","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2409.20537","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 2080 Ti","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":4.0,"wall_hours_max":4.0,"time_bound":"approximate","gpu_hours_min":4.0,"gpu_hours_max":4.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"HPT 下游小 stem/head 适配；trunk 预训练另计；20K iterations，batch 256","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2409.20537","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"HPT 用可共享 Transformer 主干连接不同本体的视觉和本体感知数据，研究异构预训练的扩展规律。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"916b83ffd17b6954235c6765b2de8dfa997588d83cf1d45fb345d6509f0240d6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0483","title":"Fluid: Scaling Autoregressive Text-to-image Generative Models with Continuous Tokens","year":"2024","venue":"","arxiv_id":"2410.13863","aliases":["library:P254"],"technical_key":"","catalog_id":"","library_id":"P254","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2410.13863","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Fluid 对比离散或连续 token、固定或随机顺序，并采用连续 token 的随机顺序自回归生成。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"80c1be5c679ba115b6f0fc0ab5c564ec37f4d8ed57b4069a4b508b5a98602d4e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0484","title":"Is Noise Conditioning Necessary for Denoising Generative Models?","year":"2025","venue":"","arxiv_id":"2502.13129","aliases":["library:P275"],"technical_key":"","catalog_id":"","library_id":"P275","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2502.13129","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"研究去除噪声条件后的生成性能与误差，并给出理论和图像实验分析。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"febad3905e249296385c845878094e4f84279393710c279fdb7d14fdc21c2fab","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0485","title":"Fractal Generative Models","year":"2025","venue":"","arxiv_id":"2502.17437","aliases":["library:P277"],"technical_key":"","catalog_id":"","library_id":"P277","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2502.17437","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Fractal Generative Models 通过递归调用生成模块形成自相似结构，并以逐像素自回归图像生成进行验证。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"dafa2e2c8a6aee546b6b4a03a6230e9ae094de32f9c8a45fa02863a39527a85b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0486","title":"Denoising Hamiltonian Network for Physical Reasoning","year":"2025","venue":"","arxiv_id":"2503.07596","aliases":["library:P282"],"technical_key":"","catalog_id":"","library_id":"P282","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.07596","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Denoising Hamiltonian Network 将物理算子推广成可学习的神经算子，覆盖前向模拟之外的物理推理任务。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"55725dbd09cd5710820ac97daa299fbc4b7b010be60efbb3b36af7eb6a78cee8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0487","title":"Mean Flows for One-step Generative Modeling","year":"2025","venue":"","arxiv_id":"2505.13447","aliases":["library:P293"],"technical_key":"","catalog_id":"","library_id":"P293","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2505.13447","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MeanFlow 以区间平均速度代替仅建模瞬时速度，构建不依赖预训练蒸馏的一步生成框架。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"540b84586f3c5b2319d6cee815141e326107dc1503b8d70d8c6bd77d874c92fb","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0488","title":"Diffuse and Disperse: Image Generation with Representation Regularization","year":"2025","venue":"","arxiv_id":"2506.09027","aliases":["library:P298"],"technical_key":"","catalog_id":"","library_id":"P298","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2506.09027","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Dispersive Loss 鼓励隐藏表征分散，不要求正样本配对或外部表征教师，作为生成训练的附加正则。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e88b54b2f6097916b97f834849c55b02966dff045b9cb189e931f74457b8ff90","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0489","title":"Back to Basics: Let Denoising Generative Models Denoise","year":"2025","venue":"","arxiv_id":"2511.13720","aliases":["library:P312"],"technical_key":"","catalog_id":"","library_id":"P312","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2511.13720","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"JiT 以大块像素 Transformer 直接预测干净数据，研究去噪目标与流形假设，避免依赖额外 tokenizer。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f8ddeac25f8edebee1d82f83fa09f1b54b4b445ebf00853103097592742a486f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0490","title":"ARC Is a Vision Problem!","year":"2025","venue":"","arxiv_id":"2511.14761","aliases":["library:P313"],"technical_key":"","catalog_id":"","library_id":"P313","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2511.14761","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":4.8,"wall_hours_max":4.8,"time_bound":"approximate","gpu_hours_min":38.4,"gpu_hours_max":38.4,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"ARC ViT-18M offline training","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2511.14761","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":0.019444,"wall_hours_max":0.019444,"time_bound":"approximate","gpu_hours_min":0.0194,"gpu_hours_max":0.0194,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"ARC test-time training：100 epochs × 0.7s/epoch = 70s；仅适配阶段；离线模型另 8 H100×4.8h","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2511.14761","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Vision ARC 把 ARC 网格任务改写为画布上的图像到图像映射，从 ARC 数据训练，并通过测试时训练适应新任务。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"b2726446f18e3ae97f192d3bd7e78fa39fa7d1d2ffff00d370305df46491f685","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0491","title":"Improved Mean Flows: On the Challenges of Fastforward Generative Models","year":"2025","venue":"","arxiv_id":"2512.02012","aliases":["library:P318"],"technical_key":"","catalog_id":"","library_id":"P318","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.02012","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Improved MeanFlow 用平均速度参数化瞬时速度回归，并把引导作为显式条件，改善从零训练的一步生成。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5d271dfc49dcb8bac8559dda5adc1be08877ccfa8da426bddd36281e6decf72c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0492","title":"Bidirectional Normalizing Flow: From Data to Noise and Back","year":"2025","venue":"","arxiv_id":"2512.10953","aliases":["library:P324"],"technical_key":"","catalog_id":"","library_id":"P324","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.10953","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"BiFlow 放宽精确解析逆的要求，通过双向归一化流改进生成质量和采样效率。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a7d5caaf97dcafa93ec161a48cf5119c39425a93cc94ccaaa505efbe3a6558cb","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0493","title":"One-step Latent-free Image Generation with Pixel Mean Flows","year":"2026","venue":"","arxiv_id":"2601.22158","aliases":["library:P336"],"technical_key":"","catalog_id":"","library_id":"P336","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2601.22158","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Pixel MeanFlow 区分网络输出空间和损失空间，在图像流形与平均速度之间建立转换。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3070cd41ccdbb686d88ba835ed6396f776026ec4dbdc1224b190c4ff1db10511","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0494","title":"Generative Modeling via Drifting","year":"2026","venue":"","arxiv_id":"2602.04770","aliases":["library:P340"],"technical_key":"","catalog_id":"","library_id":"P340","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.04770","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Drifting Models 通过漂移场推动生成样本分布与数据分布匹配，形成单步推理的生成器。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1780eb7966d5577a70c9565441e99ec3ae3cd4fc7e8b7e00b291d4d066e3f9b7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0495","title":"GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training","year":"2026","venue":"","arxiv_id":"2602.20399","aliases":["library:P346"],"technical_key":"","catalog_id":"","library_id":"P346","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.20399","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":null,"count_max":null,"card_memory_gb":40.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"GeoPT 物理模拟预训练实验；GPU 数与耗时未定位","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2602.20399","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.19（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"GeoPT 用随机速度、粒子和动态距离等构造预训练信号，把静态几何数据扩展为动态表征学习。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c6b017cd248c10b8cb477e14c99bcc32c57e98b02fd6d44f8df715bafde5195f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0496","title":"Image Generators are Generalist Vision Learners","year":"2026","venue":"","arxiv_id":"2604.20329","aliases":["library:P363"],"technical_key":"","catalog_id":"","library_id":"P363","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2604.20329","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Vision Banana 将多种感知任务的输出编码成 RGB 图像，以图像生成骨干进行统一指令微调。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"368f30d6d5b257e622050aba3e2bd4c295422bc218d5a7fc7037b7b7d2c2cbe5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0497","title":"Video Generation Models are General-Purpose Vision Learners","year":"2026","venue":"","arxiv_id":"2607.09024","aliases":["library:P392"],"technical_key":"","catalog_id":"","library_id":"P392","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2607.09024","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"TPU v6e","count_min":256.0,"count_max":256.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Video General-Purpose Vision Learners；15K steps；batch 64；TPU 设备独立呈现","comparable":true,"is_gpu":false,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2607.09024","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"将视频扩散骨干改成干净 latent、固定 t=0 的单次前向感知模型；稠密任务映射到 RGB，稀疏任务加可学习 token，主要使用合成监督。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2524ac8d312220c6175f6d325886dacda8d751cd66b94953898a5109396d842d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0498","title":"Affordance Grounding from Demonstration Video to Target Image","year":"2023","venue":"","arxiv_id":"2303.14644","aliases":["library:P170"],"technical_key":"","catalog_id":"","library_id":"P170","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2303.14644","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Afformer 逐步细化 affordance 定位，并用 MaskAHand 自监督预训练模拟上下文变化。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"abd0cb8d3091e99d01ca6cbaf33aea459171e4295f721bf411c5b345663954cd","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0499","title":"HOSNeRF: Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video","year":"2023","venue":"","arxiv_id":"2304.12281","aliases":["library:P173"],"technical_key":"","catalog_id":"","library_id":"P173","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2304.12281","status":"已核对资源阶段","resource_note":"","stages":[{"role":"opt","device":"V100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":120.0,"wall_hours_max":120.0,"time_bound":"approximate","gpu_hours_min":480.0,"gpu_hours_max":480.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"HOSNeRF 每场景三阶段表示优化约 5 天；不是预训练视频模型 fine-tuning","comparable":false,"is_gpu":true,"role_label":"逐实例表示优化","source_url":"https://arxiv.org/abs/2304.12281","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"HOSNeRF 用对象骨骼与状态嵌入组织人体交互和场景表示。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"376863fa5d3d74f44cd66ffe57783609891bc94afac1eaf35ddef1bfbbece344","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0500","title":"EgoVLPv2: Egocentric Video-Language Pre-training with Fusion in the Backbone","year":"2023","venue":"","arxiv_id":"2307.05463","aliases":["library:P181"],"technical_key":"","catalog_id":"","library_id":"P181","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2307.05463","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":120.0,"wall_hours_max":120.0,"time_bound":"approximate","gpu_hours_min":3840.0,"gpu_hours_max":3840.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"EgoVLPv2 pretraining；20 epochs，batch 256；下游 fine-tuning 另计","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2307.05463","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"EgoVLPv2 将跨模态融合直接放入视频和语言骨干以支持不同下游任务。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"6eb39fedc39efc4952cdf9f9ac7c3264b221cdd9a405db12c73c37f8649f66ca","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0501","title":"Show-1: Marrying Pixel and Latent Diffusion Models for Text-to-Video Generation","year":"2023","venue":"","arxiv_id":"2309.15818","aliases":["library:P187"],"technical_key":"","catalog_id":"","library_id":"P187","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2309.15818","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":64.0,"count_max":64.0,"card_memory_gb":40.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Show-1 由图像模型扩展到视频的训练基础设施；WebVid-10M；未定位各模块总耗时","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2309.15818","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.8（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Show-1 先生成低分辨率像素视频，再以潜扩散放大并修复。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"8e934576b4621b11db625cbdb5d086f3e1c99f259ee8c70e32b6f979933fed30","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0502","title":"MotionDirector: Motion Customization of Text-to-Video Diffusion Models","year":"2023","venue":"","arxiv_id":"2310.08465","aliases":["library:P402"],"technical_key":"","catalog_id":"","library_id":"P402","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2310.08465","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MotionDirector 用双路径 LoRA 和去外观偏差的时间损失学习运动定制。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e221bffdb2a4f7617cd1672d5f12fd923df216db6f6b27843ed46d60208cdd59","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0503","title":"DynVideo-E: Harnessing Dynamic NeRF for Large-Scale Motion- and View-Change Human-Centric Video Editing","year":"2023","venue":"","arxiv_id":"2310.10624","aliases":["library:P190"],"technical_key":"","catalog_id":"","library_id":"P190","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2310.10624","status":"已核对资源阶段","resource_note":"","stages":[{"role":"opt","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"DynVideo-E 每场景 dynamic NeRF 编辑；包含 personalized diffusion prior；未定位各部分时长","comparable":false,"is_gpu":true,"role_label":"逐实例表示优化","source_url":"https://arxiv.org/abs/2310.10624","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"DynVideo-E 在动态 NeRF 中编辑，并通过形变场传播，结合多视角多姿态扩散先验。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c32ea2541351dc6547da4533251ba69fa0f6aec7f6a487339eb4da03c8a1ba2d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0504","title":"XAGen: 3D Expressive Human Avatars Generation","year":"2023","venue":"","arxiv_id":"2311.13574","aliases":["library:P195"],"technical_key":"","catalog_id":"","library_id":"P195","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2311.13574","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"XAGen 采用多尺度、多部位表示与渲染、判别设计实现表情和姿态控制。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"fa330da9272ca68b68fe724158b2a0f78de4c9a33024da00214aefba15f11bbe","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0505","title":"MagicAnimate: Temporally Consistent Human Image Animation using Diffusion Model","year":"2023","venue":"","arxiv_id":"2311.16498","aliases":["library:P403"],"technical_key":"","catalog_id":"","library_id":"P403","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2311.16498","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MagicAnimate 使用视频扩散、外观编码与片段融合改善人物动画。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a67dbd7dc82156e04e70805efa2c811085a0753c59a0658cae27d99ad75f0821","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0506","title":"Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives","year":"2023","venue":"","arxiv_id":"2311.18259","aliases":["library:P197"],"technical_key":"","catalog_id":"","library_id":"P197","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2311.18259","status":"已核对资源阶段","resource_note":"","stages":[{"role":"aux","device":"RTX 4090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":4.0,"wall_hours_max":4.0,"time_bound":"approximate","gpu_hours_min":4.0,"gpu_hours_max":4.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Ego-Exo4D THOR-net pose baseline manual annotations；manual+auto 为 10h；仅一个辅助 benchmark 模块","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/abs/2311.18259","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.43（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Ego-Exo4D 提供第一／第三视角活动数据及技能、姿态和视角转换任务。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"473a14e71968440c9afeb0dd16a6dd5379347169b0c21d8accadd544cdf1463f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0507","title":"DeformGS: Scene Flow in Highly Deformable Scenes for Deformable Object Manipulation","year":"2023","venue":"","arxiv_id":"2312.00583","aliases":["library:P404"],"technical_key":"","catalog_id":"","library_id":"P404","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.00583","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"DeformGS 通过 Gaussian 表示和动量、等距启发正则估计场景流。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4381c80e992f9d9145db6603c1e7bc6ecbb5ed205b2cf384f035f32daf3951e5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0508","title":"VideoSwap: Customized Video Subject Swapping with Interactive Semantic Point Correspondence","year":"2023","venue":"","arxiv_id":"2312.02087","aliases":["library:P198"],"technical_key":"","catalog_id":"","library_id":"P198","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.02087","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":40.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VideoSwap source video semantic point adapter；16 frames；单视频适配，不是通用生成器训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2312.02087","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"VideoSwap 用稀疏语义对应及拖动、删除交互控制主体身份和运动。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"15d25cd0b961e9e8210e0982ac42603b3d2dbcea441f659ea78ac0bed3ace271","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0509","title":"ASSISTGUI: Task-Oriented Desktop Graphical User Interface Automation","year":"2023","venue":"","arxiv_id":"2312.13108","aliases":["library:P206"],"technical_key":"","catalog_id":"","library_id":"P206","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.13108","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"AssistGUI 建立 Windows 软件任务并用 GUI 解析与 actor-critic 推理组织执行。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e256cdd0f4913b9cbcabebece472e524209cd863ac7878af0c7bbe6c856a9217","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0510","title":"ShowRoom3D: Text to High-Quality 3D Room Generation Using 3D Priors","year":"2023","venue":"","arxiv_id":"2312.13324","aliases":["library:P207"],"technical_key":"","catalog_id":"","library_id":"P207","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.13324","status":"已核对资源阶段","resource_note":"","stages":[{"role":"opt","device":"RTX 3090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ShowRoom3D 三阶段 NeRF 优化；10K / 15K / 5K iterations；不是 diffusion backbone fine-tuning","comparable":false,"is_gpu":true,"role_label":"逐实例表示优化","source_url":"https://arxiv.org/abs/2312.13324","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"ShowRoom3D 结合多视角扩散先验、渐进视角选择和姿态变换优化场景。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"6877a09b8ed833c679b767dfe3a9a60c5fc4bb58d780365e6c75bd57deeb6298","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0511","title":"Moonshot: Towards Controllable Video Generation and Editing with Multimodal Conditions","year":"2024","venue":"","arxiv_id":"2401.01827","aliases":["library:P209"],"technical_key":"","catalog_id":"","library_id":"P209","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2401.01827","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":16.0,"count_max":16.0,"card_memory_gb":40.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Moonshot 视频生成与编辑训练；512×320；去水印 refinement 另用 1K videos","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2401.01827","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Moonshot 用多模态视频模块与解耦交叉注意力整合条件，并接入预训练 ControlNet。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"b7e215eee1774c36dd166ea8e72ff1cf3176f605a0776b3bee0be8ad6fac9438","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0512","title":"Towards A Better Metric for Text-to-Video Generation","year":"2024","venue":"","arxiv_id":"2401.07781","aliases":["library:P211"],"technical_key":"","catalog_id":"","library_id":"P211","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2401.07781","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"T2VScore 分开建模文本视频对齐与视频质量，并用人工判断数据检验指标。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b3e403f7be7335a8fa19ae00c2543d1ffb3a8a0b4b94bce88e25cd87d4656e50","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0513","title":"Bring Your Own Character: A Holistic Solution for Automatic Facial Animation Generation of Customized Characters","year":"2024","venue":"","arxiv_id":"2402.13724","aliases":["library:P216"],"technical_key":"","catalog_id":"","library_id":"P216","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2402.13724","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 3060","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":0.5,"wall_hours_max":0.5,"time_bound":"approximate","gpu_hours_min":0.5,"gpu_hours_max":0.5,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Bring Your Own Character 每个 3D character 拓扑 adapter；约 30min；非视频生成器","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2402.13724","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.7（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Bring Your Own Character 预测表情系数进行面部重定向，并通过人在环反馈改进定制。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"eedefc47ce455258791124df7dc6b670d93f57ba21a66a5661b32a6b883bcaf0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0514","title":"DragAnything: Motion Control for Anything using Entity Representation","year":"2024","venue":"","arxiv_id":"2403.07420","aliases":["library:P219"],"technical_key":"","catalog_id":"","library_id":"P219","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2403.07420","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"DragAnything SVD 适配；100K steps；设备数/时长未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2403.07420","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"DragAnything 将实体表征与运动引导结合，支持对象和背景的轨迹控制。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"a0127a13f5cb26f7d477d43b1088da6b80e2144a2756b65d3e3bfad02988f853","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0515","title":"VideoGUI: A Benchmark for GUI Automation from Instructional Videos","year":"2024","venue":"","arxiv_id":"2406.10227","aliases":["library:P234"],"technical_key":"","catalog_id":"","library_id":"P234","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2406.10227","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"VideoGUI 以视频目标评估高层程序规划、中层动作叙述与原子执行。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"beda3470d737b57d09ab21e9e8280314e4f75d2815228be8c8a18549fc862d5e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0516","title":"VideoLLM-online: Online Video Large Language Model for Streaming Video","year":"2024","venue":"","arxiv_id":"2406.11816","aliases":["library:P235"],"technical_key":"","catalog_id":"","library_id":"P235","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2406.11816","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"VideoLLM-online 以 LIVE 目标、流式数据构造和推理优化支持时间对齐对话。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2475152b9c79b8d29c3068f96be3034b42f8e60b296b3923a6f05b019587006c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0517","title":"Learning Video Context as Interleaved Multimodal Sequences","year":"2024","venue":"","arxiv_id":"2407.21757","aliases":["library:P240"],"technical_key":"","catalog_id":"","library_id":"P240","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2407.21757","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MovieSeq 用交错多模态序列与指令微调连接画面、人物、情节和字幕。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9598c87792d534b8c348c657b5018855e342af0f18b57db024e39d0dd88bbc5c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0518","title":"Show-o: One Single Transformer to Unify Multimodal Understanding and Generation","year":"2024","venue":"","arxiv_id":"2408.12528","aliases":["library:P242"],"technical_key":"","catalog_id":"","library_id":"P242","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2408.12528","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":48.0,"count_max":48.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Show-o base multimodal training；500K + 1M steps；最终 high-quality fine-tuning 另缺","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2408.12528","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.21（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Show-o 在共享模型内结合文本建模与视觉生成，研究统一多模态能力。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"fbd9addc85bc61a8a274d59d15542ef7b3b41bad678cdcd3f7f47dfa88a8decb","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0519","title":"VideoLLM-MoD: Efficient Video-Language Streaming with Mixture-of-Depths Vision Computation","year":"2024","venue":"","arxiv_id":"2408.16730","aliases":["library:P243"],"technical_key":"","catalog_id":"","library_id":"P243","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2408.16730","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VideoLLM-MoD；原文明确所有模型用 8 A100；时长未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2408.16730","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"VideoLLM-MoD 让部分视觉 token 跳过若干层，减少长视频处理计算。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"5e1f07fc2a91e3891af4482a064134b71df870a297a328cde0d7c2af3f61d347","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0520","title":"High Quality Human Image Animation using Regional Supervision and Motion Blur Condition","year":"2024","venue":"","arxiv_id":"2409.19580","aliases":["library:P247"],"technical_key":"","catalog_id":"","library_id":"P247","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2409.19580","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":70.0,"wall_hours_max":70.0,"time_bound":"approximate","gpu_hours_min":560.0,"gpu_hours_max":560.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Human Animation spatial stage；35K iterations","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2409.19580","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":8.0,"wall_hours_max":8.0,"time_bound":"approximate","gpu_hours_min":64.0,"gpu_hours_max":64.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Human Animation appearance encoder regional-supervision fine-tuning；4K iterations；还有 temporal stages，不代表完整预算","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2409.19580","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者增加关键区域监督、显式运动模糊条件和高分辨率训练策略。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"ae738f432628a95f421e40c88b197be60c895e755833fa6e95192ef25c3a4524","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0521","title":"ControLRM: Fast and Controllable 3D Generation via Large Reconstruction Model","year":"2024","venue":"","arxiv_id":"2410.09592","aliases":["library:P252"],"technical_key":"","catalog_id":"","library_id":"P252","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2410.09592","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"V100","count_min":16.0,"count_max":16.0,"card_memory_gb":32.0,"peak_memory_gb":null,"wall_hours_min":96.0,"wall_hours_max":144.0,"time_bound":"approximate","gpu_hours_min":1536.0,"gpu_hours_max":2304.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"ControLRM-T cross-attention 适配；4–6 天；ControLRM-D 5–6 天；不同变体不相加","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2410.09592","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"ControLRM 通过条件分支和图像分支联合训练，连接二维控制与三维表示。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"4f0b124f983bd270d2c17fd9eff40b8abc005f6d41d6bdb40c33610084cf20e6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0522","title":"Skinned Motion Retargeting with Dense Geometric Interaction Perception","year":"2024","venue":"","arxiv_id":"2410.20986","aliases":["library:P255"],"technical_key":"","catalog_id":"","library_id":"P255","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2410.20986","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX A6000","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Skinned Motion Retargeting 新 motion model；30-frame clips；耗时未定位","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2410.20986","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MeshRet 用语义传感点对应和稠密网格交互场建模动作重定向。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3f8e969fdfa869d674da9850b41b4cbb696b4db2298d1aa530a0fdef62638b4c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0523","title":"ReCapture: Generative Video Camera Controls for User-Provided Videos using Masked Video Fine-Tuning","year":"2024","venue":"","arxiv_id":"2411.05003","aliases":["library:P258"],"technical_key":"","catalog_id":"","library_id":"P258","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2411.05003","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":0.083333,"wall_hours_max":0.083333,"time_bound":"approximate","gpu_hours_min":0.0833,"gpu_hours_max":0.0833,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"ReCapture 每个输入视频 spatial / temporal LoRA；400 steps，5 分钟；不是通用 camera-control 训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2411.05003","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"ReCapture 先构造新视角的带噪锚定视频，再以掩码视频微调修复。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"d598b3bbc355ef1f7050e75b9b59a152dc5e7bb49379f510380a1185fab2d007","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0524","title":"ShowUI: One Vision-Language-Action Model for GUI Visual Agent","year":"2024","venue":"","arxiv_id":"2411.17465","aliases":["library:P260"],"technical_key":"","catalog_id":"","library_id":"P260","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2411.17465","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"V100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ShowUI instruction-tuning；下游 adaptation 用 8 V100；LoRA，时长未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2411.17465","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"V100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ShowUI downstream adaptation；与 32 卡 instruction-tuning 分阶段保留","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2411.17465","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"ShowUI 结合界面引导 token 选择、交错状态动作流与高质量指令数据。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"0e3a0247ced07f5150c0a173ecc6c3856f9baa7dbe092e7b4a8cb605f0ac07d2","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0525","title":"MakeAnything: Harnessing Diffusion Transformers for Multi-Domain Procedural Sequence Generation","year":"2025","venue":"","arxiv_id":"2502.01572","aliases":["library:P269"],"technical_key":"","catalog_id":"","library_id":"P269","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2502.01572","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MakeAnything 以多领域过程序列微调 DiT，并用非对称 LoRA 和时空约束生成或反推制作过程。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"7fc79b26b27f6ca23afa3789e934701da7d1a0f901f70fd50a63235f5f2252b1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0526","title":"UniMoD: Efficient Unified Multimodal Transformers with Mixture-of-Depths","year":"2025","venue":"","arxiv_id":"2502.06474","aliases":["library:P270"],"technical_key":"","catalog_id":"","library_id":"P270","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2502.06474","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"UniMoD 在 Show-o 上 fine-tune；LLaVA-v1.5-mix-665K + image data；80% token pruning","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2502.06474","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.7（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"UniMoD 用任务感知路由选择跳过计算的 token，比较训练成本和多模态任务表现。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"5d46b7630071007b3c7127cf413157346bfec59228818baba6d0c5fd87380165","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0527","title":"WorldGUI: An Interactive Benchmark for Desktop GUI Automation from Any Starting Point","year":"2025","venue":"","arxiv_id":"2502.08047","aliases":["library:P271"],"technical_key":"","catalog_id":"","library_id":"P271","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2502.08047","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"WorldGUI 系统改变软件任务初始状态，用批评式执行框架比较恢复与适应能力。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b8dc1f4de030c5d8fad8e363e4e9b8912df8a21b1519684231b7974d76138519","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0528","title":"PhysReason: A Comprehensive Benchmark towards Physics-Based Reasoning","year":"2025","venue":"","arxiv_id":"2502.12054","aliases":["library:P274"],"technical_key":"","catalog_id":"","library_id":"P274","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2502.12054","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"PhysReason 构建物理题及答案、步骤级评分，分析定理使用、条件判断与计算问题。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d78157aaf8ec30a5060d791e5baef5f15dc47dd723da8ed32625d8e32b5bb23a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0529","title":"Difix3D+: Improving 3D Reconstructions with Single-Step Diffusion Models","year":"2025","venue":"","arxiv_id":"2503.01774","aliases":["library:P278"],"technical_key":"","catalog_id":"","library_id":"P278","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.01774","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Difix3D+ 用单步图像扩散修复渲染结果，并将修复信号反馈到三维表示。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"714f9ac83975d59dafc3ba6fe8830d7059865181bdc55462f1b6c68fc48a8f12","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0530","title":"DoraCycle: Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles","year":"2025","venue":"","arxiv_id":"2503.03651","aliases":["library:P279"],"technical_key":"","catalog_id":"","library_id":"P279","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.03651","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":10.5,"wall_hours_max":10.5,"time_bound":"approximate","gpu_hours_min":84.0,"gpu_hours_max":84.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"DoraCycle 3K steps；原文每 1K steps 3.5h，相乘得 10.5h","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2503.03651","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"DoraCycle 通过图文双向循环利用单模态数据进行领域适配。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"fece5ec215ce31fa8a219ba4edf3d15a235f3a438c61e54e24cf462915960c6f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0531","title":"Automated Movie Generation via Multi-Agent CoT Planning","year":"2025","venue":"","arxiv_id":"2503.07314","aliases":["library:P281"],"technical_key":"","catalog_id":"","library_id":"P281","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.07314","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MovieAgent 用层级多智能体思维链组织场景、镜头与角色，并调用视频生成。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"8878c60c855d5257f8b841a2b4226e8d49e8becc551908052d2b9f318ba8d29f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0532","title":"Long-Context Autoregressive Video Modeling with Next-Frame Prediction","year":"2025","venue":"","arxiv_id":"2503.19325","aliases":["library:P286"],"technical_key":"","catalog_id":"","library_id":"P286","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2503.19325","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Long-Context Next-Frame Prediction 各任务表给 H100-days；UCF 12.7 / BAIR 2.6 / Minecraft 18.2 / DMLab 17.5；不是 GPU 数","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2503.19325","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"FAR 使用远近不同的 patch 粒度组织历史帧，提高长视频建模效率。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3cba45a65a2a4308333ae6c33175548f4fff3975cd89d99a05eac60ce837119c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0533","title":"LiveCC: Learning Video LLM with Streaming Speech Transcription at Scale","year":"2025","venue":"","arxiv_id":"2504.16030","aliases":["library:P290"],"technical_key":"","catalog_id":"","library_id":"P290","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2504.16030","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":128.0,"count_max":128.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"LiveCC pretraining 和 SFT 均 batch 512 / 128 GPUs；型号和阶段时长未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2504.16030","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"LiveCC 将词语和视频帧按时间交错训练，建立直播描述数据与体育评论评价。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"large","max_reported_ft_gpus":128.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"c442ec265b1716ad97b0a3122f776e28cb9959368609f55d486446e912e4a2c1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0534","title":"UniRL: Self-Improving Unified Multimodal Models via Supervised and Reinforcement Learning","year":"2025","venue":"","arxiv_id":"2505.23380","aliases":["library:P296"],"technical_key":"","catalog_id":"","library_id":"P296","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2505.23380","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"UniRL Show-o SFT 5K + GRPO 3K steps；Janus 为另一训练配置","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2505.23380","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"UniRL 以模型合成图像形成迭代数据，比较监督微调和 GRPO。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"79851d7d8c956da6f9e1a66a015ed1faf2ff8d22320b7f4452eb0229082e10b3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0535","title":"macOSWorld: A Multilingual Interactive Benchmark for GUI Agents","year":"2025","venue":"","arxiv_id":"2506.04135","aliases":["library:P297"],"technical_key":"","catalog_id":"","library_id":"P297","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2506.04135","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"macOSWorld 提供可交互软件任务，比较多类 GUI agent 的完成情况。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0d1c5bff225885f34d5cbe2a5f21aa2369755d4c0cc39c546a3dab900fc33828","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0536","title":"Show-o2: Improved Native Unified Multimodal Models","year":"2025","venue":"","arxiv_id":"2506.15564","aliases":["library:P300"],"technical_key":"","catalog_id":"","library_id":"P300","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2506.15564","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":36.0,"wall_hours_max":36.0,"time_bound":"approximate","gpu_hours_min":2304.0,"gpu_hours_max":2304.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Show-o2 初始训练段约 1.5 天；后续 40K iterations 和 stage 2 不包含在该耗时中","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2506.15564","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Show-o2 结合自回归与 flow matching，并在因果三维 VAE 空间融合视觉表示。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3f65225748b63248b58b2dd0b38da32cb1d27038648ba47a5fe3c67f4dad9984","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0537","title":"FramePrompt: In-context Controllable Animation with Zero Structural Changes","year":"2025","venue":"","arxiv_id":"2506.17301","aliases":["library:P301"],"technical_key":"","catalog_id":"","library_id":"P301","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2506.17301","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"FramePrompt 适配；结构不改不等于 training-free；耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2506.17301","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.14（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"FramePrompt 把参考外观、骨架运动和目标视频组织为统一视觉序列，减少专用结构修改。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"95cb3aff58446419f5e6bdebf3a4811254bce9077e3ed095f1f3d00564c406de","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0538","title":"Code2Video: A Code-centric Paradigm for Educational Video Generation","year":"2025","venue":"","arxiv_id":"2510.01174","aliases":["library:P305"],"technical_key":"","catalog_id":"","library_id":"P305","topic_code":"X0","topic":"教程、技能反馈与空间指导的应用近邻","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2510.01174","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Code2Video 组合规划、代码生成与视觉检查，并用学科视频和知识测验代理指标评价。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"445128aa7f126b6f1b9462eb0dc45fb3bcf6480d1fb59b1167148016b40d0f31","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0539","title":"Paper2Video: Automatic Video Generation from Scientific Papers","year":"2025","venue":"","arxiv_id":"2510.05096","aliases":["library:P306"],"technical_key":"","catalog_id":"","library_id":"P306","topic_code":"X0","topic":"教程、技能反馈与空间指导的应用近邻","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2510.05096","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Paper2Video 构建论文展示基准，PaperTalker 用多智能体协调幻灯片、语音、字幕和讲者。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0457f26a30bf406d17cb3e444b9a4e73c6bd852e89c179aeb45b9db5f7cd68a8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0540","title":"Computer-Use Agents as Judges for Generative User Interface","year":"2025","venue":"","arxiv_id":"2511.15567","aliases":["library:P314"],"technical_key":"","catalog_id":"","library_id":"P314","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2511.15567","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者提出 AUI-Gym，以 Coder 与计算机使用 agent 协作，通过可执行任务和可视化反馈改进界面。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b73ea64b24e04682d4c45bb8cdb52852c6bc0c5d57b3b5da987bd5eb36bd03e3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0541","title":"WorldWander: Bridging Egocentric and Exocentric Worlds in Video Generation","year":"2025","venue":"","arxiv_id":"2511.22098","aliases":["library:P317"],"technical_key":"","catalog_id":"","library_id":"P317","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2511.22098","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":192.0,"gpu_hours_max":192.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"WorldWander egocentric-exocentric translation；约 2 天；推理 3min 不计作训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2511.22098","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"WorldWander 以同步三元组数据训练上下文视角转换，并构建 EgoExo-8K。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"b5f565fe223ff1036c0360eeb33be6bf15591b8f5078563e11c797623a2b422a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0542","title":"X-Humanoid: Robotize Human Videos to Generate Humanoid Videos at Scale","year":"2025","venue":"","arxiv_id":"2512.04537","aliases":["library:P319"],"technical_key":"","catalog_id":"","library_id":"P319","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.04537","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":14.0,"wall_hours_max":14.0,"time_bound":"approximate","gpu_hours_min":56.0,"gpu_hours_max":56.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"X-Humanoid 1.3B；2,500 steps；14h","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2512.04537","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H200","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":16.0,"wall_hours_max":16.0,"time_bound":"approximate","gpu_hours_min":64.0,"gpu_hours_max":64.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"X-Humanoid 14B；16h；不同模型不相加","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2512.04537","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"X-Humanoid 使用合成配对数据适配 Wan 2.2，再将真实人类视频机器人化。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"91a09d75b52e5e96c84df98cef8b0aa136e28c26cefb2c6fbe43541280308bd8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0543","title":"H2R-Grounder: A Paired-Data-Free Paradigm for Translating Human Interaction Videos into Physically Grounded Robot Videos","year":"2025","venue":"","arxiv_id":"2512.09406","aliases":["library:P321"],"technical_key":"","catalog_id":"","library_id":"P321","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.09406","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"H2R-Grounder in-context model；200 steps；mini-batch 4，accumulation 2；VACE baseline 2 epochs 另计","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2512.09406","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"H2R-Grounder 用背景修补及夹爪位置、方向视觉提示连接人类与机器人视频，并微调 Wan 模型。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"c9a5c1bfce2693e66c6b35afba0f7cdf29cfabb825cb0007cf47b74611316fb3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0544","title":"Mitty: Diffusion-based Human-to-Robot Video Generation","year":"2025","venue":"","arxiv_id":"2512.17253","aliases":["library:P327"],"technical_key":"","catalog_id":"","library_id":"P327","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.17253","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":2.0,"count_max":2.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Mitty LoRA rank 96；5B 20K steps / 14B 10K；416×224，41 frames；不等于低总算力","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2512.17253","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Mitty 利用视频上下文学习与自动合成人机配对数据进行 Human2Robot 转换。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":2.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"deed0ad1879df16e1e333bba9528824d618aa03d1d0f4a4e0d050c3135ae30e3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0545","title":"ShowUI-$π$: Flow-based Generative Models as GUI Dexterous Hands","year":"2025","venue":"","arxiv_id":"2512.24965","aliases":["library:P331"],"technical_key":"","catalog_id":"","library_id":"P331","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.24965","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"ShowUI-π；SmolVLM 初始化；per-GPU batch 64；耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2512.24965","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.12（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"ShowUI-π 用 flow 模型生成连续鼠标调整，并构建拖拽数据与 ScreenDrag 基准。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"f4f5abeb94c3ce63d7d3327d8df15107777c219419c5b9c5301381dc8115496f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0546","title":"ShowUI-Aloha: Human-Taught GUI Agent","year":"2026","venue":"","arxiv_id":"2601.07181","aliases":["library:P335"],"technical_key":"","catalog_id":"","library_id":"P335","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2601.07181","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"ShowUI-Aloha 结合录屏和精确交互记录，整理可执行任务与示范数据。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2245d1d5eebec0f59d4f6f9e48cb7b7ff2507fc407e0dd5d65a3f657121c509e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0547","title":"World-VLA-Loop: Closed-Loop Learning of Video World Model and VLA Policy","year":"2026","venue":"","arxiv_id":"2602.06508","aliases":["library:P341"],"technical_key":"","catalog_id":"","library_id":"P341","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.06508","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":25.0,"wall_hours_max":35.0,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"World-VLA-Loop 每个 task 的 RL 约 25–35h；原文称 single H100 node，不能当作 single GPU","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2602.06508","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"World-VLA-Loop 以状态感知视频模型预测帧和二元奖励，形成世界模型与 VLA 的闭环训练。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"e6bf9537ea817621302227dedb92b5518075a7f1550e31bcb8dd270f585a8461","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0548","title":"Kiwi-Edit: Versatile Video Editing via Instruction and Reference Guidance","year":"2026","venue":"","arxiv_id":"2603.02175","aliases":["library:P348"],"technical_key":"","catalog_id":"","library_id":"P348","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2603.02175","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Kiwi-Edit 构建参考图引导的编辑四元组数据，并以多阶段训练融合语义查询和视觉特征。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9abdca7c41a3c1405833f78d0e171ccc0df6852b026a78ad905f61f95d98f13f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0549","title":"P-Flow: Prompting Visual Effects Generation","year":"2026","venue":"","arxiv_id":"2603.22091","aliases":["library:P355"],"technical_key":"","catalog_id":"","library_id":"P355","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2603.22091","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"P-Flow 探索免训练的动态特效定制，在文本和图像条件视频任务中评估效果。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"53eda04693f7a2cede09ab71c036fec9ad83bd7cd7ee26c6518f5006ca79c3ac","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0550","title":"OpenWorldLib: A Unified Codebase and Definition of Advanced World Models","year":"2026","venue":"","arxiv_id":"2604.04707","aliases":["library:P359"],"technical_key":"","catalog_id":"","library_id":"P359","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2604.04707","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"OpenWorldLib 提供多任务推理框架和能力分类，连接感知、交互与长期记忆模块。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f444bd94bda45752ef70a223331640100df3603f13162ea81a90ad6d4d2a9c0d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0551","title":"GameWorld: Towards Standardized and Verifiable Evaluation of Multimodal Game Agents","year":"2026","venue":"","arxiv_id":"2604.07429","aliases":["library:P361"],"technical_key":"","catalog_id":"","library_id":"P361","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2604.07429","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"GameWorld 统一浏览器游戏动作接口，并用可检验状态度量任务结果。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2fce3ea62f196b551e9df92320712c2705a319e290df05e311c772d9f31eb0cc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0552","title":"Sparkle: Realizing Lively Instruction-Guided Video Background Replacement via Decoupled Guidance","year":"2026","venue":"","arxiv_id":"2605.06535","aliases":["library:P365"],"technical_key":"","catalog_id":"","library_id":"P365","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.06535","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Sparkle 构建解耦引导和质量筛选的数据管线，并发布背景替换数据与基准。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a33a794ed54def4c8993e51a3263d394f0b6f8f2bb7b4fa09ae58f842e5bd1f8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0553","title":"OmniHumanoid: Streaming Cross-Embodiment Video Generation with Paired-Free Adaptation","year":"2026","venue":"","arxiv_id":"2605.12038","aliases":["library:P367"],"technical_key":"","catalog_id":"","library_id":"P367","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.12038","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":2.0,"count_max":2.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"OmniHumanoid Wan2.2-5B LoRA rank 64；10K steps；每 GPU batch 1","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2605.12038","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.7（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"OmniHumanoid 将共享运动迁移与外形适配器分开，并支持流式人形视频生成。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":2.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"6f561b1cf2e921a4849abe1a3207a630f1e3c4d5ad61e24175a0c47a2cc3f655","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0554","title":"World Action Models: The Next Frontier in Embodied AI","year":"2026","venue":"","arxiv_id":"2605.12090","aliases":["library:P368"],"technical_key":"","catalog_id":"","library_id":"P368","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.12090","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"World Action Models 综述区分级联和联合建模，梳理数据、条件和动作解码方式。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2c3b31e6f142e6b616f3d0c7f0e42fb43c3b688e72d961d39c5ad4f0caddddbd","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0555","title":"AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation","year":"2026","venue":"","arxiv_id":"2605.13724","aliases":["library:P369"],"technical_key":"","catalog_id":"","library_id":"P369","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.13724","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"AnyFlow training cost 按 8 H100 单步 profiling；总训练时长未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2605.13724","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"AnyFlow 学习任意时间区间的 flow map，并通过在策略分布上的模拟蒸馏改善少步采样。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"774446e76822e109aef7925f873a8afabc5ff58d9082b58a2a09e864c47a1178","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0556","title":"StreamingEffect: Real-Time Human-Centric Video Effect Generation","year":"2026","venue":"","arxiv_id":"2605.17019","aliases":["library:P370"],"technical_key":"","catalog_id":"","library_id":"P370","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.17019","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"StreamingEffect bidirectional teacher LoRA；8K steps；causal SFT / DMD 后续成本另缺","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2605.17019","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.7（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"StreamingEffect 将双向教师蒸馏为因果学生并减少采样步数，支持在线注入关键帧特效。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"ce14ccc020e7bbfddd9e6054f753ddc7c2119de4b32bf6ae5b21b192134f0b4b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0557","title":"VISTA: Triplet-Supervised Video Style Transfer with Diffusion Transformers","year":"2026","venue":"","arxiv_id":"2605.17312","aliases":["library:P371"],"technical_key":"","catalog_id":"","library_id":"P371","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.17312","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VISTA style transfer；20K steps；mixed precision / checkpointing","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2605.17312","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"VISTA 使用运动对齐的三元组数据与风格适配器训练视频扩散模型。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"e483fae601937bb01ab5ce68cbb3dab0f6e36ecbaef41ac7ed5c3873f1746b6a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0558","title":"SWEET: Sparse World Modeling with Image Editing for Embodied Task Execution","year":"2026","venue":"","arxiv_id":"2605.19319","aliases":["library:P373"],"technical_key":"","catalog_id":"","library_id":"P373","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.19319","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H20","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"SWEET FLUX.1-Kontext-dev DiT LoRA rank 32；6K steps；生成关键帧，不是 dense video world model","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2605.19319","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.6（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"SWEET 在相同机器人数据设置下比较视频生成和图像编辑，再用逐次编辑产生任务关键帧。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"c7039749fac2167d8b83dc59c9e69f8731a291081bbd266afab017b76b678ada","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0559","title":"PAI-Studio: Cinematic Video Background Replacement with Camera-Aware Motion","year":"2026","venue":"","arxiv_id":"2606.01399","aliases":["library:P377"],"technical_key":"","catalog_id":"","library_id":"P377","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2606.01399","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":135.0,"wall_hours_min":120.0,"wall_hours_max":120.0,"time_bound":"approximate","gpu_hours_min":960.0,"gpu_hours_max":960.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"PAI-Studio；1080p，97 frames；6K steps，约 5 天；训练 peak 135GB/GPU","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2606.01399","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"PAI-Studio 用参考条件与双向注意力联合建模前景动态和背景外观，并构建电影来源训练数据。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"41433876d52d950185ba5bf47f1e92d32b1dfb3b507db8c86969bd14c0b56529","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0560","title":"Dream.exe: Can Video Generation Models Dream Executable Robot Manipulation?","year":"2026","venue":"","arxiv_id":"2606.04811","aliases":["library:P380"],"technical_key":"","catalog_id":"","library_id":"P380","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2606.04811","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"主要贡献是评价框架；包含通用闭源、开源及机器人专用模型，并另测在测试集外数据微调的 Wan。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a99000842361fe3c0dfac331ada713dc4d69d75dc06e8a63db98cba8a44e2c28","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0561","title":"Supervise What Survives: Geometry-Guided VLA Adaptation from Synthetic Robot Videos","year":"2026","venue":"","arxiv_id":"2606.24448","aliases":["library:P388"],"technical_key":"","catalog_id":"","library_id":"P388","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2606.24448","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Supervise What Survives VLA LoRA rank 32；stage 1 5K / stage 2 10K steps；合成视频成本另计","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2606.24448","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"GRA 将合成数据中的几何运动提炼为二维末端路径，用辅助预测和表征对齐适配 VLA。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"14e6eabaeef43f2b6782073ec463302f41a2c2c2baf5baadf26b9b0ad1441b6d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0562","title":"Show-Harness: Just a VLM Agent Can Play Robots","year":"2026","venue":"","arxiv_id":"2609.10522","aliases":["library:P401"],"technical_key":"","catalog_id":"","library_id":"P401","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2609.10522","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":2.0,"wall_hours_max":2.0,"time_bound":"upper","gpu_hours_min":2.0,"gpu_hours_max":2.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Show-Harness Qwen3.5-2B 适配 <2h，为上界；7.9K examples，40 epochs","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2609.10522","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.25（同文多阶段请结合上方范围）"},{"role":"claim","device":"GPU（型号未定位）","count_min":1.0,"count_max":1.0,"card_memory_gb":24.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Show-Harness 作者称可在 24GB-class GPU 适配；实际报告的时长来自 H200，不作 24GB 实测","comparable":false,"is_gpu":true,"role_label":"作者声明的最低可运行配置","source_url":"https://arxiv.org/abs/2609.10522","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Show-Harness 通过语义动作接口连接意图与执行，并以 GUMI 支持 GUI 示范采集。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"74377b8b4ab2b565f6b3aa189a61c248c8bc412de26401708c415be5b23b944f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0563","title":"Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture","year":"2023","venue":"","arxiv_id":"2301.08243","aliases":["library:P167"],"technical_key":"","catalog_id":"","library_id":"P167","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2301.08243","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"upper","gpu_hours_min":1152.0,"gpu_hours_max":1152.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"I-JEPA ViT-Huge/14 ImageNet pretraining 少于 72h，为上界","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2301.08243","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.1（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"I-JEPA 用上下文预测多个目标区域的表示，研究掩码规模和上下文信息对语义学习的影响。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d8c8860ed0c04d0c38fc6193dec88e0379aa0d9ca9ff3e1621d312d5e9b08828","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0564","title":"Introduction to Latent Variable Energy-Based Models: A Path Towards Autonomous Machine Intelligence","year":"2023","venue":"","arxiv_id":"2306.02572","aliases":["library:P178"],"technical_key":"","catalog_id":"","library_id":"P178","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2306.02572","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"这篇介绍性论文梳理能量模型、潜变量和分层 JEPA 的关系，提供架构层面的研究动机。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"99aed2b82f8af1a525f40bb46a215673de62372efe8df943f956bba0e5139d84","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0565","title":"Self-Supervised Learning with Lie Symmetries for Partial Differential Equations","year":"2023","venue":"","arxiv_id":"2307.05432","aliases":["library:P180"],"technical_key":"","catalog_id":"","library_id":"P180","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2307.05432","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者利用 Lie 对称性构造联合嵌入自监督任务，评估方程参数回归与神经求解器时间步推进。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"25b543f5da2ed1f4087c0bc27b50faae54404071b7db293bb92824daff2ea35d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0566","title":"MC-JEPA: A Joint-Embedding Predictive Architecture for Self-Supervised Learning of Motion and Content Features","year":"2023","venue":"","arxiv_id":"2307.12698","aliases":["library:P183"],"technical_key":"","catalog_id":"","library_id":"P183","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2307.12698","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"V100","count_min":8.0,"count_max":8.0,"card_memory_gb":32.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"MC-JEPA self-supervised encoder / flow estimator；batch 384；时长未定位","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2307.12698","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.15（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MC-JEPA 将光流估计与内容表征目标联合训练，评估运动和分割任务。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5ad9d7fd2177882c62ee3a9907135d970ba177f7b99f15ae51423f9edde8caa7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0567","title":"GAIA: a benchmark for General AI Assistants","year":"2023","venue":"","arxiv_id":"2311.12983","aliases":["library:P194"],"technical_key":"","catalog_id":"","library_id":"P194","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2311.12983","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"GAIA 以需要检索、工具和综合推理的现实问题构建助手评测。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d44cddf9dfb23e63e8817581c9e422cf0f682170dc7a0803faf61d246cda0b68","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0568","title":"Gradient-based Planning with World Models","year":"2023","venue":"","arxiv_id":"2312.17227","aliases":["library:P208"],"technical_key":"","catalog_id":"","library_id":"P208","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2312.17227","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者比较梯度优化、其他模型预测控制和策略方法，考察样本效率与规划效果。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c26f70319b7faf818fb52de31906c73a565303a6124ce6f043571c5d6c436ff6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0569","title":"Learning and Leveraging World Models in Visual Representation Learning","year":"2024","venue":"","arxiv_id":"2403.00504","aliases":["library:P217"],"technical_key":"","catalog_id":"","library_id":"P217","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2403.00504","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Image World Models 把 JEPA 从遮挡补全扩展到给定全局光度变化的潜空间预测，并研究条件和容量的影响。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2e1ec86789a5483edbdc4b02929e31d0c6fee935a748a9219f4d092016d31801","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0570","title":"Revisiting Feature Prediction for Learning Visual Representations from Video","year":"2024","venue":"","arxiv_id":"2404.08471","aliases":["library:P223"],"technical_key":"","catalog_id":"","library_id":"P223","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2404.08471","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":null,"count_max":null,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"V-JEPA representation pretraining；GPU 数和总耗时未定位","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2404.08471","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"V-JEPA 从视频特征预测学习表征，以冻结骨干评估运动、外观和图像任务。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"50f56b87c4fbdff9717fdaa84c955410a401eb0f5f8e77f40e98f6954c31fc97","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0571","title":"EgoPet: Egomotion and Interaction Data from an Animal's Perspective","year":"2024","venue":"","arxiv_id":"2404.09991","aliases":["library:P224"],"technical_key":"","catalog_id":"","library_id":"P224","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2404.09991","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"EgoPet 收集宠物运动与交互视频，设计动物行为任务并评估其对四足机器人预训练的帮助。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"dd7f70c832b1b47f16ff6ed4b2482702dc8728e84183e9f2393fc628b97c7170","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0572","title":"Fine-Tuning Large Vision-Language Models as Decision-Making Agents via Reinforcement Learning","year":"2024","venue":"","arxiv_id":"2405.10292","aliases":["library:P228"],"technical_key":"","catalog_id":"","library_id":"P228","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2405.10292","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":40.0,"wall_hours_min":36.0,"wall_hours_max":36.0,"time_bound":"upper","gpu_hours_min":288.0,"gpu_hours_max":288.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"VLM as Decision-Making Agents；每条训练曲线最多 36h、VRAM <40GB；均为上界；full trainable components","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2405.10292","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者将任务描述、语言思维链和文本动作接入强化学习，通过任务奖励优化视觉语言决策。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"844b4805d4d3efdb4bd3bc40c3b01ca426b4efdc577b192fa7202ff9c9984dd3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0573","title":"Hierarchical World Models as Visual Whole-Body Humanoid Controllers","year":"2024","venue":"","arxiv_id":"2405.18418","aliases":["library:P231"],"technical_key":"","catalog_id":"","library_id":"P231","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2405.18418","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 3090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":288.0,"wall_hours_max":288.0,"time_bound":"approximate","gpu_hours_min":288.0,"gpu_hours_max":288.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Hierarchical Humanoid world model 的 tracking model；12 天","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2405.18418","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.6（同文多阶段请结合上方范围）"},{"role":"pre","device":"RTX 3090","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":96.0,"wall_hours_max":96.0,"time_bound":"approximate","gpu_hours_min":96.0,"gpu_hours_max":96.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Hierarchical Humanoid puppeteering model；4 天；单卡但训练总时长不小","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2405.18418","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.6（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者让高层视觉策略产生指令、低层控制器执行，并以奖励联合训练模拟人形机器人的多个任务。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5870ab189ff057743cc1d3a26322d28deb46fd7d0cd15129f99103c3893df230","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0574","title":"Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs","year":"2024","venue":"","arxiv_id":"2406.16860","aliases":["library:P237"],"technical_key":"","catalog_id":"","library_id":"P237","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2406.16860","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"TPU v4-512","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Cambrian-1 最终模型少于 4 天；保留 TPU pod 规格，不转换 NVIDIA 数量","comparable":true,"is_gpu":false,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2406.16860","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Cambrian-1 系统比较视觉编码器，引入 CV-Bench 与空间视觉聚合器，研究视觉中心的多模态学习。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"06830ad4ce459f3ac05a41963f2353deb2d2086de543a159e5c3c7850cf1601f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0575","title":"PooDLe: Pooled and dense self-supervised learning from naturalistic videos","year":"2024","venue":"","arxiv_id":"2408.11208","aliases":["library:P241"],"technical_key":"","catalog_id":"","library_id":"P241","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2408.11208","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":30.0,"wall_hours_max":30.0,"time_bound":"approximate","gpu_hours_min":480.0,"gpu_hours_max":480.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"PooDLe BDD100K 100 epochs；Walking Tours 每 epoch 40min，不能沿用 BDD 耗时","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2408.11208","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.16（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"PooDLe 结合全局不变表征与光流变换下的稠密等变目标，评估驾驶和行走视频中的特征。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c8f85d5571f0bc02685334ffee50f7526ec104c3403e1e1611d50bc5de65ff57","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0576","title":"DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning","year":"2024","venue":"","arxiv_id":"2411.04983","aliases":["library:P256"],"technical_key":"","catalog_id":"","library_id":"P256","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2411.04983","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"DINO-WM 在预训练视觉特征空间学习动力学，以目标条件搜索完成多类环境中的零样本规划。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0aeacd07ed902640510a6a650bee41229d1825293cea6e2f21a0a6b310213039","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0577","title":"Navigation World Models","year":"2024","venue":"","arxiv_id":"2412.03572","aliases":["library:P262","reasoning:CVPR-2025-2277"],"technical_key":"","catalog_id":"","library_id":"P262","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2412.03572","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Navigation World Models XL；8 machines × 8 H100；未定位时长","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2412.03572","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Navigation World Models 预测给定导航动作后的观察，再生成或排序抵达目标的候选轨迹。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1a3b985ae5d34838a5184f5c74c44f12ce758a287a893a6f39fccadd18830b7c","reasoning_id":"CVPR-2025-2277","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0578","title":"Video Representation Learning with Joint-Embedding Predictive Architectures","year":"2024","venue":"","arxiv_id":"2412.10925","aliases":["library:P264"],"technical_key":"","catalog_id":"","library_id":"P264","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2412.10925","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 8000","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"upper","gpu_hours_min":48.0,"gpu_hours_max":48.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Video Representation JEPA 小数据实验；单卡，各实验 <48h 为上界；MovingMNIST / CLEVRER","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2412.10925","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.8（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"VJ-VCR 以方差协方差正则学习视频表征，并探索潜变量对不确定未来的表达。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"097c58421ba6b9b1d1028e368a9669a570bb1e924913a1a81b1c956a50cae65b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0579","title":"MetaMorph: Multimodal Understanding and Generation via Instruction Tuning","year":"2024","venue":"","arxiv_id":"2412.14164","aliases":["library:P265"],"technical_key":"","catalog_id":"","library_id":"P265","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2412.14164","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"MetaMorph instruction tuning；1 epoch；GPU 数量/耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2412.14164","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.16（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"MetaMorph 在语言模型中加入视觉 token 预测，研究理解数据与生成能力的相互作用。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"a35e5dac3aa1cd061d6bf8de381ff50cee525b7862b76e734d675bf50ea05c33","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0580","title":"Intuitive physics understanding emerges from self-supervised pretraining on natural videos","year":"2025","venue":"","arxiv_id":"2502.11831","aliases":["library:P273"],"technical_key":"","catalog_id":"","library_id":"P273","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2502.11831","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":null,"count_max":null,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Intuitive Physics 的 V-JEPA pretraining 表；下游物理 probe 与预训练成本分开","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2502.11831","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.14（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者使用预期违背实验，测试潜空间视频预测器对物体持续性和形状一致性等性质的敏感性。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"60d335e1705e18006599623dbe4c5b4a7c19d64825704e9214b345b76f2f9faf","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0581","title":"Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models","year":"2025","venue":"","arxiv_id":"2502.14819","aliases":["library:P276"],"technical_key":"","catalog_id":"","library_id":"P276","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2502.14819","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"V100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":24.0,"wall_hours_max":24.0,"time_bound":"approximate","gpu_hours_min":24.0,"gpu_hours_max":24.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"PLDM 每个实验最多 1 天；完整研究含约 500–2,000 GPU-days，不把单次 run 当完整研究成本","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2502.14819","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者在导航任务中系统改变数据质量、多样性和环境布局，比较 JEPA 动力学规划与强化学习。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ea1ab7b1428fcd3187af8a6a62b30bb769d9f1ea2c376f7c31b72805c9b65b94","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0582","title":"OSVI-WM: One-Shot Visual Imitation for Unseen Tasks using World-Model-Guided Trajectory Generation","year":"2025","venue":"","arxiv_id":"2505.20425","aliases":["library:P294"],"technical_key":"","catalog_id":"","library_id":"P294","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2505.20425","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":4.0,"count_max":4.0,"card_memory_gb":40.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"OSVI-WM 所有模型训练用 4 A100；RTX A4000 / 2080 Ti 分别是仿真与真实部署，非训练","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2505.20425","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"OSVI-WM 根据专家示范和当前观察生成潜状态、动作轨迹，以世界模型指导视觉模仿。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"a8f4eb2a687d9e5f8f5d953ce4f3b7198ebc85eff80a63c5dc84437d084a9adf","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0583","title":"V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning","year":"2025","venue":"","arxiv_id":"2506.09985","aliases":["library:P299"],"technical_key":"","catalog_id":"","library_id":"P299","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2506.09985","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"V-JEPA 2 先学习视频潜表征，再以机器人数据训练动作条件预测器；语言对齐与机器人规划分别评价。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"38eb40b557173a20fd08193ef5aa140a206e05d84072e225daf0fc0870e5e000","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0584","title":"Whole-Body Conditioned Egocentric Video Prediction","year":"2025","venue":"","arxiv_id":"2506.21552","aliases":["library:P303"],"technical_key":"","catalog_id":"","library_id":"P303","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2506.21552","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"PEVA 用身体关节层级组织的三维姿态条件训练自回归扩散模型，并在 Nymeria 上设计分层预测与控制评价。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ecd377b1c5b65c7cbc6f1a8f04f6216cbc7284ed00125884c2b5d98aa981c0fa","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0585","title":"Back to the Features: DINO as a Foundation for Video World Models","year":"2025","venue":"","arxiv_id":"2507.19468","aliases":["library:P304"],"technical_key":"","catalog_id":"","library_id":"P304","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2507.19468","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"DINO-world 在 DINOv2 潜空间预测未来，并考察视频预测与观测—动作轨迹适配。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"bfef464ed58a73e4446acab849042209e9a1a390ba2abdfda6fc31a71ea00ff7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0586","title":"Cambrian-S: Towards Spatial Supersensing in Video","year":"2025","venue":"","arxiv_id":"2511.04670","aliases":["library:P310"],"technical_key":"","catalog_id":"","library_id":"P310","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2511.04670","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Cambrian-S 提出空间 supersensing 框架、VSI-SUPER 基准与 VSI-590K 数据，考察持续空间回忆和计数。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"731179a63ff9c902bbfb45d39b217c0430163afe4a4b138d67ee505ad22a6bfb","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0587","title":"LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics","year":"2025","venue":"","arxiv_id":"2511.08544","aliases":["library:P311"],"technical_key":"","catalog_id":"","library_id":"P311","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2511.08544","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"LeJEPA 提出 Sketched Isotropic Gaussian Regularization，研究表示分布与下游预测风险的联系。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d08cc82c9859ee2dc9ee2137dff80e6ad7ef07c353f4c6e3bd1f9b47583e20d9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0588","title":"From Generated Human Videos to Physically Plausible Robot Trajectories","year":"2025","venue":"","arxiv_id":"2512.05094","aliases":["library:P320"],"technical_key":"","catalog_id":"","library_id":"P320","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.05094","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 4090","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"GenMimic robot tracking policy；IsaacGym，1.5B samples；4060 Mobile 为部署","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2512.05094","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"GenMimic 将视频生成、4D 人体重建、动作重定向和物理强化学习控制组合起来，并构建动作基准。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"37888430bebaec47e1e7577e2db3713e8df2116b5e53fdc95d3ffe194443f95f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0589","title":"Closing the Train-Test Gap in World Models for Gradient-Based Planning","year":"2025","venue":"","arxiv_id":"2512.09929","aliases":["library:P322"],"technical_key":"","catalog_id":"","library_id":"P322","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.09929","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"B200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Adversarial World Modeling PushT；1 epoch；20K rollouts","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2512.09929","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.18（同文多阶段请结合上方范围）"},{"role":"ft","device":"B200","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Adversarial World Modeling PointMaze / Wall；1 / 2 epochs；不同任务不相加","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2512.09929","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.18（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者以面向规划的数据合成和训练设计，改善 JEPA 在梯度动作规划中的使用效果。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"edd8ca28a36cdaa80ab1235af0f75902f613c4b9e012774d02e5dfd503c8be6b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0590","title":"VL-JEPA: Joint Embedding Predictive Architecture for Vision-language","year":"2025","venue":"","arxiv_id":"2512.10942","aliases":["library:P323"],"technical_key":"","catalog_id":"","library_id":"P323","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.10942","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H200","count_min":192.0,"count_max":192.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":672.0,"wall_hours_max":672.0,"time_bound":"approximate","gpu_hours_min":129024.0,"gpu_hours_max":129024.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"VL-JEPA BASE video pretraining；24 nodes × 8；4 周；下游 SFT 另计","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2512.10942","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.4（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"VL-JEPA 将视觉语言预测放在语义嵌入空间，研究选择性解码与视频理解。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9c96edffa425dfb0105ce9673253f3bbc439320fbad261653292d31c08358b37","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0591","title":"World Models for Learning Dexterous Hand-Object Interactions from Human Videos","year":"2025","venue":"","arxiv_id":"2512.13644","aliases":["library:P326"],"technical_key":"","catalog_id":"","library_id":"P326","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.13644","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"DexWM 使用手指关键点动作表示与潜状态、手部一致性目标，学习可用于灵巧规划的动态表示。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9eda95ec48d76218cce3e6b009ba05c13deae281a338f37a73da56a01b57470f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0592","title":"What Drives Success in Physical Planning with Joint-Embedding Predictive World Models?","year":"2025","venue":"","arxiv_id":"2512.24497","aliases":["library:P329"],"technical_key":"","catalog_id":"","library_id":"P329","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.24497","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H100","count_min":16.0,"count_max":16.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Physical Planning JEPA 的 Maze / Wall training；PushT / Metaworld 另用 32 H100；每 epoch 时间需按模型区分","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2512.24497","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.27（同文多阶段请结合上方范围）"},{"role":"pre","device":"H100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Physical Planning JEPA 的 PushT / Metaworld training；不是上行同一任务","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2512.24497","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.27（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者系统比较表示、训练目标和规划器配置，并在仿真和真实任务中检查效果。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"de9712a46bcad938526485f3d91d3134e17a4130ba5a70824c27f2d133853068","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0593","title":"Value-guided action planning with JEPA world models","year":"2025","venue":"","arxiv_id":"2601.00844","aliases":["library:P332"],"technical_key":"","catalog_id":"","library_id":"P332","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2601.00844","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Value-guided action planning 使表示距离近似目标条件价值，从而为动作优化提供更有用的信号。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2c3d1f5e7eec696b50a1d0536eba1c1b8fb8993e355083348d4d2e1879d9fbc1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0594","title":"Learning Latent Action World Models In The Wild","year":"2026","venue":"","arxiv_id":"2601.05230","aliases":["library:P334","reasoning:ICML-2026-1668"],"technical_key":"","catalog_id":"","library_id":"P334","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2601.05230","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":12.0,"wall_hours_max":12.0,"time_bound":"approximate","gpu_hours_min":768.0,"gpu_hours_max":768.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Latent Action World Models latent dynamics / inverse dynamics；约 12h，encoder 预训练不包含","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2601.05230","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.17（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者从自然视频学习受约束的连续潜动作并用于动作条件世界建模，减少对离散动作码的依赖。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a22e2f51466baa758d63a407eb61cefc565cf5b94f4e6fdbcbe6f92ad38d7b4f","reasoning_id":"ICML-2026-1668","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0595","title":"Parallel Stochastic Gradient-Based Planning for World Models","year":"2026","venue":"","arxiv_id":"2602.00475","aliases":["library:P337"],"technical_key":"","catalog_id":"","library_id":"P337","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.00475","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"GRASP 引入并行优化的虚拟状态和软动力学约束，缓解逐步展开的规划成本。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b6a08bd1e3eb3231c1656802848a175c486260b494f0efbb084c17791203bb0a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0596","title":"Rectified LpJEPA: Joint-Embedding Predictive Architectures with Sparse and Maximum-Entropy Representations","year":"2026","venue":"","arxiv_id":"2602.01456","aliases":["library:P338"],"technical_key":"","catalog_id":"","library_id":"P338","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.01456","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"L40S","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":55.0,"wall_hours_max":55.0,"time_bound":"approximate","gpu_hours_min":55.0,"gpu_hours_max":55.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Rectified LpJEPA ImageNet-100；1,000 epochs，2天7h；单 seed","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2602.01456","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.41（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Rectified LpJEPA 以整流广义高斯分布约束表示，研究稀疏性与视觉表征学习。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f18a6f6a2a23101be8aaef327e29deed452dd167ba7194854148316a95af47ee","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0597","title":"A Lightweight Library for Energy-Based Joint-Embedding Predictive Architectures","year":"2026","venue":"","arxiv_id":"2602.03604","aliases":["library:P339"],"technical_key":"","catalog_id":"","library_id":"P339","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.03604","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"GPU（型号未定位）","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"轻量 EB-JEPA library 示例；原文称数小时，无精确耗时；不是 foundation model 训练","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2602.03604","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"EB-JEPA 整理表征学习与世界建模的模块化实现和实验接口。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"6b397e00f6e731a66f8f53d33f3b0279f513576d1e79f85be7453d9b4c9324c1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0598","title":"stable-worldmodel-v1: Reproducible World Modeling Research and Evaluation","year":"2026","venue":"","arxiv_id":"2602.08968","aliases":["library:P343"],"technical_key":"","catalog_id":"","library_id":"P343","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2602.08968","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"stable-worldmodel-v1 提供统一模型与规划评测，讨论 DINO-WM 等方法对分布变化的表现。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4d2665b7f0e5b302dfdbc70cc9a9f36140f8d8b3fcd362234043c84e83cf17e9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0599","title":"Temporal Straightening for Latent Planning","year":"2026","venue":"","arxiv_id":"2603.12231","aliases":["library:P350"],"technical_key":"","catalog_id":"","library_id":"P350","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2603.12231","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Temporal Straightening 用轨迹曲率约束改善潜空间距离和规划优化条件。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"20895a8796104399cd504c36ba34ca4766906e64beeb02fc2bac24d7040272c2","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0600","title":"Representation Learning for Spatiotemporal Physical Systems","year":"2026","venue":"","arxiv_id":"2603.13227","aliases":["library:P351"],"technical_key":"","catalog_id":"","library_id":"P351","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2603.13227","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者用物理参数估计等下游任务评估时空物理系统的表示，比较预测式与通用自监督学习。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c5a27ad8bdbc472bcfe76869b1ca8e9c6aee182303ff6904d250a8aaff0047f2","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0601","title":"V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning","year":"2026","venue":"","arxiv_id":"2603.14482","aliases":["library:P352"],"technical_key":"","catalog_id":"","library_id":"P352","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2603.14482","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"V-JEPA 2.1 通过图像与视频上的自监督学习及中间层监督改善稠密视觉表示。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"7c54a4b0ba04e3403161bf468c7712563a758a4e3ca9d76db021cef22963c66f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0602","title":"LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels","year":"2026","venue":"","arxiv_id":"2603.19312","aliases":["library:P353"],"technical_key":"","catalog_id":"","library_id":"P353","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2603.19312","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"GPU（型号未定位）","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"LeWorldModel 15M end-to-end；原文称数小时；设备型号和精确时长需另核","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2603.19312","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"LeWorldModel 联合学习编码与未来潜状态预测，并以高斯正则约束表示分布，考察规划和物理异常信号。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9c8e59c19e563b36912d520a833210e4de641588fcef6785c2c4961b122312d8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0603","title":"Hierarchical Planning with Latent World Models","year":"2026","venue":"","arxiv_id":"2604.03208","aliases":["library:P358"],"technical_key":"","catalog_id":"","library_id":"P358","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2604.03208","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"HWM 使用分层世界模型与模型预测控制，在不同层级组织目标和动作搜索。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"6bb8cffc77e0ea5ee73cc9ac02c696e29f92623dc6d1ea1f6164ed339196ef39","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0604","title":"On Training in Imagination","year":"2026","venue":"","arxiv_id":"2605.06732","aliases":["library:P366"],"technical_key":"","catalog_id":"","library_id":"P366","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.06732","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者分析动力学和奖励误差对策略学习的影响，讨论噪声、方差与计算预算之间的关系。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4f97ea9a7dde688af958de2dfc4d616c792fe2b638fcbc4e67045ef504684bcb","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0605","title":"stable-worldmodel: A Platform for Reproducible World Modeling Research and Evaluation","year":"2026","venue":"","arxiv_id":"2605.21800","aliases":["library:P374"],"technical_key":"","catalog_id":"","library_id":"P374","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.21800","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"stable-worldmodel 提供数据、模型和规划器接口，以及视觉、几何和物理变化的统一实验平台。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"320be8dd79c84d5c54319f6bd4c70d2548198cc6cf3db6825f4fd3430dee0ed9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0606","title":"When Does LeJEPA Learn a World Model?","year":"2026","venue":"","arxiv_id":"2605.26379","aliases":["library:P375"],"technical_key":"","catalog_id":"","library_id":"P375","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2605.26379","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"作者分析 LeJEPA 的线性可辨识性以及状态分布、噪声假设与规划性能之间的关系。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d00becef5a63bae985d65f9e7cba76a5539992f5e539b9fa81691594d4bbd7b3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0607","title":"Unifying Object-Centric World Models and Diffusion Policy: A Hierarchical Framework for Multi-Stage Robotic Tasks","year":"2026","venue":"","arxiv_id":"2606.08775","aliases":["library:P381"],"technical_key":"","catalog_id":"","library_id":"P381","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2606.08775","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"WorldDP 单个 GPU cluster node；每模型 5 epochs；不能把 node 当成 1 GPU","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2606.08775","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.9（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"WorldDP 在运行时提出可行的子目标，由低层 diffusion policy 执行，连接世界预测与多阶段机器人任务。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5a3faae71fe1d1097bbd0b5f0efe64aa0293d519973c1217139a50cbd64698c8","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0608","title":"You Don't Need Strong Assumptions: Visual Representation Learning via Temporal Differences","year":"2026","venue":"","arxiv_id":"2606.15956","aliases":["library:P382"],"technical_key":"","catalog_id":"","library_id":"P382","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2606.15956","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H100","count_min":2.0,"count_max":2.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":96.0,"gpu_hours_max":96.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Temporal Differences Visual Learning 在 SSv2 预训练；20 epochs，48h","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2606.15956","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":1.0,"count_max":1.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"UperNet segmentation adaptation on ADE20K / Cityscapes；耗时另未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2606.15956","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"这项工作研究利用时间差分学习视觉表示，重新审视常见预训练假设。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"78ec102bd5d81f9b3fc8561c5e7cf8a4da8c5e929710c1bd311fcdf4a295c0b2","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0609","title":"SkyJEPA: Learning Long-Horizon World Models for Zero-Shot Sim-to-Real Control of Quadrotors","year":"2026","venue":"","arxiv_id":"2606.23444","aliases":["library:P386"],"technical_key":"","catalog_id":"","library_id":"P386","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2606.23444","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"SkyJEPA 将隐空间动力学、物理启发的状态探针与采样规划结合，研究仿真到真实的飞行控制。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"bee08cd59925d84840c05a848164bfb3012868510b717fedba661a6a3be4be45","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0610","title":"Patch Policy: Efficient Embodied Control via Dense Visual Representations","year":"2026","venue":"","arxiv_id":"2607.18236","aliases":["library:P395"],"technical_key":"","catalog_id":"","library_id":"P395","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2607.18236","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Patch Policy 使用预训练视觉特征中的局部 token，减少对完整视觉语言模型的依赖。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a54a8924af0c6618d941e6be99098bc2bb1978b775a547043e9d69eea91d5429","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0611","title":"LpWM: A Case for Sparse Representations in World Models","year":"2026","venue":"","arxiv_id":"2608.22764","aliases":["library:P397"],"technical_key":"","catalog_id":"","library_id":"P397","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2608.22764","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"LpWM 从稀疏表征与动力学近似的关系出发研究世界模型的潜空间结构。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c7e1c0d239b96427d3f39b8e152cb0c951673bc81807bfe8f5fda80f15e47a29","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0612","title":"LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics","year":"2026","venue":"","arxiv_id":"2608.27395","aliases":["library:P399"],"technical_key":"","catalog_id":"","library_id":"P399","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2608.27395","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 5080","count_min":1.0,"count_max":1.0,"card_memory_gb":16.0,"peak_memory_gb":null,"wall_hours_min":12.0,"wall_hours_max":12.0,"time_bound":"approximate","gpu_hours_min":12.0,"gpu_hours_max":12.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"LeVJEPA ViT-Tiny consumer-hardware 实验；8 Walking Tours videos / 620K frames；不是完整大模型配置","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/2608.27395","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"LeVJEPA 将 LeJEPA 的表征学习目标扩展到视频，并比较外观与运动任务中的表示质量。 本条方法定位来自摘要，训练细节待核。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ecbc9da4afe9186f4af4de009029a22d601f08cf6637fd977037fc43cdd46afe","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0613","title":"Reasoning with Language Model is Planning with World Model","year":"2023","venue":"","arxiv_id":"","aliases":["library:P104"],"technical_key":"","catalog_id":"","library_id":"P104","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://github.com/maitrix-org/llm-reasoners","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"定义状态、动作和奖励，MCTS选择/扩展/回传，LLM模拟后果并评估中间收益，输出一个选定计划。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4dfa9f8ab2fcbd4bd4d1a796193a23996c350372fc6e685212d198a590a96c6c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0614","title":"New Evaluation, Library, and Analysis of Step-by-Step Reasoning with Large Language Models","year":"2024","venue":"","arxiv_id":"2404.05221","aliases":["library:P105"],"technical_key":"","catalog_id":"","library_id":"P105","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2404.05221v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"自动归纳任务错误/规则形成rubric，再评估链；将CoT、ToT、RAP统一为搜索/奖励/转移，研究宽度、深度和提示。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"163029c88573270d9c71f729b36aa6e1f6f4210e247fb9d5b95773a98c987900","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0615","title":"Pandora: Towards General World Model with Natural Language Actions and Video States","year":"待核实","venue":"","arxiv_id":"","aliases":["library:P106"],"technical_key":"","catalog_id":"","library_id":"P106","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"未归入视频任务轴","directions":[],"source_url":"https://world-model.ai","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Vicuna7B接收历史视觉状态与动作，输出条件embedding供DynamiCrafter生成下一段，额外使用上一段末4帧维持连续性。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"dcfbd2bff898d2011befa792be7895731996ad04ad6b420b21e22784c4779bda","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0616","title":"Offline Reinforcement Learning for LLM Multi-step Reasoning","year":"2025","venue":"","arxiv_id":"","aliases":["library:P107"],"technical_key":"","catalog_id":"","library_id":"P107","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://aclanthology.org/2025.findings-acl.464/","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"最大熵RL下优化策略/价值的Bellman一致性，支持token或响应层次目标、迭代数据更新和价值引导搜索。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"541f86aa61d9f0537892c8f2bb521c4c9969047bfbf7a131f6a56dd2bc009cbd","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0617","title":"Flow of Reasoning: Training LLMs for Divergent Problem Solving with Minimal Examples","year":"未注明（匿名稿）","venue":"","arxiv_id":"2406.05673","aliases":["library:P108"],"technical_key":"","catalog_id":"","library_id":"P108","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2406.05673","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Flow of Reasoning BlocksWorld 6-step、10 epochs；时长表需另核；单卡实测","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2406.05673","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.7（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"以Markov推理状态与路径奖励构造流匹配/轨迹训练，从有限问题反复探索高质量路径并微调LLM。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"3f8b9fabbf3b1bc7dcbebbebbc098f728e885b8b4071dd4b65d8da044cb8ff20","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0618","title":"Understanding the Sources of Uncertainty for Large Language and Multimodal Models","year":"2025","venue":"","arxiv_id":"","aliases":["library:P110"],"technical_key":"","catalog_id":"","library_id":"P110","topic_code":"M5","topic":"不确定性、决策损失与鲁棒优化","task":"未归入视频任务轴","directions":[],"source_url":"https://scholar.google.com/scholar?q=Understanding%20the%20Sources%20of%20Uncertainty%20for%20Large%20Language%20and%20Multimodal%20Models","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"用条件熵与互信息逐层分解总熵，对提示/输入变体采样，比较分量与错误关联；用于VQA及SVAMP自训练。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0858c37fdba10cb2d633fbc7db5dd3faa22f7ef5b14b431c092f4be66e0c65e4","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0619","title":"CocoaBench: Evaluating Unified Digital Agents in the Wild","year":"2026","venue":"","arxiv_id":"2604.11201","aliases":["library:P111"],"technical_key":"","catalog_id":"","library_id":"P111","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2604.11201v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"只给任务说明，以最终产物函数判定成功；分别测完整系统和相同Cocoa-Agent下的模型，分析规划/执行/视觉错误。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2e310d1649694128608e9519d1abafc1366b514c504e520cd02f71cb0f90d9bf","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0620","title":"Emergence of Superposition: Unveiling the Training Dynamics of Chain of Continuous Thought","year":"2026","venue":"","arxiv_id":"2509.23365","aliases":["library:P113"],"technical_key":"","catalog_id":"","library_id":"P113","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2509.23365v3","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"推导梯度训练下局部搜索logit变化，用课程训练跟踪注意力及潜状态，检验理论预测。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c1137f5f79b65c86f096fc4c0f5a8566162065fa0ee2a2410e918599e595fd6d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0621","title":"LLM Pretraining with Continuous Concepts","year":"2026 / 2025","venue":"","arxiv_id":"2502.08524","aliases":["library:P114"],"technical_key":"","catalog_id":"","library_id":"P114","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2502.08524v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"从教师隐状态经SAE提取稀疏概念，学生同时预测token和概念，将压缩概念向量插入隐状态流。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1e59f9cb869e9710fd6a98f1f11ef5effc9ec87f15229d054a15b05ea6b848e1","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0622","title":"ToolkenGPT: Augmenting Frozen Language Models with Massive Tools via Tool Embeddings","year":"2023","venue":"","arxiv_id":"","aliases":["library:P115"],"technical_key":"","catalog_id":"","library_id":"P115","topic_code":"M1","topic":"Agent 规划、工具、澄清与经验","task":"未归入视频任务轴","directions":[],"source_url":"https://github.com/Ber666/ToolkenGPT","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 3090","count_min":1.0,"count_max":1.0,"card_memory_gb":24.0,"peak_memory_gb":null,"wall_hours_min":0.033333,"wall_hours_max":0.033333,"time_bound":"approximate","gpu_hours_min":0.0333,"gpu_hours_max":0.0333,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"ToolkenGPT LLaMA-7B 工具 embedding 优化；FuncQA；2min；冻结 LLM，非 full fine-tuning","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://github.com/Ber666/ToolkenGPT","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"把工具embedding矩阵拼到输出词表，联合选择文字/工具；触发工具后切参数提示，执行结果回填上下文。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"9b0ce59a5a33355758ac13e869b5ed7b3e123c40642062a1e0fad1a3ccc4171b","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0623","title":"Reasoning by Superposition: A Theoretical Perspective on Chain of Continuous Thought","year":"2025","venue":"","arxiv_id":"","aliases":["library:P116"],"technical_key":"","catalog_id":"","library_id":"P116","topic_code":"B2","topic":"理论、机制分析与负面证据","task":"未归入视频任务轴","directions":[],"source_url":"https://github.com/Ber666/reasoning-by-superposition","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"在连续向量中叠加搜索前沿，用两层注意力迭代扩展；ProsQA合成子集上课程训练与潜态/注意力分析。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c45ea2e248c31ec92a42f84cc0c8f010cb5b066877624e0dbd199e97c1ece5c9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0624","title":"Calibrating Reasoning in Language Models with Internal Consistency","year":"2024","venue":"","arxiv_id":"2405.18711","aliases":["library:P118"],"technical_key":"","catalog_id":"","library_id":"P118","topic_code":"M5","topic":"不确定性、决策损失与鲁棒优化","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2405.18711v2","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"用中间层潜预测与末层比较，计算一致性；对多条CoT/L2M路径加权，提高更可信答案的权重。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5dad716e7db35c7ecbc3d6493929d2030ec3598be376059793c795a56f55bb30","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0625","title":"Reasoning Models Know When They’re Right: Probing Hidden States for Self-Verification","year":"2025","venue":"","arxiv_id":"2504.05419","aliases":["library:P119"],"technical_key":"","catalog_id":"","library_id":"P119","topic_code":"M5","topic":"不确定性、决策损失与鲁棒优化","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2504.05419v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"收集中间答案隐状态与标签，训练MLP探针，验证集调参/校准，推理中达到阈值时退出并输出答案。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f8103357fd642ee44cccc4a5a5dddf66cd5c0f2f9bbe83afa16aa7251da45df6","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0626","title":"Imagine This! Scripts to Compositions to Videos","year":"2018","venue":"","arxiv_id":"1804.03608","aliases":["library:P120"],"technical_key":"","catalog_id":"","library_id":"P120","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/1804.03608v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"逐实体布局预测、联合文本/场景/位置检索，triplet与辅助分类学习检索空间，融合时空片段。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c335f5310e5a22a670ecf4cd1094c2d9b459e9c3b3e16a6c67b881ad59f07aad","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0627","title":"World Models","year":"2018","venue":"","arxiv_id":"1803.10122","aliases":["library:P121"],"technical_key":"","catalog_id":"","library_id":"P121","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/1803.10122v4","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"GPU（型号未定位）","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":1.0,"wall_hours_max":1.0,"time_bound":"upper","gpu_hours_min":1.0,"gpu_hours_max":1.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"原始 World Models 的 VAE / MDN-RNN 各 <1h 为上界；不含 controller evolution 全部成本","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/abs/1803.10122v4","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"无监督学习V和M，控制器用z及RNN隐状态选动作，由进化策略优化；在随机梦境模型中训练并调温防利用模型缺陷。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"324ec6324af26664162fe37db0e7803b328f67aa0171f54fb2ab184eb2355325","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0628","title":"Phenaki: Variable Length Video Generation from Open Domain Textual Descriptions","year":"2023","venue":"","arxiv_id":"","aliases":["library:P122"],"technical_key":"","catalog_id":"","library_id":"P122","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://phenaki.github.io/","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"因果时间注意力压缩视频token，T5文本条件下迭代补全被掩token，解码后以已有帧/新提示生成后续片段。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"748dc25dacfff0c7f2f6f611adbd71f243671526c2cf8c44916b594990689b2f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0629","title":"iVideoGPT: Interactive VideoGPTs are Scalable World Models","year":"2024","venue":"","arxiv_id":"2405.15223","aliases":["library:P124","reasoning:NeurIPS-2024-2391"],"technical_key":"","catalog_id":"","library_id":"P124","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2405.15223v1","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100 / RTX 3090","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"iVideoGPT 每实验 4 GPUs；64² 实验每卡 24GB 可运行；不同任务 GPU-days 见表 2，不能统一成一项","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2405.15223v1","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.25（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"用上下文帧帮助大幅压缩后续帧token，将动作奖励交错建模；下游微调模型/编码器，并以模型生成经验供策略学习。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"8d9c4484f1e03583f901bb8a0ff51c7d3720e693fa8fa19686cc026158c7c09f","reasoning_id":"NeurIPS-2024-2391","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0630","title":"Diffusion-Based Visual Art Creation: A Survey and New Perspectives","year":"2025","venue":"","arxiv_id":"","aliases":["library:P126"],"technical_key":"","catalog_id":"","library_id":"P126","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://doi.org/10.1145/3728459","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"明确研究问题和纳入条件，结构/时间分析论文编码，归纳技术与人机协作趋势。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"49f3a91f236e28f8772078e2d8d7026cf78d317b044561686afb9c0c54cd4755","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0631","title":"MovieDreamer: Hierarchical Generation for Coherent Long Visual Sequences","year":"2025","venue":"","arxiv_id":"2407.16655","aliases":["library:P127"],"technical_key":"","catalog_id":"","library_id":"P127","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2407.16655v3","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":288.0,"gpu_hours_max":288.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"MovieDreamer global autoregressive model；128 context frames；3 天；local diffusion 其他阶段另计","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2407.16655v3","locator":"正文／附录的 Training / Implementation / Compute 设置；关键词定位 PDF p.17（同文多阶段请结合上方范围）"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"扩散自动编码器产生紧凑关键帧token，自回归预测下一个keyframe，文本/FaceID条件解码，再生成帧间视频。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"40035786e69bacd5241fffb9a68fa9e8548cb2b2406181615f5fce8f7fdb18b3","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0632","title":"StoryMem: Multi-shot Long Video Storytelling with Memory","year":"2025","venue":"","arxiv_id":"2512.19539","aliases":["library:P128"],"technical_key":"","catalog_id":"","library_id":"P128","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2512.19539v1","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"历史镜头选相关美观关键帧写入记忆，潜空间拼接注入模型；MI2V/MR2V支持过渡和定制。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"7c22f22a1db2c2ae6a3c80ed253d69b0812f6884ecd5b392e3af055cda8b4b1a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0633","title":"StoryReasoning Dataset: Using Chain-of-Thought for Scene Understanding and Grounded Story Generation","year":"2025","venue":"","arxiv_id":"2505.10292","aliases":["library:P129"],"technical_key":"","catalog_id":"","library_id":"P129","topic_code":"B4","topic":"数据、仿真与可复现实验基础","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2505.10292v2","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":2.0,"count_max":2.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":6.0,"wall_hours_max":12.0,"time_bound":"approximate","gpu_hours_min":12.0,"gpu_hours_max":24.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"StoryReasoning 的 Qwen Storyteller；4 epochs；LoRA / full 配置 6–12h","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/abs/2505.10292v2","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"视觉相似性与人脸识别匹配对象，表格组织跨帧场景，CoT生成并将词语链接到实体，微调Qwen2.5VL7B。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":2.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"1cd52140256a0e29fcc7ca833d91b20be727e923e10c85eae8a65965b9f60157","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0634","title":"Controllable Video Generation: A Survey","year":"2025 / 2026","venue":"","arxiv_id":"2507.16869","aliases":["library:P131"],"technical_key":"","catalog_id":"","library_id":"P131","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2507.16869v3","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"PRISMA式四阶段筛选，按基础模型、条件类型、去噪注入方式和应用分析研究。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0f3c830be6212bb9fc143cd3a1e62a66bc01e89bc23d799dd44345800f3fcd1a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0635","title":"Pre-trained Video Generative Models as World Simulators","year":"2025","venue":"","arxiv_id":"2502.07825","aliases":["library:P132"],"technical_key":"","catalog_id":"","library_id":"P132","topic_code":"M3","topic":"世界模型、闭环规划与控制","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/abs/2502.07825","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"轻量动作模块对齐动作与视觉变化，以运动损失强化动态；按价值选想象轨迹供PPO/离线策略训练。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c897ffaed334212f21a13368c960902b550293cab8cbfd741bf8f58fa41b8689","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0636","title":"StoryBench: A Multifaceted Benchmark for Continuous Story Visualization","year":"2023","venue":"","arxiv_id":"","aliases":["library:P136"],"technical_key":"","catalog_id":"","library_id":"P136","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://github.com/google/storybench","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"TPU v4 chips","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":24.0,"wall_hours_max":24.0,"time_bound":"upper","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"StoryBench 每模型 500K steps 少于一天，为上界；4×4×4 TPU chips；不等于 GPU","comparable":true,"is_gpu":false,"role_label":"微调／适配训练","source_url":"https://github.com/google/storybench","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"人工标注时间段/动作及诊断标签；用小型Phenaki基线进行零样本、单任务、多任务微调；比较自动指标与人类评价。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b259c3c21acaee948a8bb592ddd8130359677ae89bcf9464ef199e6673e264ff","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0637","title":"Long-Context State-Space Video World Models","year":"2025","venue":"","arxiv_id":"","aliases":["library:P137","reasoning:ICCV-2025-1556"],"technical_key":"","catalog_id":"","library_id":"P137","topic_code":"G4","topic":"动作／物理条件的世界生成","task":"未归入视频任务轴","directions":[],"source_url":"https://www.openaccess.thecvf.com/content/ICCV2025/papers/Po_Long-Context_State-Space_Video_World_Models_ICCV_2025_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"Mamba状态压缩长历史，block-wise扫描与frame-local attention结合，逐帧扩散生成。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5067632b583a5b0ce770b123399adaca5ccfe4893e5e146b0c8a459017d9126d","reasoning_id":"ICCV-2025-1556","reasoning_role":"R3","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0638","title":"StoryGAN: A Sequential Conditional GAN for Story Visualization","year":"2019","venue":"","arxiv_id":"","aliases":["library:P138"],"technical_key":"","catalog_id":"","library_id":"P138","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://github.com/yitong91/StoryGAN","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"故事编码初始化状态，Text2Gist融合当前句与历史，逐图生成；两级对抗损失约束单图和全序列。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"64914287dac328f2a8009e23d03d8d3a680bade3ef33b687f292bcd6f8d4f6ff","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0639","title":"Is Your World Simulator a Good Story Presenter? A Consecutive Events-Based Benchmark for Future Long Video Generation","year":"2025","venue":"","arxiv_id":"","aliases":["library:P139"],"technical_key":"","catalog_id":"","library_id":"P139","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://ypwang61.github.io/project/StoryEval","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"检索真实视频/人工构思提示，VLM如GPT4o和LLaVAOV72B核验每一事件并投票，与人工评价检查一致性。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0c91610bde9af8bb4521519e2fde2d2356276adce9ea1dfd6b98cefb3b6ab658","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0640","title":"Look-Before-Move: Narrative-Grounded World Visual Attention in Dynamic 3D Story Worlds","year":"2026","venue":"","arxiv_id":"2606.26964","aliases":["library:VG18"],"technical_key":"","catalog_id":"","library_id":"VG18","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2606.26964","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较只优化相机轨迹平滑，它先决定必须看见什么证据，再求实现这些观察的镜头运动。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"aaa28356bf7da2969198a75345e4f6dc69ed2d534200e084220da396bfe28b25","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0641","title":"Sabre: A Narrative Planner Supporting Intention and Deep Theory of Mind","year":"2021","venue":"","arxiv_id":"","aliases":["library:VG01"],"technical_key":"","catalog_id":"","library_id":"VG01","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://cdn.aaai.org/ojs/18896/18896-52-22662-1-2-20211004.pdf","status":"无需本文权重训练","resource_note":"Sabre 是符号状态空间 narrative planner；运行搜索，不涉及神经网络 fine-tuning。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较只管情节可达或角色全知的规划器，它把错误信念与深层心智理论纳入可解释行动规划。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"200585634950c7094b4f9635d47bf4b1340fc9c01e998cf77a611669c77c2b1c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0642","title":"StoryBlender: Inter-Shot Consistent and Editable 3D Storyboard with Spatial-Temporal Dynamics","year":"2026","venue":"","arxiv_id":"2604.03315","aliases":["library:VG14"],"technical_key":"","catalog_id":"","library_id":"VG14","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2604.03315","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较二维分镜的身份漂移与重采样式编辑，它通过全局资产和镜头变量解耦获得跨镜头一致性。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e48666ba62554c12cf3a3ed872a091495795e60ad872c4c99559ce3bb5abc1fc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0643","title":"VideoAuteur: Towards Long Narrative Video Generation","year":"2025","venue":"","arxiv_id":"2501.06173","aliases":["library:VG03"],"technical_key":"","catalog_id":"","library_id":"VG03","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2501.06173","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较只生成几秒片段，它专门处理多步骤内容的语义推进与视觉衔接。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c2f8b6b45d2d3fe7f604927599015614c02f58b7869be7802da20def2e7b1ba9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0644","title":"VideoGen-Eval: Agent-based System for Video Generation Evaluation","year":"2025","venue":"","arxiv_id":"2503.23452","aliases":["library:VG05"],"technical_key":"","catalog_id":"","library_id":"VG05","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2503.23452","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较固定指标和简单提示，它让评测随内容变化，并改善与人类偏好的对齐。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5ef402b749b019b1b947f9d024293c61ace2a6fb363101b641cfe2fd4df904fe","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0645","title":"Reasoning with Sampling: Your Base Model is Smarter Than You Think","year":"2025","venue":"","arxiv_id":"2510.14901","aliases":["library:VG06"],"technical_key":"","catalog_id":"","library_id":"VG06","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2510.14901","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较把推理增益完全归于 RL，它展示单靠改变采样分布也能释放已有能力并保留多样性。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"16f29eea22ecaeb25689b1f6a7d58be0fc48ff14eabe6b787c9c8cbf0da409cc","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0646","title":"Thinking with Video: Video Generation as a Promising Multimodal Reasoning Paradigm","year":"2025","venue":"","arxiv_id":"2511.04570","aliases":["library:VG07","reasoning:CVPR-2026-2789"],"technical_key":"","catalog_id":"","library_id":"VG07","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2511.04570","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较只用文本或静态图像表达推理，它考察动态视觉过程能否成为统一的推理载体。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f929964266efbc10cac200fc319bb6c9294a26e7a4d185792b1135ad39e8d9e1","reasoning_id":"CVPR-2026-2789","reasoning_role":"R2","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0647","title":"Learning to Reason in 4D: Dynamic Spatial Understanding for Vision Language Models（DSR Suite）","year":"2025","venue":"","arxiv_id":"2512.20557","aliases":["library:VG08"],"technical_key":"","catalog_id":"","library_id":"VG08","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2512.20557","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较静态空间题或直接堆入大量 3D 特征，它专门学习多对象、视角变化和细粒度时间关系。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4226672ec2144aa6f41a4987f90cbd89245903d064f54f6ea8fa88241e72695c","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0648","title":"Stereo World Model: Camera-Guided Stereo Video Generation（StereoWorld）","year":"2026","venue":"","arxiv_id":"2603.17375","aliases":["library:VG12"],"technical_key":"","catalog_id":"","library_id":"VG12","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2603.17375","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H20","count_min":24.0,"count_max":24.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"StereoWorld；20K steps，batch 24；49 frames，480×640","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.17375","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较先生成单目再估深转双目，它直接在 RGB 生成中约束视差与视角一致性。","training_signal":"不确定","hardware_bin":"medium","max_reported_ft_gpus":24.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"e22c23d6f4b2d3def8e16a04e82dbb2f186fe6fb6071b3acc346e73d37dccfa9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0649","title":"Training Object Permanence in World Models（WROP）","year":"2026","venue":"","arxiv_id":"2609.28654","aliases":["library:VG21"],"technical_key":"","catalog_id":"","library_id":"VG21","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2609.28654","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较小规模或依赖 VLM 评分的物理题，它提供可训练的 3D 任务及人评支持的恒存测量。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"2a3793de5610445637a05842f7b0242b2b34f837638fc4c437f74c1dc0cef338","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0650","title":"From Scores to Samples: Elastic Forcing for Autoregressive Video Generation","year":"2026","venue":"","arxiv_id":"2609.35491","aliases":["library:VG23"],"technical_key":"","catalog_id":"","library_id":"VG23","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2609.35491","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":23.0,"wall_hours_max":23.0,"time_bound":"approximate","gpu_hours_min":184.0,"gpu_hours_max":184.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Elastic Forcing Wan-14B post-training；5-step AR generator；184 GPU-hours","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2609.35491","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较依赖双评分网络的 DMD，它把少步自回归后训练改为基于样本的分布匹配。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"6c33c73918272576bb6011b0c0357e8a940ffe51b92a244afabfc66eca9f7932","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0651","title":"Scaling Video Generation for Reasoning: At What Cost?","year":"2026","venue":"","arxiv_id":"2609.36599","aliases":["library:VG24"],"technical_key":"","catalog_id":"","library_id":"VG24","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2609.36599","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H200","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Scaling Video Generation for Reasoning 受控训练每 run 1 H200；1M episodes；论文外推 H100-equivalent 成本是估计值，不当实测","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2609.36599","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较把低生成损失视为强推理，它显示外观拟合与正确跟踪隐状态可显著脱节，符号监督也能带来收益。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9de09bc32830529b4b1154f2f9012ed6b2f2dca8cd27d6f10369f1cca4e317da","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0652","title":"LongLive-Plug: Once-for-All Distillation","year":"2026","venue":"","arxiv_id":"2609.38154","aliases":["library:VG25"],"technical_key":"","catalog_id":"","library_id":"VG25","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2609.38154","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":2.5,"wall_hours_max":2.5,"time_bound":"approximate","gpu_hours_min":80.0,"gpu_hours_max":80.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"LongLive-Plug 一次 base distillation；700 iterations，80 GPU-hours；后续任务复用不再逐个蒸馏","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2609.38154","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较每个任务各蒸馏一次，它在骨干家族内复用加速能力，并保留一定 guidance 调节能力。","training_signal":"不确定","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"0c281d5761bde168f553854623c05f2aca8c6efccb42ef7afb6c3930f2ac86b9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0653","title":"LoopVL: Recurrent Visual Intelligence","year":"2026","venue":"","arxiv_id":"2609.38426","aliases":["library:VG26"],"technical_key":"","catalog_id":"","library_id":"VG26","topic_code":"M6","topic":"迭代、潜空间与组合推理","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2609.38426","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较只增加不共享的网络深度，它探索用参数复用获得更深的多模态计算与注意力调整。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"6d0ad699ff7eb5552e58bf15f5cb80b5f9bfb01a4137ed1465b2158c5c95c7e5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0654","title":"Finetuning with Sampling: SFT Learns Better Than You Think","year":"2026","venue":"","arxiv_id":"2610.02140","aliases":["library:VG27"],"technical_key":"","catalog_id":"","library_id":"VG27","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2610.02140","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较直接模仿外部轨迹或只靠在线 RL 探索，它从数据分布适配入手兼顾新能力学习与保持。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"629c1e5cc53448b22de0642cbad8f6dd1c1130e51fe42c80d8d543386f54d2a7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0655","title":"RPGAgent: Driving Coherent Story-to-Play Generation with an LLM-Based Multi-Agent System","year":"2026","venue":"","arxiv_id":"","aliases":["library:VG28"],"technical_key":"","catalog_id":"","library_id":"VG28","topic_code":"H3","topic":"人机共创与可编辑工作流","task":"未归入视频任务轴","directions":[],"source_url":"https://doi.org/10.1145/3772318.3790326","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较零散生成故事或资产，它面向新手打通从创意到可玩原型的工作流。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0656","title":"Procedural Generation of Narrative Worlds","year":"2022","venue":"","arxiv_id":"","aliases":["library:VG29"],"technical_key":"","catalog_id":"","library_id":"VG29","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://doi.org/10.1109/TG.2022.3216582","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较仅凭常见物体共现布置静态场景，它加入跨地点、随剧情变化的对象需求。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0657","title":"Controllable Texture Tiling via Diffusion Transformers with Transformed Rotary Embeddings","year":"2026","venue":"","arxiv_id":"2606.22945","aliases":["library:VG32"],"technical_key":"","catalog_id":"","library_id":"VG32","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2606.22945","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Texture Tiling LoRA rank 128 + modified input embedder；backbone 冻结；batch 32","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.22945","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较只能语义迁移材质或破坏图案的像素重采样，它解耦空间变换与内容生成以保留纹理结构。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"7b1a9c4cc2b7053804f994294d57eb165705893444ef77db0e1df0bc74e056b5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0658","title":"DuetSVG: Unified Multimodal SVG Generation with Internal Visual Guidance","year":"2025","venue":"","arxiv_id":"2512.10894","aliases":["library:VG33"],"technical_key":"","catalog_id":"","library_id":"VG33","topic_code":"G2","topic":"叙事、创作代理与持续生成","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2512.10894","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":336.0,"wall_hours_max":336.0,"time_bound":"approximate","gpu_hours_min":21504.0,"gpu_hours_max":21504.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"DuetSVG 全训练 pipeline 约 14 天；multi-task SFT 300K steps；不是单个输出优化","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2512.10894","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较只按文本序列写 SVG，它补入解码中的视觉约束，改善语义和几何一致性。","training_signal":"不确定","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"3c108d61d3e783dd875702fa02dceb5b1f2fcdc8efeddb9736a3840f0dcd0d93","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0659","title":"Identity-Preserving Human Reconstruction from a Single Image via 3D Token Reasoning（IPRM）","year":"未核实","venue":"","arxiv_id":"","aliases":["library:VG34"],"technical_key":"","catalog_id":"","library_id":"VG34","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://openreview.net/pdf?id=QlIteJADE6","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。当前来源是 OpenReview 稿件；本轮不据此确认录用状态。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较主要在二维特征中推理人体，它显式区分已观察身份与待补全部分来改善 3D 一致性。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0660","title":"Learning Stable Canonical Worlds for Novel View Synthesis and Beyond（CanonicalGS）","year":"2026","venue":"","arxiv_id":"2606.23027","aliases":["library:VG35"],"technical_key":"","catalog_id":"","library_id":"VG35","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2606.23027","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX A6000","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"CanonicalGS；RE10K 300K steps + DL3DV fine-tuning 100K steps；耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2606.23027","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较视角依赖预测随更多输入累积冗余噪声，它将证据归并到共享世界坐标并抑制不可靠观测。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"e07a74446fe0b7bcf242e6db9d6447679f12bdbc3a9efb405e8e27f481af01bf","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0661","title":"MeshFlow: Mesh Generation with Equivariant Flow Matching","year":"2026","venue":"","arxiv_id":"2606.23489","aliases":["library:VG36"],"technical_key":"","catalog_id":"","library_id":"VG36","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2606.23489","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":288.0,"gpu_hours_max":288.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"MeshFlow mesh flow-matching model；约 3 天；batch 256","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2606.23489","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较依赖任意序列顺序的网格生成，它把表示的置换结构直接纳入模型与匹配过程。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a06a1f12c9efbb6dac3dae00e0aee654ceeb992d6d8b1b77ad7eaf7912e499aa","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0662","title":"Video-Holmes: Can MLLM Think like Holmes for Complex Video Reasoning?","year":"2025","venue":"","arxiv_id":"2505.21374","aliases":["library:VG38"],"technical_key":"","catalog_id":"","library_id":"VG38","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2505.21374","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较单个显式线索即可回答的感知题，它要求主动定位、整合跨片段证据。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"83e2bc8d56cf31cf6a5b97acbd99e4226fcc9587098a4c727ae389d819f90ae0","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0663","title":"X2Video: Adapting Diffusion Models for Multimodal Controllable Neural Video Rendering","year":"2025","venue":"","arxiv_id":"2510.08530","aliases":["library:VG39"],"technical_key":"","catalog_id":"","library_id":"VG39","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2510.08530","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 5880 Ada","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"X2Video 全参数；50K steps，640×448；batch 32","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2510.08530","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较仅提示控制，它把多种可解释场景属性接入生成器并支持全局与局部控制。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"1b75dbeb0092ea3b7690e6bec7b1eadbe655a49f0f7d82bc967e8e5f7c62b3c7","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0664","title":"EffectMaker: Unifying Reasoning and Generation for Customized Visual Effect Creation","year":"2026","venue":"","arxiv_id":"2603.06014","aliases":["library:VG40"],"technical_key":"","catalog_id":"","library_id":"VG40","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2603.06014","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H20","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"EffectMaker；约 50K steps；81 frames，短边 704；耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.06014","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较每个特效分别微调，它通过语义和视觉双路条件支持参考驱动的定制与泛化。","training_signal":"不确定","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"0d431ee35188d764815e16d21e8f52412c527713175c4185723b7ddcfb20189d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0665","title":"FlexTraj: Image-to-Video Generation with Flexible Point Trajectory Control","year":"2025","venue":"","arxiv_id":"2510.08527","aliases":["library:VG41"],"technical_key":"","catalog_id":"","library_id":"VG41","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2510.08527","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H800","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":168.0,"wall_hours_max":168.0,"time_bound":"approximate","gpu_hours_min":1344.0,"gpu_hours_max":1344.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"FlexTraj HunyuanVideo-I2V self-attention LoRA rank 128；约一周；batch 1","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2510.08527","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较要求固定密度或严格对齐的轨迹控制，它让不同粒度的轨迹能被同一视频模型执行。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"897b5229ac32f5e1f8608e92076e3f7028bdab85449933904dd1dffb05b3364d","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0666","title":"Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond","year":"2026","venue":"","arxiv_id":"2604.22748","aliases":["library:VG42"],"technical_key":"","catalog_id":"","library_id":"VG42","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2604.22748","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较不同社区各用一套世界模型定义，它把预测、长程模拟与新证据驱动的模型修订明确分开。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e49430c87e4212c59fa0c238c553ecbfeb3dc6e948f42c254a2b85032da0abd5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0667","title":"Efficient Diffusion Language Models: A Comprehensive Survey","year":"2026","venue":"","arxiv_id":"","aliases":["library:VG44"],"technical_key":"","catalog_id":"","library_id":"VG44","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://raw.githubusercontent.com/MessiX77/Awesome-Efficient-dLLMs/main/files/Efficient_dLLMs_EMNLP2026.pdf","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较只比较解码步数，它讨论并行更新、缓存和部署的整体吞吐—质量取舍。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"48e721374ffa425a5adf35b39d61a98046dc96d9b63859064580e631a57f0106","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0668","title":"Generalized Kullback-Leibler Divergence Loss","year":"2025","venue":"","arxiv_id":"2503.08038","aliases":["library:VG45"],"technical_key":"","catalog_id":"","library_id":"VG45","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2503.08038","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"RTX 3090","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Generalized KL Loss 的 adversarial classifier training；不等同于生成器 fine-tuning","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2503.08038","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较直接使用逐样本 KL，它从损失结构解释并缓解教师分数造成的优化偏差。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c33c8751fbb03a88f4b01d8a5f7c460e34126fb2955ce7069493773d878639b9","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0669","title":"Easier Painting Than Thinking: Can Text-to-Image Models Set the Stage, but Not Direct the Play?（T2I-CoReBench）","year":"2025","venue":"","arxiv_id":"2509.03516","aliases":["library:VG47"],"technical_key":"","catalog_id":"","library_id":"VG47","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2509.03516","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较低密度、简单一对一推理测试，它分开诊断“画出指定元素”与“推断未明说元素”。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"5d59cf19bddf56dbcf7c846a99084d456750f2b5906c6457e4805fbe4b4dad8f","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0670","title":"QeRL: Quantization-Enhanced Low-Rank Reinforcement Learning for LLMs","year":"2025","venue":"","arxiv_id":"2510.11696","aliases":["library:VG48"],"technical_key":"","catalog_id":"","library_id":"VG48","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2510.11696","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":1.0,"count_max":1.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"QeRL 实测支持 32B GRPO 单卡；量化 + low-rank RL；完整训练时长需另核","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2510.11696","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较仅将量化当压缩，它实证研究量化噪声对策略熵与训练表现的作用。","training_signal":"不确定","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"beeb8bad255396146d2d9e62843c0495bdbf429c9f7e05535bc2e4890cb919a2","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0671","title":"Is This Edit Correct? A Multi-Dimensional Benchmark for Reasoning-Aware Image Editing（RE-Edit）","year":"2026","venue":"","arxiv_id":"2606.05172","aliases":["library:VG49"],"technical_key":"","catalog_id":"","library_id":"VG49","topic_code":"B1","topic":"能力诊断与评测基准","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2606.05172","status":"已核对资源阶段","resource_note":"","stages":[{"role":"aux","device":"RTX PRO 6000","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"RE-EditBench reasoning agent SFT；AdaLoRA 最终 rank 8，120 epochs；不是图像编辑器训练","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2606.05172","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"H100","count_min":6.0,"count_max":6.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"RE-EditBench reasoning agent GRPO；1 epoch；与 SFT 分阶段","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/html/2606.05172","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较只看视觉质量或字面指令，它检查改一处后隐含上下文约束是否仍成立。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"a97cc61737d56aaeb189975832480331806adb452cb82327b348f7584f2d4f16","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0672","title":"Learning to See through Illumination Extremes with Event Streaming in Multimodal Large Language Models（Event-MLLM）","year":"2026","venue":"","arxiv_id":"2603.27558","aliases":["library:VG51"],"technical_key":"","catalog_id":"","library_id":"VG51","topic_code":"M4","topic":"视觉证据、空间推理与流式感知","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2603.27558","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"RTX 5090D / A800","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Event Streaming MLLM；illumination correction 30 epochs，LoRA 1 epoch；数量/耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2603.27558","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较只依赖过曝或欠曝 RGB，它用另一种传感证据弥补丢失的结构信息。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"2ec6dc421a89a9d436a17216df930bc50188ca7c91cfdd6afa26212db264c988","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0673","title":"LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation","year":"2026","venue":"","arxiv_id":"2605.18739","aliases":["library:VG52"],"technical_key":"","catalog_id":"","library_id":"VG52","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2605.18739","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GB200","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"LongLive-2.0 AR training；600 iterations；NVFP4，SP 4；后续训练段另核","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2605.18739","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较只减少采样步数，它同时处理训练布局、推理计算、通信与解码瓶颈。","training_signal":"不确定","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"7332105efaa9be1dd91e890ed649f31771f89a3d1df3fc2ef2a9d702b0fa7924","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0674","title":"Mira-Scene: Pixel-Aligned Layouts for Generative 3D Scene Reconstruction","year":"2026","venue":"","arxiv_id":"2609.23796","aliases":["library:VG53"],"technical_key":"","catalog_id":"","library_id":"VG53","topic_code":"G3","topic":"三维资产、重建与动画","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2609.23796","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":32.0,"count_max":32.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Mira-Scene geometry stream 预训练；100K steps；TRELLIS 初始化，保留作者 pretraining 阶段定义","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/html/2609.23796","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Mira-Scene fine-tuning；与 32 GPU pretraining 分阶段","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2609.23796","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较难学习的稀疏位姿回归，它以有界像素对应表征连接对象几何和场景摆放。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"e08d577fe65c2ede607830fc70fca14d70134af62cb6ecfa0faec37b01e08993","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0675","title":"MindOmni: Unleashing Reasoning Generation in Vision Language Models with RGPO","year":"2025","venue":"","arxiv_id":"2505.13031","aliases":["library:VG54"],"technical_key":"","catalog_id":"","library_id":"VG54","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2505.13031","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较固定文本编码器或隐式语义映射，它支持多模态输入并用反馈优化需要推理的生成任务。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"13fc75c62dce3f499214b860603790df0edc86966464afd2a91bf14761947cde","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0676","title":"Scaling RL to Long Videos（LongVILA-R1 / MR-SP）","year":"2025","venue":"","arxiv_id":"2507.07966","aliases":["library:VG55"],"technical_key":"","catalog_id":"","library_id":"VG55","topic_code":"M2","topic":"RL／GRPO、过程反馈与数据选择","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2507.07966","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"LongVILA-R1 系统可在单个 8 GPU node 上运行 hour-long video RL；主模型完整耗时另缺","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2507.07966","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较短视频或昂贵重复预填，它使长输入推理的 RL rollout 更可行。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"80f7fec56e69ebc6e6f3a8e65c9715f1b9a8e68e2a1948a08832f2510be5cc52","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0677","title":"UniTok: A Unified Tokenizer for Visual Generation and Understanding","year":"2025","venue":"","arxiv_id":"2502.20321","aliases":["library:VG56"],"technical_key":"","catalog_id":"","library_id":"VG56","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2502.20321","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较将两种训练目标冲突视为必然，它显示表示瓶颈是统一 tokenizer 的重要障碍。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"80bc4f238b4e2dc13f9328d088094a01cb2db1772f4d91f0b89479943a7fb629","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0678","title":"VFMTok：Vision Foundation Models as Effective Visual Tokenizers / as Generalist Tokenizers","year":"2025","venue":"","arxiv_id":"2507.08441","aliases":["library:VG57"],"technical_key":"","catalog_id":"","library_id":"VG57","topic_code":"M7","topic":"自监督表征、JEPA 与 tokenizer","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2507.08441","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H800","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":36.0,"wall_hours_max":36.0,"time_bound":"approximate","gpu_hours_min":576.0,"gpu_hours_max":576.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"VFMTok tokenizer 适配；50 epochs，1.5 天；AR generator 训练成本另计","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2507.08441","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较从零训练且冗余的重建 tokenizer，它复用语义先验以提高 token 效率和生成收敛速度。","training_signal":"不确定","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"91d2865090eeef313734d08e301a4a33a06cab1688d951664d4993c2605deb0a","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0679","title":"S2VG: 3D Stereoscopic and Spatial Video Generation via Denoising Frame Matrix","year":"2025","venue":"","arxiv_id":"2508.08048","aliases":["library:VG59"],"technical_key":"","catalog_id":"","library_id":"VG59","topic_code":"G1","topic":"视频运动、镜头、编辑与特效","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2508.08048","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较分别补洞或逐视角生成，它利用现成视频先验联合维持时间与多视角一致性，无需额外训练。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"31fa4442fdaf19113357678487002cd7f955644e93525a6ee43fd398b2d086ac","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0680","title":"Veda: Scalable Video Diffusion via Distilled Sparse Attention","year":"2026","venue":"","arxiv_id":"2605.30325","aliases":["library:VG62"],"technical_key":"","catalog_id":"","library_id":"VG62","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2605.30325","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较只提高稀疏率，它强调稀疏模式是否保留原注意力结构决定质量与实际加速。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"51eafc35e7455443d41878243e001395f15a9e24aa1a67815d8011626c799822","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0681","title":"Visual Generation in the New Era: An Evolution from Atomic Mapping to Agentic World Modeling","year":"2026","venue":"","arxiv_id":"2604.28185","aliases":["library:VG65"],"technical_key":"","catalog_id":"","library_id":"VG65","topic_code":"B3","topic":"综述、概念框架与评价立场","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2604.28185","status":"无需本文权重训练","resource_note":"综述／立场文章；不把其引用论文的 GPU 配置归为本文训练。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较仅按模型架构或画质组织工作，它把结构、状态、交互与因果能力作为研究定位轴。","training_signal":"不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0f72f85635812e8ee561c7ec17952412faaee063c2831d38a1fb341d2b6f29a5","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0682","title":"AnyStep-WAM: Budget-Aligned Distillation and Adaptive Inference for World Action Models","year":"2026","venue":"","arxiv_id":"2609.33748","aliases":["library:VG67"],"technical_key":"","catalog_id":"","library_id":"VG67","topic_code":"M8","topic":"生成基础架构、优化、蒸馏与效率","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/html/2609.33748","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"AnyStep-WAM real-robot tasks；每任务 50 episodes，8K steps；batch 32；耗时未定位","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/html/2609.33748","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"不确定；资源备注列出已核对的具体模块／底座时以该项为准","training_datasets":"不确定；未独立确认完整训练数据清单，不把 benchmark 自动视作训练集","existing_method":"相较固定步数加速，它已针对不同动作的误差敏感性自适应分配计算。","training_signal":"不确定","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"d5fd0a02ec9cdda44601ee1abc990616ca162dc252206dc09c46d2af37975c6e","reasoning_id":"","reasoning_role":"","reasoning_topic":""},{"id":"R0683","title":"TV2TV: A Unified Framework for Interleaved Language and Video Generation","year":2026,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2026-1584"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Han_TV2TV_A_Unified_Framework_for_Interleaved_Language_and_Video_Generation_CVPR_2026_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"自建 Mixture-of-Transformers：3B/8B text tower + 同规模 video tower；Cosmos-Tokenize1-CV4x8x8-360p VAE","training_datasets":"CS:GO 95 小时带 controller action 的视频；YT-Temporal-1B 的体育视频，经场景切分、过滤及交错 caption 构建","existing_method":"","training_signal":"有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"7452f7b3bc162dd32b94559eb5de65c9913b0d6dc362c4ddf834fe9c889335b1","reasoning_id":"CVPR-2026-1584","reasoning_role":"R1","reasoning_topic":"通用生成框架 / 联合音视频"},{"id":"R0684","title":"DriveLaW: Unifying Planning and Video Generation in a Latent Driving World","year":2026,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2026-1722"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Xia_DriveLaW_Unifying_Planning_and_Video_Generation_in_a_Latent_Driving_CVPR_2026_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"LTX-Video 初始化的 2B Video DiT + 133M diffusion planner","training_datasets":"nuScenes + nuPlan；两阶段视频预训练后做 trajectory fine-tuning","existing_method":"","training_signal":"有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"c11be19e46058b7cf811e6cabe91663b1cc92a5b903fd1497f797fdae454292d","reasoning_id":"CVPR-2026-1722","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0685","title":"Training-free Motion Factorization for Compositional Video Generation","year":2026,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2026-2024"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Wang_Training-free_Motion_Factorization_for_Compositional_Video_Generation_CVPR_2026_paper.pdf","status":"无需本文权重训练","resource_note":"本文采用training-free motion factorization / guidance；不更新生成器权重。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"VideoCrafter-v2.0、CogVideoX-2B","training_datasets":"无新增训练：training-free 推理方法","existing_method":"","training_signal":"Training-free","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"66b992d2d86e5a8de4dc52c53cd4cd4a3baaac3433ea59bbda43e63dc49c4f3d","reasoning_id":"CVPR-2026-2024","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0686","title":"MoReGen: Multi-Agent Motion-Reasoning Engine for Code-based Text-to-Video Synthesis","year":2026,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2026-2326"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Bai_MoReGen_Multi-Agent_Motion-Reasoning_Engine_for_Code-based_Text-to-Video_Synthesis_CVPR_2026_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"代码驱动的物理模拟与渲染；不以 diffusion video backbone 为主生成器","training_datasets":"1,200 条人工整理 text–specification pairs 对 Qwen2.5-Coder-14B 做 SFT","existing_method":"","training_signal":"有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"74ec93fc89746b4d5f6119ede29c44ef0fff6f42e4333da65d5cca1b390c7e28","reasoning_id":"CVPR-2026-2326","reasoning_role":"R1","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0687","title":"SVBench: Evaluation of Video Generation Models on Social Reasoning","year":2026,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2026-2560"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Peng_SVBench_Evaluation_of_Video_Generation_Models_on_Social_Reasoning_CVPR_2026_paper.pdf","status":"无需本文权重训练","resource_note":"Social reasoning benchmark及四agent评测构建pipeline明确training-free。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"被测：Hailuo02-S、Kling2.5-turbo、Sora2-Pro、Veo-3.1、HunyuanVideo、LongCat-Video、LTX-1.0、Wan2.2","training_datasets":"无新增生成模型训练；benchmark 的 agent-based 构建流程为 training-free","existing_method":"","training_signal":"评测／无新增训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"19b0a8cdc97c4c8d69afbc6e5b001b8ff263f8dfff56c79ac57cc6c011d9b0da","reasoning_id":"CVPR-2026-2560","reasoning_role":"R3","reasoning_topic":"评测 / 数据 / 分析 / 安全"},{"id":"R0688","title":"Active Intelligence in Video Avatars via Closed-loop World Modeling","year":2026,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2026-2680"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/He_Active_Intelligence_in_Video_Avatars_via_Closed-loop_World_Modeling_CVPR_2026_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Wanx2.2 I2V + distilled LoRA（LightX2V）","training_datasets":"无任务特定训练；预训练 VLM 的 structured prompting 与现有 I2V 模型","existing_method":"","training_signal":"Training-free","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3611846a41c78b854ad278a74b9e0d6b7aee21ef043ded5af983bf9939a6603a","reasoning_id":"CVPR-2026-2680","reasoning_role":"R1","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0689","title":"LAMP: Language-Assisted Motion Planning for Controllable Video Generation","year":2026,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2026-3054"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Kizil_LAMP_Language-Assisted_Motion_Planning_for_Controllable_Video_Generation_CVPR_2026_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"轨迹可控的视频生成后端；主文所读部分未明确统一 checkpoint；Wan2.1 明确是 text-only baseline","training_datasets":"自建程序化 motion-program 数据：文字描述、DSL 及 3D 轨迹配对；训练 planner","existing_method":"","training_signal":"有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"f0687da3d73f11b47876e65749e35edca714121fa77a0572af0f78feb758b68f","reasoning_id":"CVPR-2026-3054","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0690","title":"Lighting-grounded Video Generation with Renderer-based Agent Reasoning","year":2026,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2026-3115"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Cai_Lighting-grounded_Video_Generation_with_Renderer-based_Agent_Reasoning_CVPR_2026_paper.pdf","status":"已核对资源阶段","resource_note":"单H100约40min属于VideoFrom3D baseline，已从本文训练统计排除。","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"本方法 LoRA：100K steps，704×1280，每卡 batch 2；文中单 H100 40min 是 VideoFrom3D baseline，未计入。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Cai_Lighting-grounded_Video_Generation_with_Renderer-based_Agent_Reasoning_CVPR_2026_paper.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-5B + proxy encoder / adapter / LoRA","training_datasets":"LiVERSet：LiVER-Real + LiVER-Syn；10k 训练、1k 测试，真实／合成各半","existing_method":"","training_signal":"有训练","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"f97d8be7312322008e68285062f3da42be8e0d02f21f308cfec1e4c43c2bcca3","reasoning_id":"CVPR-2026-3115","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0691","title":"TiViBench: Benchmarking Think-in-Video Reasoning for Video Generation","year":2026,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2026-3787"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Chen_TiViBench_Benchmarking_Think-in-Video_Reasoning_for_Video_Generation_CVPR_2026_paper.pdf","status":"无需本文权重训练","resource_note":"TiViBench及VideoTPO在推理时优化prompt，不额外训练生成器。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"被测：Wan2.2-I2V-A14B、Wan2.1-I2V-14B、HunyuanVideo-I2V、CogVideoX1.5-I2V；Veo3.1-fast、Sora2、Kling2.1","training_datasets":"无新增权重训练；TiViBench 是评测集，VideoTPO 不需要额外 tuning","existing_method":"","training_signal":"评测／Training-free","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"569bc9b6b1448ac6f744da64c3f08877c64dc7b4b52724c06c0e5e122774c98f","reasoning_id":"CVPR-2026-3787","reasoning_role":"R3","reasoning_topic":"评测 / 数据 / 分析 / 安全"},{"id":"R0692","title":"Reasoning Diffusion for Unpaired Test Time Out-of-distribution Text-Image to Video Generation","year":2026,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2026-3919"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Pan_Reasoning_Diffusion_for_Unpaired_Test_Time_Out-of-distribution_Text-Image_to_Video_CVPR_2026_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"ReasonDiff：Wan2.1 系列 + AlignFormer；主文所读部分未明确所用大小","training_datasets":"WebVid；每 0.2s 取帧并由 Llama-3.2-11B-Vision-Instruct 生成 frame captions；第一阶段联合训练 base model 与 AlignFormer，第二阶段冻结 base model、仅微调 AlignFormer","existing_method":"","training_signal":"有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"cfff1edff97cd83bc60aa6e358f23d07e19a7087bda4e675ce013a283dcfc280","reasoning_id":"CVPR-2026-3919","reasoning_role":"R1","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0693","title":"CoT-Edit: Let CoT Guide Instruction Video Editing","year":2026,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2026-4022"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Liang_CoT-Edit_Let_CoT_Guide_Instruction_Video_Editing_CVPR_2026_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Wan2.2-5B editor + Mask / Reverse Connectors","training_datasets":"Mask 分支：ADE20K、PhraseCut、YouTube-VOS、OVIS；Editor：AnyEdit、UltraEdit、SEED-Data-Edit、EditWorld、Señorita2M、Ditto 和自建编辑数据；联合阶段约 100k 内部有 mask 的 before/after pairs","existing_method":"","training_signal":"有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"d6a8dc99ed95f79a9d3eb94bcf129c5e0f39d4625ae73fdf713d8e192213334d","reasoning_id":"CVPR-2026-4022","reasoning_role":"R1","reasoning_topic":"Editing / 修改已有视频"},{"id":"R0694","title":"Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning","year":2026,"venue":"ICLR","arxiv_id":"2601.16163","aliases":["reasoning:ICLR-2026-1627"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2601.16163","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":3072.0,"gpu_hours_max":3072.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"LIBERO：Cosmos-Predict2 全参数微调，40K steps，global batch 1920。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2601.16163","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":1536.0,"gpu_hours_max":1536.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"RoboCasa：45K steps，global batch 800；与 LIBERO 为替代实验，不相加。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2601.16163","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":384.0,"gpu_hours_max":384.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"ALOHA：185 demonstrations / 4 tasks，50K steps；真实机器人采集与搜索推理另计。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2601.16163","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Cosmos-Predict2-2B，全参数 post-training","training_datasets":"目标平台 robot demonstrations：LIBERO、RoboCasa、真实 ALOHA（四任务共 185 demos）；规划阶段另用本策略收集的 rollout 数据","existing_method":"","training_signal":"有训练","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"60e521a2cd1505fee0515354e2e2663bb725d7bc34e6c866a2b333f2fa0f7d2a","reasoning_id":"ICLR-2026-1627","reasoning_role":"R2","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0695","title":"World-In-World: World Models in a Closed-Loop World","year":2026,"venue":"ICLR","arxiv_id":"2510.18135","aliases":["reasoning:ICLR-2026-2063"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2510.18135","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":29,"gpu_hours_max":29,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"Table 8：SVD1.5B domain post-training约40K clips，peak84GB；设备数未明确，84GB不推定单卡。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2510.18135","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":5,"gpu_hours_max":5,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"Table 8：LTX-Video2B，peak61GB；设备数未明确。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2510.18135","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":74,"gpu_hours_max":74,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"Table 8：Wan2.1-14B LoRA，peak57GB；设备数未明确。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2510.18135","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":15,"gpu_hours_max":15,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"Table 8：Cosmos-Predict2-2B，peak71GB；设备数未明确。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2510.18135","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"被测视频模型包括 SVD-1.5B、LTX-Video-2B、Hunyuan-13B、Wan2.1-14B、Wan2.2-5B/A14B、Cosmos-Predict2-2B、NWM、Runway-Gen4；另有图像 WM 对照","training_datasets":"有 post-training 实验：Habitat-Sim 自采 action-observation 数据；RLBench 每 manipulation task 200 demonstrations；不是所有被测模型都训练","existing_method":"","training_signal":"基准含模型训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"e144b1778af367c9afe342f13664257475e2a16c63fbb3b5a30f771265415b7b","reasoning_id":"ICLR-2026-2063","reasoning_role":"R3","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0696","title":"Dual-IPO: Dual-Iterative Preference Optimization for Text-to-Video Generation","year":2026,"venue":"ICLR","arxiv_id":"2502.02088","aliases":["reasoning:ICLR-2026-2954"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2502.02088","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":128.0,"count_max":128.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":336.0,"wall_hours_max":336.0,"time_bound":"approximate","gpu_hours_min":43008.0,"gpu_hours_max":43008.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"T2V generator alignment：batch 512，每轮约两周，每轮约 100K preference pairs；不是 critic 的 8 A100 配置。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2502.02088","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"GPU（型号未定位）","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VILA 13B/40B critic warm-start：5 epochs，16 video frames；不冒充生成器微调资源。","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/pdf/2502.02088","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":24.0,"wall_hours_max":24.0,"time_bound":"approximate","gpu_hours_min":192.0,"gpu_hours_max":192.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"SRPO critic update：每轮约一天，可在不同生成器之间复用；与 generator alignment 分开。","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/pdf/2502.02088","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CogVideoX-2B / 5B 与 WanX-1.3B（按本文命名）","training_datasets":"约 1M 组合式合成 captions；每轮生成约 100K top-bottom preference pairs；VidGen-1M 子集用于 real-data regularization；critic 用少量 CoT warm-start 与每轮约 20K pseudo-labels","existing_method":"","training_signal":"reward model 与 generator 交替 post-training","hardware_bin":"large","max_reported_ft_gpus":128.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"988845a802a62621de8bde840a6156ab43e313a5cb9a561a72889bc80938bfc9","reasoning_id":"ICLR-2026-2954","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0697","title":"BindWeave: Subject-Consistent Video Generation via Cross-Modal Integration","year":2026,"venue":"ICLR","arxiv_id":"2510.00438","aliases":["reasoning:ICLR-2026-3626"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2510.00438","status":"已核对资源阶段","resource_note":"原文512 xPUs保留原设备单位，不归入GPU卡数分档。","stages":[{"role":"ft","device":"xPU（型号未明确）","count_min":512.0,"count_max":512.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"BindWeave 两阶段训练，1M 筛选视频对，global batch 512；原文写 xPU，不能当作 512 张 NVIDIA GPU。","comparable":true,"is_gpu":false,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2510.00438","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Wan 系列 DiT video foundation model；本文所读配置段落未明确精确版本／规模","training_datasets":"OpenS2V-5M 过滤得到约 1M video-text pairs；两阶段 curriculum fine-tuning","existing_method":"","training_signal":"有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"0b649d0348218dcce06861e3cd68682420da171dce2e29174da5b80a07beecc9","reasoning_id":"ICLR-2026-3626","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0698","title":"Captain Cinema: Towards Short Movie Generation","year":2026,"venue":"ICLR","arxiv_id":"2507.18634","aliases":["reasoning:ICLR-2026-4101"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2507.18634","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":2304.0,"gpu_hours_max":2304.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Appendix Table 4 的 keyframe generator 配置；正文 40K steps 与表中分阶段 20K steps 需结合阅读。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2507.18634","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":256.0,"count_max":256.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":200.0,"wall_hours_max":200.0,"time_bound":"approximate","gpu_hours_min":51200.0,"gpu_hours_max":51200.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Appendix Table 4 的 Seaweed-3B video generator；正文写 15K steps，表中写 10K total，存在版本内不一致，耗时按表记录。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2507.18634","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"FLUX.1-dev 做 interleaved keyframe generation；Seaweed-3B 做 bottom-up video generation","training_datasets":"专门整理的 cinematic interleaved text-image/video 数据；确切来源组成／数量在所读部分不确定","existing_method":"","training_signal":"有训练","hardware_bin":"large","max_reported_ft_gpus":256.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"f6078ca2bb29a33421429a48250095b1b2666d176459227dde266716dcaa0ee7","reasoning_id":"ICLR-2026-4101","reasoning_role":"R1","reasoning_topic":"长视频 / 多镜头 / 记忆"},{"id":"R0699","title":"FilMaster: Bridging Cinematic Principles and Generative AI for Automated Film Generation","year":2026,"venue":"ICLR","arxiv_id":"2506.18899","aliases":["reasoning:ICLR-2026-4273"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2506.18899","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Kling Elements，支持多张参考图；ElevenLabs 用于 voice-over","training_datasets":"未报告新增模型权重训练；440k film clips 为 RAG 参考库，46,826 audio assets 为检索音频库","existing_method":"","training_signal":"检索／推理流程","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"72074a1f20e75577a35a964071835952c55e6ceab5b7f45333679e5906f67dba","reasoning_id":"ICLR-2026-4273","reasoning_role":"R1","reasoning_topic":"长视频 / 多镜头 / 记忆"},{"id":"R0700","title":"AUHead: Realistic Emotional Talking Head Generation via Action Units Control","year":2026,"venue":"ICLR","arxiv_id":"2602.09534","aliases":["reasoning:ICLR-2026-4471"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2602.09534","status":"已核对资源阶段","resource_note":"","stages":[{"role":"aux","device":"A100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Stage 1 Qwen-Audio-Chat：原文 24 GPU-hours，不是 24 小时墙钟耗时。","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/pdf/2602.09534","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Stage 2 Hallo / MEMO 的 AU adapter：原文 12 GPU-hours；两个底座实验不自动求和。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2602.09534","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"A100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":24,"gpu_hours_max":24,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"Stage 1：原文直接报告总GPU-hours，4卡；不再乘4。","comparable":true,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/pdf/2602.09534","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":12,"gpu_hours_max":12,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"Stage 2：原文直接报告总GPU-hours，4卡；不再乘4。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2602.09534","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Hallo V1、MEMO + AU adapter","training_datasets":"MEAD、CREMA 的 subject-independent train splits；先训练 AU prediction，再训练视频 AU adapter","existing_method":"","training_signal":"有训练","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"b6b477f03ee6d287554acc08c7ba973f2b86a7af9cb3a93d6442bef5afb39bea","reasoning_id":"ICLR-2026-4471","reasoning_role":"R1","reasoning_topic":"人物动画 / Talking Head / HOI"},{"id":"R0701","title":"Compositional Visual Planning via Inference-Time Diffusion Scaling","year":2026,"venue":"ICLR","arxiv_id":"2603.02646","aliases":["reasoning:ICLR-2026-4937"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2603.02646","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H200","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"训练短时 DiT-L / XL 视觉模型；数量和耗时未定位。inference-time composition 本身不需要额外训练。","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/pdf/2603.02646","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"自训短程 DiT-L / DiT-XL video diffusion model + Cosmos tokenizer","training_datasets":"ManiSkill 派生的组合操作 demonstrations，以及真实操作示范；短程模型从长程 demos 抽短片训练，长程组合无需再训练","existing_method":"","training_signal":"底座有训练／组合推理 Training-free","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"39c6419dc12663a4c30089b2a91a13b62dab15654059b598ec1820c13369ad3a","reasoning_id":"ICLR-2026-4937","reasoning_role":"R2","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0702","title":"ReGen: Agentic Video World Modeling with Synergized Reasoning and Generation","year":2026,"venue":"NeurIPS","arxiv_id":"","aliases":["reasoning:NeurIPS-2026-0460"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openreview.net/forum?id=QMvrZxy8W9","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"causal diffusion generator；具体 checkpoint 不确定","training_datasets":"自建 action-grounded video dataset，超过 2M segments；具体源数据与训练拆分尚未核实","existing_method":"","training_signal":"有训练／细节待核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"NeurIPS-2026-0460","reasoning_role":"R1","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0703","title":"Seeing the World through Any Eyes","year":2026,"venue":"NeurIPS","arxiv_id":"","aliases":["reasoning:NeurIPS-2026-0902"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openreview.net/forum?id=GEPHFHPezV","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"EgoEye；底座型号不确定","training_datasets":"EgoScape：19.5K 同步 exo-ego pairs + 24.2K in-the-wild egocentric videos；实际训练划分不确定","existing_method":"","training_signal":"有训练／细节待核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"NeurIPS-2026-0902","reasoning_role":"R1","reasoning_topic":"Editing / 修改已有视频"},{"id":"R0704","title":"Bridging Scene Generation and Planning: Driving with World Model via Unifying Vision and Motion Representation","year":2026,"venue":"NeurIPS","arxiv_id":"2603.14948","aliases":["reasoning:NeurIPS-2026-2519"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2603.14948","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"TA-DWM 两阶段 domain training：nuPlan 200K + nuScenes 100K steps；只明确 CogVideoX VAE 和结构，不能推定全部 diffusion 权重初始化。","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/pdf/2603.14948","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"RTX 3090","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"冻结编码器，NAVSIM planner 50 epochs；FAR 10 epochs 的单独资源未拆出。","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://arxiv.org/pdf/2603.14948","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"WorldDrive：CogVideoX 初始化的 3D VAE / DiT + trajectory-aware driving WM","training_datasets":"WM：nuPlan 训练 200k iterations 后 nuScenes 100k；planner：NAVSIM navtrain 50 epochs","existing_method":"","training_signal":"有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"97c3ec4e8fcaa7712ce127752debb4fc9087e6c18b695f4e3e8b711d3ade5647","reasoning_id":"NeurIPS-2026-2519","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0705","title":"Hierarchical Denoising For Multi-Step Visual Reasoning","year":2026,"venue":"NeurIPS","arxiv_id":"2607.15278","aliases":["reasoning:NeurIPS-2026-2747"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2607.15278","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":384.0,"gpu_hours_max":384.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"预训练 Wan2.2-5B-TI2V + hierarchical latent tree；18K 推理视频，224×224，100K steps，每个 main run。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2607.15278","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-5B-TI2V + hierarchical latent tree / sparse attention","training_datasets":"自建 18,000 条 synthetic reasoning videos；maze、Hanoi、one-line drawing、sliding puzzle、Sokoban、water pouring 各 3,000","existing_method":"","training_signal":"有训练","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"5e6e6061382a19cc39d309f59bb04d57e7e46e7ecc28cedc70ab7c23afb403e9","reasoning_id":"NeurIPS-2026-2747","reasoning_role":"R2","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0706","title":"OptiWorld: Optimal Control for Video World Generation under Physical Constraints","year":2026,"venue":"NeurIPS","arxiv_id":"","aliases":["reasoning:NeurIPS-2026-3007"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openreview.net/forum?id=ElToIAx2OM","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"视频生成 backbone 不确定；OptiWorld 是外置 optimal-control framework","training_datasets":"摘要称控制层作用于 inference time；是否有其他组件训练及 dataset 尚未核实","existing_method":"","training_signal":"训练情况不确定","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"NeurIPS-2026-3007","reasoning_role":"R1","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0707","title":"PlasticMem: Adding Temporal Reasoning to Diffusions for Consistent Long Video Generation","year":2026,"venue":"NeurIPS","arxiv_id":"","aliases":["reasoning:NeurIPS-2026-3426"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openreview.net/forum?id=xTHR2rP4s6","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"PlasticMem diffusion renderer；底座型号不确定","training_datasets":"短视频训练；数据集名称、数量和来源尚未核实","existing_method":"","training_signal":"有训练／数据待核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"NeurIPS-2026-3426","reasoning_role":"R1","reasoning_topic":"长视频 / 多镜头 / 记忆"},{"id":"R0708","title":"Agentic Video Editing from Underspecified Requests","year":2026,"venue":"NeurIPS","arxiv_id":"","aliases":["reasoning:NeurIPS-2026-3727"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openreview.net/forum?id=pGPKhPBqR0","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"Aurora unified video DiT；具体 backbone 不确定","training_datasets":"executable edit planning / reference selection 的监督数据，以及 tool use / instruction refinement preference pairs；名称及规模未核实","existing_method":"","training_signal":"有训练／数据待核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"NeurIPS-2026-3727","reasoning_role":"R1","reasoning_topic":"Editing / 修改已有视频"},{"id":"R0709","title":"Bernini: Latent Semantic Planning for Video Diffusion","year":2026,"venue":"NeurIPS","arxiv_id":"2605.22344","aliases":["reasoning:NeurIPS-2026-4117"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2605.22344","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":40.0,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Video editing training：FSDP 后 per-GPU memory 从 72GB 降到 40GB；这里是峰值占用，不是卡容量；硬件数量、型号和耗时未定位。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2605.22344","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-A14B DiT renderer","training_datasets":"20M 自构 video pairs；内部 T2I/T2V/I2I 与自构编辑／reference-video 数据；公开部分包括 UniREdit-100K、Pico-Banana-400K、OpenS2V-Top200K、Señorita-Controllable、SynCamVideo 等；四阶段训练","existing_method":"","training_signal":"有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"1022fd1a03f7193e9a7413e04090cba4d778c88facfc67421b9ea69bd72ced2e","reasoning_id":"NeurIPS-2026-4117","reasoning_role":"R1","reasoning_topic":"通用生成框架 / 联合音视频"},{"id":"R0710","title":"Auteur: Language-Driven Cinematographic Framing for Human-Centric Video Generation","year":2026,"venue":"NeurIPS","arxiv_id":"2606.01900","aliases":["reasoning:NeurIPS-2026-4145"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2606.01900","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":8.0,"wall_hours_max":8.0,"time_bound":"approximate","gpu_hours_min":8.0,"gpu_hours_max":8.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Qwen2.5-VL-7B-Instruct，LoRA rank 128 + multimodal merger；只训练 director，不训练后端视频模型；正文 34K aligned trajectories，附录约 300K image-text samples / 1 epoch。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2606.01900","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"VerseCrafter 视频后端；另一接口组合 Kimodo 人体动作生成与视频渲染","training_datasets":"程序化合成 + CondensedMovies 实景的 text-motion-DSL 数据；读取的 arXiv 版称 34K trajectories，训练约 300K image-text samples","existing_method":"","training_signal":"有训练","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"0433535acfc37f059638058c4dd50c47543d481d921f84243b851a3e5fe69e17","reasoning_id":"NeurIPS-2026-4145","reasoning_role":"R1","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0711","title":"UNIVERSE: Unified Video Action Models for Autonomous Driving with Flexible Mask-Modulated Modality Generation","year":2026,"venue":"NeurIPS","arxiv_id":"2607.05133","aliases":["reasoning:NeurIPS-2026-4719"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2607.05133","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H20","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Wan2.2-TI2V-5B adaptation：20K + 10K steps，832×480；卡数和耗时未定位。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2607.05133","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-TI2V-5B，统一 mask-modulated DiT；沿用 text encoder / causal VAE","training_datasets":"NAVSIM / OpenScene 的驾驶训练数据；共同监督 future video 与 trajectories","existing_method":"","training_signal":"有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"8f5d9d7aeef7a30667436c6b7408c172163ee14712faf145da10be7b8494e6ca","reasoning_id":"NeurIPS-2026-4719","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0712","title":"How Far Are Video Models from True Multimodal Reasoning?","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-0039"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/12354.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"被测：Wan2.2-14B、HunyuanVideo1.5、UniVideo、CogVideoX1.5-5B、LTX-2；Seedance2.0/1.5、Veo3.1-Fast、Wan2.7、Sora2","training_datasets":"不训练生成模型；AVE 用少量人工标注进行 prompt 优化，非权重训练","existing_method":"","training_signal":"评测／prompt 优化","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"dca175676a3d2454597d9f29e1e7b01ff7957565d6ff05e5a711d3f1ffd206db","reasoning_id":"ECCV-2026-0039","reasoning_role":"R3","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0713","title":"UniDrive-WM: Unified Understanding, Planning and Generation World Model For Autonomous Driving","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-0106"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/4407.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"UniDrive-WM 全部实验配置；Vicuna 1.5 LoRA rank16，含 AR / AR+Diff 分支；各阶段耗时未拆出。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/4407.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"自建 VLM-centric UniDrive-WM；MoVQGAN discrete AR 或 continuous AR+diffusion decoder","training_datasets":"Bench2Drive base split：950 scenes 训练／50 验证；nuScenes 官方 train/val；第二阶段加入 VQA 数据（具体源名待核实）","existing_method":"","training_signal":"有训练","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"6d0da76deb4df99d797eaea3412a68b7ca0713280c7ac8c17dace8c05e12faf6","reasoning_id":"ECCV-2026-0106","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0714","title":"What if? Emulative Simulation with World Models for Situated Reasoning","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-0260"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/7191.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"HunyuanVideo、CogVideoX、Wan 系列；细节实验包含 Wan2.1 LoRA / Wan2.2 prompt extension","training_datasets":"WanderDream-Gen 的两类训练子集，主要 LoRA 8 epochs；源自 HM3D、ScanNet++ 的路径视频与真实捕获；总数据 15.8K videos，训练数量以 split 为准","existing_method":"","training_signal":"有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e968c1d8bc428aded4d9165fc445434603dd74a019030aeab78a3533c2f3b74f","reasoning_id":"ECCV-2026-0260","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0715","title":"Nexus-Vid: Efficient Frequency Bridging with Homogeneous Latent Space for Video Unified Models","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-0267"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/8809.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Omni-Video 已对齐的 Wan2.1-T2V-1.3B small generator；Wan2.1-T2V-14B large generator 在推理时接续","training_datasets":"复用 unified small-generator 训练／对齐框架；具体训练 corpus 未从所读主文核实","existing_method":"","training_signal":"有训练／数据待核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"fbfcb1bbfed47dd811da8f4c95084ba9f7334c3bea254aae9d4545421e6e0d35","reasoning_id":"ECCV-2026-0267","reasoning_role":"R1","reasoning_topic":"效率 / 架构 / 蒸馏 / 实时"},{"id":"R0716","title":"LangDriveCTRL: Natural Language Controllable Driving Scene Editing with Multi-modal Agents","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-0278"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/6406.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":192.0,"gpu_hours_max":192.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"TrajectoryCrafter 初始化；阶段 1：33 frames，40K iterations，batch4。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/6406.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":4.0,"count_max":4.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":192.0,"gpu_hours_max":192.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"阶段 2：81 frames，20K iterations，batch4；与阶段 1 顺序执行，整个两阶段约 96h。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/6406.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CogVideoX，初始化 TrajectoryCrafter checkpoint，再训练 video harmonization","training_datasets":"训练使用 mesh 替换 3DGS 渲染对和 cycle-reconstruction 生成的 noisy-rendering pairs；实际源训练数据名称尚未核实","existing_method":"","training_signal":"有训练","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"6481a7c697302087567286f6b3def4dfa297cd560c6dccc7c12d841bf78f9a01","reasoning_id":"ECCV-2026-0278","reasoning_role":"R1","reasoning_topic":"Editing / 修改已有视频"},{"id":"R0717","title":"DIVER: Disentangling Camera–Object and Active–Passive Motion for Video Generation","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-0321"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/11272.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"GPU（型号未定位）","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Wan2.1-14B：camera encoder / trajectory encoder / self-attention，15K iterations，batch16。单 A100 15min 为推理配置。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/11272.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.1-14B；更新 camera/trajectory encoders 与 self-attention","training_datasets":"Panda-70M + Wild-SDG1M 选出 76K static-view videos；生成 43K paired dynamic-view videos；另 3.4K SyncMaster synthetic interaction videos","existing_method":"","training_signal":"有训练","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"04d08e92672f7e776437c23cb3b27e7464d967b9c5b226233cac657dd9b349cf","reasoning_id":"ECCV-2026-0321","reasoning_role":"R2","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0718","title":"InteractiveAvatar: Real-Time Streaming Video Generation for Consistent and Intent-Aware Avatars","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-0382"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/1287.pdf","status":"已核对资源阶段","resource_note":"H100的FPS/latency表用于推理评测，不能据此推断训练卡数。","stages":[{"role":"ft","device":"H100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"InteractiveAvatar四阶段训练，576p；Stage1 50K、Stage2 30K、Stage3/4各20K steps，FSDP。训练资源依据§4.1，FPS表另为推理。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/1287.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-5B，四阶段 audio-image-to-video / autoregressive distillation；LSVM","training_datasets":"约 3M 清洗 clips：HDTF、VFHQ、VoxCeleb2、CelebV-Text、AVSpeech；OpenHumanVid 影视数据；自有长对话／丰富动作数据","existing_method":"","training_signal":"有训练","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"45220f122f2591b60c4c80fff65307b79589d84bb62f2127ee34e99727a080d5","reasoning_id":"ECCV-2026-0382","reasoning_role":"R1","reasoning_topic":"人物动画 / Talking Head / HOI"},{"id":"R0719","title":"PhyGDPO: Physics-Aware Groupwise Direct Preference Optimization for Physically Consistent Text-to-Video Generation","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-0574"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/4366.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":144.0,"wall_hours_max":144.0,"time_bound":"approximate","gpu_hours_min":1152.0,"gpu_hours_max":1152.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"PhyGDPO：480×832，batch8，6天，BF16，LoRA-Switch Reference；数据构建成本另计。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/4366.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.1-14B + LoRA-Switch Reference；groupwise DPO","training_datasets":"自建 PhyVidGen-135K；包含物理增强数据构建后的真实视频／文本与生成偏好组","existing_method":"","training_signal":"有训练","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"edb5f7a2930a9368753dd1cfe32885988b3910407b624ce5bc000352ba3d0aed","reasoning_id":"ECCV-2026-0574","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0720","title":"Benchmarking Scientific Understanding and Reasoning for Video Generation using VideoScience-Bench","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-0677"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/11226.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"被测：Sora2、Veo3、Seedance1.0-Pro、Kling-v2.5-Turbo-Pro、Hailuo2.3、Ray2、Wan2.5-T2V-Preview","training_datasets":"无新增视频模型训练；专家构建测试任务与评分规则","existing_method":"","training_signal":"评测／无新增训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"4efd9565081d8ce5e27eac21ec55d2f31a4e1dcdc9c3926f07ca19a6a74d459a","reasoning_id":"ECCV-2026-0677","reasoning_role":"R3","reasoning_topic":"评测 / 数据 / 分析 / 安全"},{"id":"R0721","title":"Making Avatars Interact: Towards Text-Driven Human-Object Interaction for Controllable Talking Avatars","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-1141"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/4488.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"InteractAvatar：PIM motion stream + AIM audio-video stream；精确 pretrained backbone 未从主文核实","training_datasets":"人体／物体互动视频经检测生成 motion supervision；joint-generation : external-driving 训练比例 4:1；原始训练 corpus 名称待补充材料核实","existing_method":"","training_signal":"有训练／数据源待核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1a2031a396d8221cf190051aad3fc1634570e01a795ad62bc2f4665d1fe09d6d","reasoning_id":"ECCV-2026-1141","reasoning_role":"R1","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0722","title":"Better Call CineCrew: Consistent Ultra-Long Narrative-to-Film Generation","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-1189"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/2445.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"off-the-shelf keyframe / I2V generators；支持异构后端，主文所读实验未明确统一 checkpoint","training_datasets":"无新增生成器训练；通过 structured QA 与 prompt / plan 修订执行修复","existing_method":"","training_signal":"Training-free orchestration","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"7269da7a9fb5ce39590208d62ff04928cd7b8e154a339e5e92dcd290a620487b","reasoning_id":"ECCV-2026-1189","reasoning_role":"R1","reasoning_topic":"长视频 / 多镜头 / 记忆"},{"id":"R0723","title":"DriveVA: Video Action Models are Zero-Shot Drivers","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-1203"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/611.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H20","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Wan2.2-TI2V-5B，20K + 10K steps，832×480；没有由 global batch 推断卡数。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/611.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-TI2V-5B，统一 video-action DiT 与 causal VAE","training_datasets":"NAVSIM 驾驶训练；基于现成 video backbone full fine-tuning","existing_method":"","training_signal":"有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"33698482ab6f7d2f617edc9b2bbca526f71348c0fcfb1d5f7c7ef29a38365583","reasoning_id":"ECCV-2026-1203","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0724","title":"Target-Bench: Can Video World Models Achieve Mapless Path Planning with Semantic Targets?","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-1501"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/9345.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A800","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Wan2.2-TI2V-5B LoRA，DiffSynth-Studio；RTX PRO6000 / RTX4090 是后续推理与评测。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/9345.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"被测 Sora2、Veo3.1/fast、Wan2.5-I2V-preview、Wan2.2-plus/flash、Wan2.1-plus/turbo；另微调 Wan2.2-TI2V-5B","training_datasets":"Target dataset 的 train split，121-frame clips；比较普通 LoRA 与 shifting-frame 四倍增广","existing_method":"","training_signal":"基准含模型训练","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"fb2197d717098f0a189e48c0b17ae701fc025139b42bc90162c3b50c81de8523","reasoning_id":"ECCV-2026-1501","reasoning_role":"R3","reasoning_topic":"评测 / 数据 / 分析 / 安全"},{"id":"R0725","title":"Wan-R1: Verifiable-Reinforcement Learning for Generalizable Video Reasoning","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-1512"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/13964.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Wan-R1 LoRA，原文只定位 H200 machines；卡数和墙钟耗时未定位。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/13964.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-TI2V-5B；以 maze demonstrations 微调的 Wan-SFT 为 RL 起点","training_datasets":"VR-Bench 的迷宫示范用于 SFT / RL；机器人路径任务使用 Target 数据与参考轨迹；具体 split 数量未核实","existing_method":"","training_signal":"SFT + RL","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"4288940fb84c844ce77f519d4bf4e232e6ef24c3706456e518210522d6ea7143","reasoning_id":"ECCV-2026-1512","reasoning_role":"R2","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0726","title":"DeforM: Reasoning-Guided Physics-Aware Video Generation via Spatial-Temporal Masking","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-1597"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/1472.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Wan2.2-TI2V-5B","training_datasets":"DeforM-Injection：WISA-80K 中筛选约 6K deformation clips，LoRA rank 256；DeforM-Free 不新增训练","existing_method":"","training_signal":"Training-free 与 LoRA 两个版本","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"bee8f42c5fc36dd90a3c490ce01eb356bdb75c65ef3a1154b09718f7647b1d15","reasoning_id":"ECCV-2026-1597","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0727","title":"EvoWorld: A World-Model-Centric Framework for Continuous Self-Evolution of Modular Embodied Skills","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-1602"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/10502.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"action-conditioned latent world model；具体生成 backbone 未核实","training_datasets":"共享 offline experience repository 与验证后的 imagined trajectories；文中以 DROID 为场景 grounding 示例；精确预训练数据配方与 split 未核实","existing_method":"","training_signal":"模块化迭代训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"26e3ad274d0468b50fc060e1bf5c43ebe15ff8265e47b392cea4ccfdd989247e","reasoning_id":"ECCV-2026-1602","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0728","title":"From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-1910"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/5598.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Vid-PRE prompt rewriter 的 SFT + GRPO；1.5K difficult examples，约 3K GRPO steps，ZeRO-3；不是对视频生成器做同样微调。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/5598.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Vid-PRE 接入 Wan2.2、Wan2.5、Wan2.6-Flash；精确 Wan2.2 checkpoint 未核实","training_datasets":"自建约 20K 高质量样本做 SFT；1.5K 高难样本做 GRPO，使用 text-based reward；训练的是 reasoner，不是视频生成器","existing_method":"","training_signal":"Reasoner SFT + RL；生成器冻结","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"ce29e90aece8e25da9b4d83ea6c1d4b64e6753486dbe1941e33470082e4c8d15","reasoning_id":"ECCV-2026-1910","reasoning_role":"R3","reasoning_topic":"评测 / 数据 / 分析 / 安全"},{"id":"R0729","title":"Evaluating Reasoning Coherence in Video Generative Models with Text and Visual Hints","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-1966"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/9137.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"被测 Veo-3.1、Veo-3.1-fast、Sora-2、Seedance-1.0-pro/fast、Kling-v2.1、Cosmos-Predict2-14B","training_datasets":"无新增训练；使用现成生成模型的默认配置","existing_method":"","training_signal":"仅评测","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"da16f57a68b654623588163c7edb81f4709abb7e601b7da518a25f400145a15a","reasoning_id":"ECCV-2026-1966","reasoning_role":"R3","reasoning_topic":"评测 / 数据 / 分析 / 安全"},{"id":"R0730","title":"LMGenDrive: Bridging Multimodal Understanding and Generative World Modeling for End-to-End Driving","year":2026,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2026-2217"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/9549.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H800","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":48.0,"wall_hours_max":48.0,"time_bound":"approximate","gpu_hours_min":384.0,"gpu_hours_max":384.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"LMGenDrive 三阶段总约两天，50K steps 按20K/20K/10K分配；Vicuna-7B + SD1.5 + AnimateDiff，ZeRO-2。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/9549.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Stable Diffusion 1.5 + AnimateDiff；ResNet / BEV 视觉编码","training_datasets":"视觉编码器：CARLA 采集的 3M expert frames；随后联合单步训练，再冻结生成器做多步 LLM 训练；后两阶段独立数据集名称 / 规模未核实","existing_method":"","training_signal":"三阶段训练","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"7aff47afeeabdc8839ecfdd35e0bc1a5a7157be69efe2df8786ae183fd1dd442","reasoning_id":"ECCV-2026-2217","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0731","title":"DreamDojo: A Real-Time Robot World Model from Large-Scale Human Videos","year":2026,"venue":"ICML","arxiv_id":"2602.06949","aliases":["reasoning:ICML-2026-1741"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2602.06949","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H100","count_min":256.0,"count_max":256.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"DreamDojo 2B/14B pretraining：各140K steps，batch1024；不算后续 embodiment fine-tuning。","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/pdf/2602.06949","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":128.0,"count_max":128.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"DreamDojo embodiment post-training：默认50K steps，batch512，全参数；墙钟耗时未定位。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2602.06949","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Cosmos-Predict2.5，DreamDojo 2B / 14B；700M latent action Transformer","training_datasets":"世界模型：In-lab、EgoDex、DreamDojo-HV，采样比 1:2:10；latent action model 还使用 Unitree G1、Fourier GR-1、AgiBot、YAM 内部机器人数据；目标机器人数据用于 post-training","existing_method":"","training_signal":"预训练 + 机器人 post-training + 蒸馏","hardware_bin":"large","max_reported_ft_gpus":128.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"834acda2949873fbd02724687810d04e8b9a50017b1d44553e1a22c87540aea3","reasoning_id":"ICML-2026-1741","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0732","title":"Bridging Your Imagination with Audio-Video Generation via a Unified Director","year":2026,"venue":"ICML","arxiv_id":"","aliases":["reasoning:ICML-2026-1784"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openreview.net/forum?id=FHMsceUzfS","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"UniMAGE unified director，Mixture-of-Transformers；底层 audio-video generator checkpoint 未核实","training_datasets":"摘要明确 interleaved text-image data 与 first-interleaving-then-disentangling training；具体 dataset 名称、规模未核实","existing_method":"","training_signal":"有训练，细节未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"ICML-2026-1784","reasoning_role":"R1","reasoning_topic":"通用生成框架 / 联合音视频"},{"id":"R0733","title":"DiLA: Disentangled Latent Action World Models","year":2026,"venue":"ICML","arxiv_id":"2605.15725","aliases":["reasoning:ICML-2026-1911"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2605.15725","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":4.0,"count_max":4.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":24.0,"wall_hours_max":24.0,"time_bound":"approximate","gpu_hours_min":96.0,"gpu_hours_max":96.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"DiLA：123M trainable modules + frozen DINOv2/RAE；teacher-forcing 30K steps约24h，后续1K rollout adaptation另计。","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/pdf/2605.15725","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":4.0,"count_max":4.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"同节点 latent rollout fine-tuning 1K steps，时长未单列；不能把第一阶段24h当作本阶段耗时。下游VP2适配亦有单独协议。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2605.15725","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"DiLA：DINOv2-base-with-registers encoder、RAE ViT-XL decoder 冻结；约 123M 可训练模块","training_datasets":"Something-Something-v2、RT-1、RECON、LoopNav；下游 Robosuite 5,000 trajectories、RoboDesk 35,000 perturbed scripted trajectories 做 action adaptation / fine-tuning","existing_method":"","training_signal":"训练 + action adaptation","hardware_bin":"small","max_reported_ft_gpus":4.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"c40be47f3e594ab031d56ad3d76cdb513ea156ca2e035b4b904e33178c03f422","reasoning_id":"ICML-2026-1911","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0734","title":"MVISTA-4D: View-Consistent 4D World Model with Test-Time Action Inference for Robotic Manipulation","year":2026,"venue":"ICML","arxiv_id":"","aliases":["reasoning:ICML-2026-3051"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://mercerai.github.io/MVISTA-4D/icml2026.pdf","status":"未定位可比微调配置","resource_note":"单RTX4090用于真实机器人数据采集工作站，不能归为训练GPU。已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.2-TI2V-5B DiT；RGB / depth 双分支与多视角融合","training_datasets":"分别在 RLBench（超过 8,000 trajectories）、RoboTwin2（超过 10,000 trajectories）与自采真实机器人数据上训练；真实数据规模未核实","existing_method":"","training_signal":"SFT + action VAE 训练；推理时优化","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"455e16c6912d62055adcf0771912e87e284de616e339040b681eeaa324676ee5","reasoning_id":"ICML-2026-3051","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0735","title":"Co-Evolving Latent Action World Models","year":2026,"venue":"ICML","arxiv_id":"2510.26433","aliases":["reasoning:ICML-2026-4190"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2510.26433","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H200","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":100.0,"wall_hours_max":100.0,"time_bound":"approximate","gpu_hours_min":800.0,"gpu_hours_max":800.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"CoLA-World joint WARM8K + E2E52K；预训练 OpenSora 与 LAM 联合优化。两阶段baseline约75h / 110h属对照，不与joint相加。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2510.26433","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"主要使用 OpenSora；另验证 Wan2.1-T2V-1.3B","training_datasets":"混合约 30% OXE、20% AgiBot、50% human video；后者列出 SSv2、RH20T、Ego4D、EgoPAT3D、EGTEA Gaze+、HOI4D、EPIC-KITCHENS、HO-Cap、HoloAssist；LIBERO 下游训练 split 用于 probing / adaptation","existing_method":"","training_signal":"warm-up + joint training + 下游适配","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"d17e77a4317527eee7d0b943115c42653f2e0c5714438c3b66ef5fbc199010bf","reasoning_id":"ICML-2026-4190","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0736","title":"ReaForest: Fostering Generative Video Reasoning for Spatial Planning","year":2026,"venue":"ICML","arxiv_id":"","aliases":["reasoning:ICML-2026-5339"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openreview.net/forum?id=Ktvb8rqFV7","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"Video Generation Model backbone 未核实","training_datasets":"ReaGen-27k，摘要明确用于激活空间规划推理能力；训练配方未核实","existing_method":"","training_signal":"有训练 + test-time scaling，细节未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"ICML-2026-5339","reasoning_role":"R2","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0737","title":"MotiMotion: Motion-Controlled Video Generation with Visual Reasoning","year":2026,"venue":"ICML","arxiv_id":"2605.22818","aliases":["reasoning:ICML-2026-6035"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2605.22818","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Wan2.2-I2V-A14B","training_datasets":"OpenVid：先 5K steps 学 motion control，再 3K steps 对 50% 样本扰动轨迹，学习 confidence-aware control","existing_method":"","training_signal":"视频控制模块 fine-tuning；reasoner training-free","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"7c94b100ca9dc314fea0d22ff868cb15813a7b151aa1f8c2b52766c41c79fddb","reasoning_id":"ICML-2026-6035","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0738","title":"VideoStudio: Generating Consistent-Content and Multi-Scene Videos","year":2024,"venue":"ECCV","arxiv_id":"","aliases":["reasoning:ECCV-2024-1178"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/07783.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VideoStudio-Img，20K iterations，batch512；后续video训练分开。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/07783.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VideoStudio-Vid，SDXL加temporal layers，WebVid-10M + HD-VG-130M，480K iterations，batch128。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/07783.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"VideoStudio-Img 基于 SD2.1；VideoStudio-Vid 基于 SDXL 加 temporal attention / convolution","training_datasets":"VideoStudio-Img 新 cross-attention：LAION-2B；VideoStudio-Vid：WebVid-10M、HD-VG-130M","existing_method":"","training_signal":"图像与视频模块训练；LLM 推理","hardware_bin":"medium","max_reported_ft_gpus":64.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"1b7164ab2211055773317d36f974c030f4cf60e98bde635ebc59a99df916f1bf","reasoning_id":"ECCV-2024-1178","reasoning_role":"R1","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0739","title":"LLM-grounded Video Diffusion Models","year":2024,"venue":"ICLR","arxiv_id":"2309.17444","aliases":["reasoning:ICLR-2024-1470"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2309.17444","status":"无需本文权重训练","resource_note":"LLM-grounded Video Diffusion明确training-free；attention guidance的采样成本另计。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"ModelScope text-to-video diffusion","training_datasets":"无新增权重训练；WebVid-2M 用于 retrieval baseline / in-context examples，不是本文模型训练集","existing_method":"","training_signal":"Training-free","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"be35bbbe44a27df092eaa2d3020d0de4e6470ef317cfd881ab294ebe0596f0f3","reasoning_id":"ICLR-2024-1470","reasoning_role":"R1","reasoning_topic":"通用生成框架 / 联合音视频"},{"id":"R0740","title":"Video Language Planning","year":2024,"venue":"ICLR","arxiv_id":"2310.10625","aliases":["reasoning:ICLR-2024-1627"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2310.10625","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"TPUv3 pods","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"每domain的基础video diffusion训练3天；高分辨率superresolution另1天。原文计数单位pods，非GPU。","comparable":false,"is_gpu":false,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/pdf/2310.10625","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"TPUv3 pods","count_min":64.0,"count_max":64.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":24.0,"wall_hours_max":24.0,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"每domain的12B PaLM-E heuristic/policy fine-tuning；不换算GPU-hours。","comparable":true,"is_gpu":false,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2310.10625","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"aux","device":"TPUv3 pods","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":24.0,"wall_hours_max":24.0,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"每domain的goal-conditioned LAVA policy训练；不与video model混计。","comparable":false,"is_gpu":false,"role_label":"辅助模型训练","source_url":"https://arxiv.org/pdf/2310.10625","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"UniPi 风格多尺度 pixel-space video diffusion：24×40 → 48×80 → 192×320","training_datasets":"Language Table；RT-1；约 1,200 双臂 kitchen stacking teleoperation demonstrations；generalization 实验另混合自采 7DoF、Bridge、RT-2、Ego4D、EPIC-KITCHEN、LAION-400M","existing_method":"","training_signal":"按 domain 训练 video / VLM / policy","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"639d05df7a7c94c11786a21b8fab609f9a33469627d7854dbbc3927c366ea650","reasoning_id":"ICLR-2024-1627","reasoning_role":"R2","reasoning_topic":"通用生成框架 / 联合音视频"},{"id":"R0741","title":"GEVRM: Goal-Expressive Video Generation Model For Robust Visual Manipulation","year":2025,"venue":"ICLR","arxiv_id":"2502.09268","aliases":["reasoning:ICLR-2025-2467"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2502.09268","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"预训练 text-guided video DiT（引用 Zheng et al., 2024）；T5 encoder；精确 checkpoint 未核实","training_datasets":"CALVIN、Bridge 的 training split；真机自采 demonstration 数据","existing_method":"","training_signal":"video planner / VAE / goal policy 训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b7341535bbda9704ea8e448e48dbf4e50fd55502675a7464cfc76135baa2aee0","reasoning_id":"ICLR-2025-2467","reasoning_role":"R2","reasoning_topic":"通用生成框架 / 联合音视频"},{"id":"R0742","title":"FLIP: Flow-Centric Generative Planning as General-Purpose Manipulation World Model","year":2025,"venue":"ICLR","arxiv_id":"2412.08261","aliases":["reasoning:ICLR-2025-3077"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2412.08261","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"flow-conditioned latent video diffusion Transformer，使用 SDXL VAE；另有 flow diffusion proposal model","training_datasets":"LIBERO-LONG（50×10 videos）、FMB（1K single-object / 100 multi-object videos）、自采 folding / unfolding（各 40 videos）；另有 Bridge-V2 模块实验、LIBERO-90 预训练变体；policy 使用少量 action-labeled demos","existing_method":"","training_signal":"flow / dynamics / value / policy 训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"af872c4c8adc31087a3cb7ec9499f0c845f644a1c930cc9c37392140157c99a1","reasoning_id":"ICLR-2025-3077","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0743","title":"VSTAR: Generative Temporal Nursing for Longer Dynamic Video Synthesis","year":2025,"venue":"ICLR","arxiv_id":"2403.13501","aliases":["reasoning:ICLR-2025-3138"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2403.13501","status":"无需本文权重训练","resource_note":"VSTAR调整attention与初始噪声，无需fine-tuning；噪声优化不是权重更新。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"以 VideoCrafter2 为主要展示的 pretrained T2V diffusion；适用于 temporal attention 架构","training_datasets":"无新增模型训练","existing_method":"","training_signal":"Training-free inference","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"9c7d05032271269bffc7255e4893e180edefd23f92102512525fb4e73af280ed","reasoning_id":"ICLR-2025-3138","reasoning_role":"R1","reasoning_topic":"长视频 / 多镜头 / 记忆"},{"id":"R0744","title":"RoboDreamer: Learning Compositional World Models for Robot Imagination","year":2024,"venue":"ICML","arxiv_id":"","aliases":["reasoning:ICML-2024-1859"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://raw.githubusercontent.com/mlresearch/v235/main/assets/zhou24f/zhou24f.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"V100","count_min":100.0,"count_max":100.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"RoboDreamer base video diffusion，约100卡，batch256；时长未定位，非已训练视频大模型的轻量适配。","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://raw.githubusercontent.com/mlresearch/v235/main/assets/zhou24f/zhou24f.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"自训 compositional video diffusion U-Net；64² base + 128² / 256² upsamplers","training_datasets":"RT-1 robot demonstrations，约 70K；按组合条件训练 video world model；inverse dynamics 用相邻帧与机器人状态 / 动作监督","existing_method":"","training_signal":"world model + inverse dynamics 训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b85b3dd2b004e0874101ad222982dab2e48f31dd727dfe057baafae986dbc39f","reasoning_id":"ICML-2024-1859","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0745","title":"Position: Video as the New Language for Real-World Decision Making","year":2024,"venue":"ICML","arxiv_id":"","aliases":["reasoning:ICML-2024-2015"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://raw.githubusercontent.com/mlresearch/v235/main/assets/yang24z/yang24z.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"多种示例：Transformer-XL + VQ-VAE / MaskGIT；pixel-space 3D U-Net diffusion；Genie 风格 masked dynamics","training_datasets":"Position paper 汇集多种例子；正文明确有 Open X-Embodiment 训练的生成示例；其他实例的完整语料与划分未核实","existing_method":"","training_signal":"立场论文，含多个训练模型示例","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"581fe9d04f64ed32bebe520199e691b5efc5cf49fc1a15a7c117ba16313bb1f2","reasoning_id":"ICML-2024-2015","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0746","title":"Empowering World Models with Reflection for Embodied Video Prediction","year":2025,"venue":"ICML","arxiv_id":"","aliases":["reasoning:ICML-2025-0897"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://raw.githubusercontent.com/mlresearch/v267/main/assets/chi25b/chi25b.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A800","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"EVA video diffusion训练表：20K steps；Ensemble-LoRA / domain adaptation，未定位全pipeline总时间。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://raw.githubusercontent.com/mlresearch/v267/main/assets/chi25b/chi25b.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"DynamiCrafter 视频生成器，EVA-Generator / EVA","training_datasets":"EVA-Instruct 500K：来源含 OXE / RT-1、RoboVQA、Ego4D、Ego-Exo4D、CALVIN；VLM instruction stage 另含 MIMIC-IT、LLaVA instruction、VideoChatGPT","existing_method":"","training_signal":"多阶段 VLM / generator / adapter 训练","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"a39572a82b119d9803132bfdbbb3bbccbf243f53df41a7d5718f6324857c9efa","reasoning_id":"ICML-2025-0897","reasoning_role":"R1","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0747","title":"Closed-Loop Visuomotor Control with Generative Expectation for Robotic Manipulation","year":2024,"venue":"NeurIPS","arxiv_id":"","aliases":["reasoning:NeurIPS-2024-0388"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://proceedings.neurips.cc/paper_files/paper/2024/file/fad8962279154544ed69bb63eb14d677-Paper-Conference.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"aux","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":10.0,"wall_hours_max":10.0,"time_bound":"approximate","gpu_hours_min":80.0,"gpu_hours_max":80.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"CLOVER反馈控制policy：CALVIN ABC 10 epochs，batch128；不作为video diffusion主微调时长。","comparable":false,"is_gpu":true,"role_label":"辅助模型训练","source_url":"https://proceedings.neurips.cc/paper_files/paper/2024/file/fad8962279154544ed69bb63eb14d677-Paper-Conference.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CLOVER 自训 RGB-D video diffusion；Imagen-style U-Net + causal temporal attention","training_datasets":"CALVIN（policy 明确 ABC training split）；真机 50 collected demonstrations 微调 diffusion planner","existing_method":"","training_signal":"video planner + feedback policy 训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"076e26f59da33238ee5fe12d9f89baafdaf2b254d9b76f27e0ff9c543773e0c7","reasoning_id":"NeurIPS-2024-0388","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0748","title":"Compositional 3D-aware Video Generation with LLM Director","year":2024,"venue":"NeurIPS","arxiv_id":"","aliases":["reasoning:NeurIPS-2024-1367"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://proceedings.neurips.cc/paper_files/paper/2024/file/edbeca7811f9365c924c72a8a9bce83a-Paper-Conference.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"opt","device":"A100","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"3D scene / Gaussian与路径优化实验；不是视频backbone微调，1000 iterations/location不换算耗时。","comparable":false,"is_gpu":true,"role_label":"逐实例表示优化","source_url":"https://proceedings.neurips.cc/paper_files/paper/2024/file/edbeca7811f9365c924c72a8a9bce83a-Paper-Conference.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"LucidDreamer 生成 3D scene、HumanGaussian 生成 humanoid、Motion-X 动作；Stable Diffusion 提供 SDS prior","training_datasets":"未训练新的通用视频 backbone；使用预训练专家并对每个场景执行 SDS / 位置尺度优化；Motion-X 在此为动作资源","existing_method":"","training_signal":"预训练专家 + 每场景优化","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"913d0cbcbf4d24198283d439a1dff72134e89401b7c5ffb1fd186b36b89ab4dc","reasoning_id":"NeurIPS-2024-1367","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0749","title":"TalkCuts: A Large-Scale Dataset for Multi-Shot Human Speech Video Generation","year":2025,"venue":"NeurIPS","arxiv_id":"","aliases":["reasoning:NeurIPS-2025-2377"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openreview.net/forum?id=4a0w7AkrY7","status":"全文获取受限","resource_note":"已检查公开入口；未取得可检索全文。不能由摘要或项目介绍推定 fine-tuning 为零。","stages":[],"no_weight_update":false,"has_fulltext":false,"screening":"公开页面核查；全文获取受限","hardware_candidate":false,"model":"Orator multimodal video generation baseline；具体 checkpoint 未核实","training_datasets":"TalkCuts：164K clips、超过 500 小时、10K+ identities；摘要明确训练使用 TalkCuts，具体 train split 未核实","existing_method":"","training_signal":"有训练，细节未核实","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"","reasoning_id":"NeurIPS-2025-2377","reasoning_role":"R1","reasoning_topic":"评测 / 数据 / 分析 / 安全"},{"id":"R0750","title":"UniTransfer: Video Concept Transfer via Progressive Spatio-Temporal Decomposition","year":2025,"venue":"NeurIPS","arxiv_id":"2509.21086","aliases":["reasoning:NeurIPS-2025-2664"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2509.21086","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"UniTransfer self-supervised OpenVid pretraining 200K iterations；与下面监督适配分开。","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://arxiv.org/pdf/2509.21086","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"监督concept transfer：TikTok / UBC Fashion 10K iterations；animal-specific适配另按文中设置，耗时未定位。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2509.21086","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CogVideoX-based dual-to-single-stream DiT；精确参数规模未核实","training_datasets":"OpenVid 自监督预训练；TikTok + UBC Fashion 人体任务微调；OpenAnimal 动物任务微调","existing_method":"","training_signal":"self-supervised pretraining + supervised fine-tuning","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"eff36d02981e688de599ed6d64619938d375cadc06d6fb8e822ca27b6428daf7","reasoning_id":"NeurIPS-2025-2664","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0751","title":"MindJourney: Test-Time Scaling with World Models for Spatial Reasoning","year":2025,"venue":"NeurIPS","arxiv_id":"2507.12508","aliases":["reasoning:NeurIPS-2025-4954"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2507.12508","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":576.0,"gpu_hours_max":576.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"MindJourney所用video diffusion world models约3天；test-time search仍需另外运行，不属于完全零训练的整体研究。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2507.12508","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Search World Model：Wan2.2-TI2V-5B + ReCamMaster 风格 camera conditioning","training_datasets":"SWM camera encoder / projector：Habitat 在 HM3D 生成的 50K navigation clips + RealEstate10K + DL3DV-10K；VLM 无新增微调","existing_method":"","training_signal":"world-model 控制模块训练；VLM test-time scaling","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"a39d9799f025e492f597996a946515e4ca96520c9e09776c1275929ff2c8c7b3","reasoning_id":"NeurIPS-2025-4954","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0752","title":"Learning 3D Persistent Embodied World Models","year":2025,"venue":"NeurIPS","arxiv_id":"2505.05495","aliases":["reasoning:NeurIPS-2025-5490"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2505.05495","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"H100","count_min":12.0,"count_max":12.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":72.0,"wall_hours_max":72.0,"time_bound":"approximate","gpu_hours_min":864.0,"gpu_hours_max":864.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Persistent Embodied WM：视频domain adaptation及memory训练，约3天；阶段时长未进一步拆分。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2505.05495","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CogVideoX，加入 RGB-D / action channels 与 3D memory blocks；具体 checkpoint 规模未核实","training_datasets":"Habitat + HM3D 自采约 50K trajectories，每条最多 500 steps；场景分 train / test；先适配 backbone，再冻结 backbone 训练 memory blocks","existing_method":"","training_signal":"两阶段 world-model / memory 训练","hardware_bin":"medium","max_reported_ft_gpus":12.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"e975935d86b93fc40c267b02db53a5421d79e2e727dff1f9d1e9f79cce0dc58b","reasoning_id":"NeurIPS-2025-5490","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0753","title":"Dysen-VDM: Empowering Dynamics-aware Text-to-Video Diffusion with LLMs","year":2024,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2024-0659"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2024/papers/Fei_Dysen-VDM_Empowering_Dynamics-aware_Text-to-Video_Diffusion_with_LLMs_CVPR_2024_paper.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Dysen-VDM训练设置；含原文direct fine-tuning条件，未拆出每个dataset/run的时间。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://openaccess.thecvf.com/content/CVPR2024/papers/Fei_Dysen-VDM_Empowering_Dynamics-aware_Text-to-Video_Diffusion_with_LLMs_CVPR_2024_paper.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Latent VDM autoencoder + 3D U-Net；加入 RGTrm dynamic scene graph encoder","training_datasets":"WebVid 预训练 latent VDM 与 DSG-conditioned VDM；另报告 UCF-101 / MSR-VTT 的直接 fine-tuning 设置","existing_method":"","training_signal":"video / graph encoder 训练；LLM in-context reasoning","hardware_bin":"medium","max_reported_ft_gpus":16.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"6c5309df480bd6c4b3ac8112b692661752c451d01605beaf6d1a6258847919aa","reasoning_id":"CVPR-2024-0659","reasoning_role":"R1","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0754","title":"Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving","year":2024,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2024-1231"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2024/papers/Wang_Driving_into_the_Future_Multiview_Visual_Forecasting_and_Planning_with_CVPR_2024_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Stable Diffusion 初始化 multiview temporal diffusion（Drive-WM）；具体 SD 版本未核实","training_datasets":"nuScenes：700 training scenes；生成 counterfactual driving videos 另用于 planner fine-tuning","existing_method":"","training_signal":"world-model 训练；另有 planner 数据增强实验","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b18f6cb247fdb1b51e3e22700222cc35d58faef6554c6661f33442fdd04ab9fd","reasoning_id":"CVPR-2024-1231","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0755","title":"BlobGEN-Vid: Compositional Text-to-Video Generation with Blob Video Representations","year":2025,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2025-0149"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2025/papers/Feng_BlobGEN-Vid_Compositional_Text-to-Video_Generation_with_Blob_Video_Representations_CVPR_2025_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"VideoCrafter2（U-Net）与 CogVideoX-5B（DiT）两个 backbone","training_datasets":"约 1M dense-blob annotated videos，来自 OpenVid-1M、VidGEN-1M、少量 HD-VILA；indoor variant 使用 BlobGEN-3D 处理的 ScanNet++；CogVideoX 变体只用约一半主数据","existing_method":"","training_signal":"冻结 backbone，训练新增 control layers","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"69c27f036a76043b6a811f089b9ac37952f00fdda734aa34702caa76760fea6b","reasoning_id":"CVPR-2025-0149","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0756","title":"PhyT2V: LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation","year":2025,"venue":"CVPR","arxiv_id":"","aliases":["reasoning:CVPR-2025-1103"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/CVPR2025/papers/Xue_PhyT2V_LLM-Guided_Iterative_Self-Refinement_for_Physics-Grounded_Text-to-Video_Generation_CVPR_2025_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"CogVideoX（2B / 5B）、OpenSora 1.2、VideoCrafter","training_datasets":"无新增模型训练；不需要额外训练数据","existing_method":"","training_signal":"Training-free iterative refinement","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ca2937d067a03f7c54182ef90d03a757890a4b944955bad17e60f55d18466940","reasoning_id":"CVPR-2025-1103","reasoning_role":"R1","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0757","title":"Aether: Geometric-Aware Unified World Modeling","year":2025,"venue":"ICCV","arxiv_id":"","aliases":["reasoning:ICCV-2025-0423"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Zhu_Aether_Geometric-Aware_Unified_World_Modeling_ICCV_2025_paper.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":80.0,"count_max":80.0,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":336.0,"wall_hours_max":336.0,"time_bound":"approximate","gpu_hours_min":26880.0,"gpu_hours_max":26880.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"AETHER从CogVideoX-5B-I2V初始化，两阶段训练约两周，global batch320；不是单卡重建推理成本。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Zhu_Aether_Geometric-Aware_Unified_World_Modeling_ICCV_2025_paper.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CogVideoX-5B-I2V 初始化 Aether；新增 depth / raymap input-output channels","training_datasets":"按照 DA-V / The-Matrix 路线采集并标注的 synthetic RGB-D videos；精确数据源清单 / 规模未核实","existing_method":"","training_signal":"synthetic post-training","hardware_bin":"large","max_reported_ft_gpus":80.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"85e52558fa11368e93ff7573ef560ae3b661b1a6a42acc6f904ba0e2caec2353","reasoning_id":"ICCV-2025-0423","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0758","title":"VLIPP: Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior","year":2025,"venue":"ICCV","arxiv_id":"","aliases":["reasoning:ICCV-2025-0426"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Yang_VLIPP_Towards_Physically_Plausible_Video_Generation_with_Vision_and_Language_ICCV_2025_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Go-with-the-Flow image-to-video model","training_datasets":"所读方法与实验采用现成模型执行推理，未发现本文新增权重训练；底座原始训练数据未核实","existing_method":"","training_signal":"现成模型推理；新增训练未见明确报告","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"bcf8950a373f812c01c88632edd9112419e4766f7783ea0a132c1ef2391794f2","reasoning_id":"ICCV-2025-0426","reasoning_role":"R1","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0759","title":"Causal-Entity Reflected Egocentric Traffic Accident Video Synthesis","year":2025,"venue":"ICCV","arxiv_id":"","aliases":["reasoning:ICCV-2025-1440"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Li_Causal-Entity_Reflected_Egocentric_Traffic_Accident_Video_Synthesis_ICCV_2025_paper.pdf","status":"已核对资源阶段","resource_note":"原文RTX3090每卡22GB为含糊报告，未自动更正为24GB容量。","stages":[{"role":"ft","device":"RTX 3090","count_min":2.0,"count_max":2.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Causal-VidSyn UNet训练，224×224；原文写每卡22GB RAM，与常见24GB卡容量不一致，未填卡容量。1×A80080GB属于DiT对照。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Li_Causal-Entity_Reflected_Egocentric_Traffic_Accident_Video_Synthesis_ICCV_2025_paper.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Causal-VidSyn 3D U-Net video diffusion + CTS / CTG modules；具体预训练 checkpoint 未核实","training_datasets":"MM-AU 的 9,727 事故视频中采 6,492 near-crash-to-crash clips；Drive-Gaze 提供 driver fixation 监督；三个 progressive training stages","existing_method":"","training_signal":"三阶段 video / causal grounding 训练","hardware_bin":"small","max_reported_ft_gpus":2.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"8cde22bad096a8b684096d198a7fc2596b5917dfc42217944669c10f7c214f74","reasoning_id":"ICCV-2025-1440","reasoning_role":"R1","reasoning_topic":"质量 / 物理 / 推理 / 训练优化"},{"id":"R0760","title":"DrivingGPT: Unifying Driving World Modeling and Planning with Multi-modal Autoregressive Transformers","year":2025,"venue":"ICCV","arxiv_id":"","aliases":["reasoning:ICCV-2025-1517"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Chen_DrivingGPT_Unifying_Driving_World_Modeling_and_Planning_with_Multi-modal_Autoregressive_ICCV_2025_paper.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"L20","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":12.0,"wall_hours_max":12.0,"time_bound":"upper","gpu_hours_min":96.0,"gpu_hours_max":96.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"DrivingGPT 110M主模型完整训练100K iterations，batch16；within12h为上界，非大型视频模型fine-tuning。","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Chen_DrivingGPT_Unifying_Driving_World_Modeling_and_Planning_with_Multi-modal_Autoregressive_ICCV_2025_paper.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"DrivingGPT：从头训练 110M autoregressive Transformer；LlamaGen visual tokenizer","training_datasets":"NAVSIM navtrain；另对比均匀采样 nuPlan 651K sequences 与 NAVSIM train 104K sequences 的数据配方","existing_method":"","training_signal":"Transformer from scratch；使用预训练 tokenizer","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"abb83e9f4421f715723f72d88cfc4c04061881f9ebd517b92b85b65fff8dc182","reasoning_id":"ICCV-2025-1517","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0761","title":"VEGGIE: Instructional Editing and Reasoning Video Concepts with Grounded Generation","year":2025,"venue":"ICCV","arxiv_id":"","aliases":["reasoning:ICCV-2025-1565"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/ICCV2025/supplemental/Yu_VEGGIE_Instructional_Editing_ICCV_2025_supplemental.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":16.0,"count_max":16.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VEGGIE Stage 1：MagicBrush/SD1.5，25K steps，约800M trainable。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://openaccess.thecvf.com/content/ICCV2025/supplemental/Yu_VEGGIE_Instructional_Editing_ICCV_2025_supplemental.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"A100","count_min":32.0,"count_max":32.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"VEGGIE Stage 2：LLM LoRA rank64 + temporal video model，2.5K steps，8frames；约1.3B trainable。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://openaccess.thecvf.com/content/ICCV2025/supplemental/Yu_VEGGIE_Instructional_Editing_ICCV_2025_supplemental.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"MagicBrush 初始化 SD1.5 UNet，加 AnimateDiff 风格 temporal layers","training_datasets":"Stage 1：Seed-Data-Edit、LISA、gRefCoCo、PhraseCut、EraseDraw、MagicBrush、SmartEdit；Stage 2：ROVI、VPLM、GroundMoRe、RVoS、MeViS、InstructV2V、自建 VEG-Edit；Table 1 列约 3.438M image instructions、136.1K video instructions","existing_method":"","training_signal":"两阶段 image-to-video curriculum，端到端训练","hardware_bin":"medium","max_reported_ft_gpus":32.0,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"d37dba097ea3ceae7fdd8651600c5c20907ec0f58452c45f8dd6c12d1dc682e0","reasoning_id":"ICCV-2025-1565","reasoning_role":"R1","reasoning_topic":"Editing / 修改已有视频"},{"id":"R0762","title":"Epona: Autoregressive Diffusion World Model for Autonomous Driving","year":2025,"venue":"ICCV","arxiv_id":"","aliases":["reasoning:ICCV-2025-1714"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Zhang_Epona_Autoregressive_Diffusion_World_Model_for_Autonomous_Driving_ICCV_2025_paper.pdf","status":"已核对资源阶段","resource_note":"","stages":[{"role":"pre","device":"A100","count_min":48.0,"count_max":48.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":336.0,"wall_hours_max":336.0,"time_bound":"approximate","gpu_hours_min":16128.0,"gpu_hours_max":16128.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"Epona端到端from-scratch，600K iterations，batch96，512×1024，nearly两周；不算fine-tuning。","comparable":false,"is_gpu":true,"role_label":"预训练／从零训练","source_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Zhang_Epona_Autoregressive_Diffusion_World_Model_for_Autonomous_Driving_ICCV_2025_paper.pdf","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Epona 2.5B：1.3B spatiotemporal Transformer + 1.2B next-frame diffusion + 50M trajectory diffusion；temporal-aware DCAE","training_datasets":"nuPlan 公开视频 + nuScenes 700 scenes；world model 从头训练，DCAE 基于预训练组件适配","existing_method":"","training_signal":"world model from scratch + autoencoder 适配","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"b9c4acbfd43cca8beae0e81f32e2c799583bf06eb6f33cf77719c4a0f6aba2d8","reasoning_id":"ICCV-2025-1714","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0763","title":"IRASim: A Fine-Grained World Model for Robot Manipulation","year":2025,"venue":"ICCV","arxiv_id":"","aliases":["reasoning:ICCV-2025-1897"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Zhu_IRASim_A_Fine-Grained_World_Model_for_Robot_Manipulation_ICCV_2025_paper.pdf","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"IRASim diffusion Transformer + frame-level action conditioning；小数据 policy 实验初始化 OpenSora","training_datasets":"RT-1、Bridge、Language-Table、RoboNet 的训练划分；LIBERO / PushT 实验用 expert demos + policy post-trained rollouts（含成功 / 失败）","existing_method":"","training_signal":"world model / value training；policy rollout 数据增强","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"1759ee286432f273ad94b7227d522c2731dd5e3cfb548259b6dbb3626b5d63ea","reasoning_id":"ICCV-2025-1897","reasoning_role":"R2","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0764","title":"Modular-Cam: Modular Dynamic Camera-view Video Generation with LLM","year":2025,"venue":"AAAI","arxiv_id":"2504.12048","aliases":["reasoning:AAAI-2025-1293"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2504.12048","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":null,"count_max":null,"card_memory_gb":40.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"Modular-Cam的CamOperator LoRA rank4 / AdaControlNet；原文只写A10040G GPU，数量未明确，不凭单数措辞定为单卡。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2504.12048","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"SD1.5 + temporal Transformer；CamOperator LoRA 与 AdaControlNet","training_datasets":"WebVid-10M 训练 temporal layers；约 50 条按运动模式增强的 videos 训练 CamOperator；WebVid-10M 100K subset 训练 AdaControlNet","existing_method":"","training_signal":"分模块训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"a0afbf69b8c7ff466d6a41458d3c2f6729caa0b1f86c9036addd1c2c30cf559d","reasoning_id":"AAAI-2025-1293","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0765","title":"DriveDreamer-2: LLM-Enhanced World Models for Diverse Driving Video Generation","year":2025,"venue":"AAAI","arxiv_id":"2403.06845","aliases":["reasoning:AAAI-2025-2219"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2403.06845","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A800","count_min":null,"count_max":null,"card_memory_gb":80.0,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":null,"gpu_hours_max":null,"gpu_hours_basis":"","scope":"DriveDreamer-2全参数视频适配，200K iterations，8frames×6views；卡数和耗时未定位。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2403.06845","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Stable Video Diffusion + UniMVM；SD2.1 ControlNet 生成 HDMap","training_datasets":"nuScenes 700 training scenes，约 1M frames；构造 text-to-script 数据微调 GPT-3.5；HDMap 与 video generator 分别训练","existing_method":"","training_signal":"LLM / HDMap / video 多模块训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":true,"has_ft_time":false,"fulltext_sha256":"2d8bc1561e783185aff48a6abfd3f0e5969d6431dfe440823ba4409537674cd0","reasoning_id":"AAAI-2025-2219","reasoning_role":"R1","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0766","title":"DreamRunner: Fine-Grained Compositional Story-to-Video Generation with Retrieval-Augmented Motion Adaptation","year":2026,"venue":"AAAI","arxiv_id":"2411.16657","aliases":["reasoning:AAAI-2026-3029"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2411.16657","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A6000","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":0.0833333,"wall_hours_max":0.0833333,"time_bound":"approximate","gpu_hours_min":0.0833,"gpu_hours_max":0.0833,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"DreamRunner正文：每个motion prior test-time fine-tuning约5min；附录另报每motion0.2 GPU-hours，口径不一致，保留两种原文数值，不平均。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2411.16657","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"GPU（型号未定位）","count_min":null,"count_max":null,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":null,"wall_hours_max":null,"time_bound":"approximate","gpu_hours_min":0.2,"gpu_hours_max":0.2,"gpu_hours_basis":"原文报告 device-hours / device-days；days × 24 换算","scope":"附录B.3：每motion0.2 GPU-hours；15–20motions的完整故事约3–4 GPU-hours。与正文单A6000 5min/prior口径不同。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2411.16657","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"CogVideoX-2B（story）；另用 CogVideoX-5B 做 compositional T2V","training_datasets":"test-time adaptation：top-ranked retrieved motion videos 训练 temporal / spatial LoRA；参考图像训练 subject prior；检索视频库的具体 dataset 名称未核实","existing_method":"","training_signal":"test-time LoRA fine-tuning","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"96897bd7efcdc75c13f5494468247b281841898e24a481a692c67a1be51888ed","reasoning_id":"AAAI-2026-3029","reasoning_role":"R1","reasoning_topic":"长视频 / 多镜头 / 记忆"},{"id":"R0767","title":"ManipDreamer3D: Synthesizing Plausible Robotic Manipulation Video with Occupancy-aware 3D Trajectory","year":2026,"venue":"AAAI","arxiv_id":"2509.05314","aliases":["reasoning:AAAI-2026-3148"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2509.05314","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"两个实现：Stable Video Diffusion；CogVideoX-5B","training_datasets":"Bridge V1 + Bridge V2，经 curation 得 8.7K valid episodes，按 9:1 train / test 划分；辅助 detector 的独立训练数据未核实","existing_method":"","training_signal":"trajectory-conditioned generator 训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"580306cc1bf396b56d58ecab614af21bc195aeedf897d50955415accefadd83b","reasoning_id":"AAAI-2026-3148","reasoning_role":"R1","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0768","title":"FantasyHSI: Video-Generation-Centric 4D Human Synthesis in Any Scene Through a Graph-Based Multi-Agent Framework","year":2026,"venue":"AAAI","arxiv_id":"2509.01232","aliases":["reasoning:AAAI-2026-3449"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2509.01232","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"A100","count_min":8.0,"count_max":8.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":20.0,"wall_hours_max":20.0,"time_bound":"approximate","gpu_hours_min":160.0,"gpu_hours_max":160.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"FantasyHSI：Wan2.1-I2V-14B DPO，10K preference pairs，约20h；多agent推理与4D优化另计。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2509.01232","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.1-I2V-14B","training_datasets":"自建 10,000 preference pairs 做 DPO；候选来自 Veo、Hunyuan-Video、Runway、Kling，由专业标注者按指令 / 动作 / 形体 / 场景一致性标偏好","existing_method":"","training_signal":"generator DPO；agent 推理","hardware_bin":"small","max_reported_ft_gpus":8.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"c4e0ca6b4ffb695f3046a1fbf0f528f7aa7f15b30092f74598b296a6e1a44ef6","reasoning_id":"AAAI-2026-3449","reasoning_role":"R1","reasoning_topic":"人物动画 / Talking Head / HOI"},{"id":"R0769","title":"GENMAC: Compositional Text-to-Video Generation with Multi-Agent Collaboration","year":2026,"venue":"AAAI","arxiv_id":"2412.04440","aliases":["reasoning:AAAI-2026-3563"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2412.04440","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"VideoCrafter2","training_datasets":"无新增模型训练；现成生成器上用 LVD-style attention guidance","existing_method":"","training_signal":"Training-free multi-agent inference","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"ead5f0460f4455a037dddb40ce09f0fdea2955c18b5b1f29ef872e91d2c32636","reasoning_id":"AAAI-2026-3563","reasoning_role":"R1","reasoning_topic":"条件控制 / 定制 / 几何一致"},{"id":"R0770","title":"SpA2V: Harnessing Spatial Auditory Cues for Audio-driven Spatially-aware Video Generation","year":2025,"venue":"MM","arxiv_id":"2508.00782","aliases":["reasoning:MM-2025-0208"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2508.00782","status":"无需本文权重训练","resource_note":"SpA2V明确training-free的空间音频到视频pipeline。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Stable Diffusion + pretrained AnimateDiff temporal modules + MIGC grounding modules","training_datasets":"无新增权重训练，组合已有模块；in-context candidate examples 用于推理检索","existing_method":"","training_signal":"Training-free + in-context retrieval","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"36008f84a7fbcfcf6eae22009dea70745d55a5c53ebf574b4826b0a414337b40","reasoning_id":"MM-2025-0208","reasoning_role":"R1","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0771","title":"MAViD: A Multimodal Framework for Audio-Visual Dialogue Understanding and Generation","year":2026,"venue":"SIGGRAPH","arxiv_id":"2512.03034","aliases":["reasoning:SIGGRAPH-2026-0075"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2512.03034","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Creator：autoregressive audio model + Wan DiT / VAE；higgs-audio tokenizer；Creator 合计约 8B","training_datasets":"Conductor：混合 multimodal QA / human-centered instruction 数据，精确清单未核实；Creator audio stage：自采超过 10K 小时网络音频；joint stage：近 1M audio-video pairs（影片、talking body、演讲、互动）","existing_method":"","training_signal":"三阶段 Conductor / audio AR / joint AV 训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"29fc1af327d49e4e37bd5e590090add0fcc7e010bf2511964ef9d49d46c08c1b","reasoning_id":"SIGGRAPH-2026-0075","reasoning_role":"R1","reasoning_topic":"通用生成框架 / 联合音视频"},{"id":"R0772","title":"Anim-Director: A Large Multimodal Model Powered Agent for Controllable Animation Video Generation","year":2024,"venue":"SIGGRAPH-Asia","arxiv_id":"2408.09787","aliases":["reasoning:SIGGRAPH-Asia-2024-0116"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2408.09787","status":"无需本文权重训练","resource_note":"Anim-Director明确training-free；LMM与外部生成工具调用仍有成本。","stages":[],"no_weight_update":true,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":false,"model":"Midjourney 图像工具；Pika 视频工具；API 具体 snapshot 未核实","training_datasets":"无新增权重训练；通过现成 API 生成、评估、选择与修改","existing_method":"","training_signal":"Training-free agent orchestration","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"e1cfdccd2c575770e99e4538d0d902c42bfe7979a826864795dfd21cbe43b0fb","reasoning_id":"SIGGRAPH-Asia-2024-0116","reasoning_role":"R1","reasoning_topic":"人物动画 / Talking Head / HOI"},{"id":"R0773","title":"AniMaker: Multi-Agent Animated Storytelling with MCTS-Driven Clip Generation","year":2025,"venue":"SIGGRAPH-Asia","arxiv_id":"2506.10540","aliases":["reasoning:SIGGRAPH-Asia-2025-0111"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2506.10540","status":"未定位可比微调配置","resource_note":"已检索缓存正文中的训练、硬件、显存、耗时与步骤相关段落；未定位可归属于本文 fine-tuning 的明确配置。此项不等于零成本，也不等于确认全文从未报告。","stages":[],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan2.1 clip generator；FLUX.1-dev 背景与 Hunyuan3D 角色 assets","training_datasets":"生成 workflow 使用现成模型；AniEval 的 face-consistency evaluator 另在 Anime Face Dataset 上训练 InceptionNeXt","existing_method":"","training_signal":"生成器推理；evaluator 有训练","hardware_bin":"unknown","max_reported_ft_gpus":null,"has_comparable_ft":false,"has_ft_time":false,"fulltext_sha256":"3dd0ffa5fde85269f9fa96bb8df213a5f6b78f0951256b5d7fed63d03f1159c4","reasoning_id":"SIGGRAPH-Asia-2025-0111","reasoning_role":"R1","reasoning_topic":"动作预测 / 交互 World Models"},{"id":"R0774","title":"FairyGen: Storied Cartoon Video from a Single Child-Drawn Character","year":2025,"venue":"SIGGRAPH-Asia","arxiv_id":"2506.21272","aliases":["reasoning:SIGGRAPH-Asia-2025-0132"],"technical_key":"","catalog_id":"","library_id":"","topic_code":"","topic":"","task":"未归入视频任务轴","directions":[],"source_url":"https://arxiv.org/pdf/2506.21272","status":"已核对资源阶段","resource_note":"","stages":[{"role":"ft","device":"L20","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":2.0,"wall_hours_max":2.0,"time_bound":"approximate","gpu_hours_min":2.0,"gpu_hours_max":2.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"FairyGen单实例风格学习约120min；不是整个故事生成时间。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2506.21272","locator":"正文／附录的 Training / Implementation / Compute 设置"},{"role":"ft","device":"L20","count_min":1.0,"count_max":1.0,"card_memory_gb":null,"peak_memory_gb":null,"wall_hours_min":3.0,"wall_hours_max":3.0,"time_bound":"approximate","gpu_hours_min":3.0,"gpu_hours_max":3.0,"gpu_hours_basis":"[推算] 同一阶段设备数 × 墙钟小时；不是价格或归一化 FLOPs","scope":"FairyGen motion customization约180min；与风格阶段分列。","comparable":true,"is_gpu":true,"role_label":"微调／适配训练","source_url":"https://arxiv.org/pdf/2506.21272","locator":"正文／附录的 Training / Implementation / Compute 设置"}],"no_weight_update":false,"has_fulltext":true,"screening":"全文关键词检索与上下文筛查；有数字的资源条目再核对训练阶段。未定位项不进入量化比较。","hardware_candidate":true,"model":"Wan-family MMDiT image-to-video + motion customization adapter；background inpainting diffusion，精确 checkpoint 未核实","training_datasets":"以输入角色画学习 style adapter；由该角色 3D proxy 渲染的 motions 训练 identity / motion adapters（先打乱帧，后时序训练）；无需额外配对视频集","existing_method":"","training_signal":"per-character style / motion adapter 训练","hardware_bin":"single","max_reported_ft_gpus":1.0,"has_comparable_ft":true,"has_ft_time":true,"fulltext_sha256":"57f5fd3cb16dfb26e48aff1ee67e2f3ff3d5adf0fc1493a11a3afc17281df7d1","reasoning_id":"SIGGRAPH-Asia-2025-0132","reasoning_role":"R1","reasoning_topic":"通用生成框架 / 联合音视频"}],"alias_index":{"pool:P001":"R0001","pool:P002":"R0002","pool:P003":"R0003","pool:P004":"R0004","pool:P005":"R0005","pool:P006":"R0006","pool:P007":"R0007","pool:P008":"R0008","pool:P009":"R0009","pool:P010":"R0010","pool:P011":"R0011","pool:P012":"R0012","reasoning:ICCV-2025-1614":"R0012","pool:P013":"R0013","pool:P014":"R0014","pool:P015":"R0015","pool:P016":"R0016","pool:P017":"R0017","pool:P018":"R0018","pool:P019":"R0019","library:P166":"R0019","pool:P020":"R0020","pool:P021":"R0021","pool:P022":"R0022","pool:P023":"R0023","pool:P024":"R0024","pool:P025":"R0025","pool:P026":"R0026","external:XR-REF-VIDEOREPA":"R0026","pool:P027":"R0027","pool:P028":"R0028","pool:P029":"R0029","pool:P030":"R0030","pool:P031":"R0031","pool:P032":"R0032","pool:P033":"R0033","pool:P034":"R0034","pool:P035":"R0035","pool:P036":"R0036","pool:P037":"R0037","pool:P038":"R0038","pool:P039":"R0039","pool:P040":"R0040","library:P123":"R0040","pool:P041":"R0041","pool:P042":"R0042","pool:P043":"R0043","reasoning:ICML-2025-0008":"R0043","pool:P044":"R0044","library:P135":"R0044","lineage:WORLD":"R0044","pool:P045":"R0045","pool:P046":"R0046","pool:P047":"R0047","pool:P048":"R0048","pool:P049":"R0049","pool:P050":"R0050","pool:P051":"R0051","pool:P052":"R0052","pool:P053":"R0053","pool:P054":"R0054","pool:P055":"R0055","pool:P056":"R0056","pool:P057":"R0057","pool:P058":"R0058","pool:P059":"R0059","reasoning:CVPR-2026-0435":"R0059","pool:P060":"R0060","pool:P061":"R0061","pool:P062":"R0062","pool:P063":"R0063","library:P287":"R0063","pool:P064":"R0064","pool:P065":"R0065","pool:P066":"R0066","pool:P067":"R0067","pool:P068":"R0068","pool:P069":"R0069","pool:P070":"R0070","pool:P071":"R0071","pool:P072":"R0072","pool:P073":"R0073","reasoning:NeurIPS-2026-5021":"R0073","pool:P074":"R0074","lineage:PERMA":"R0074","recent:PERMA":"R0074","pool:P075":"R0075","pool:P076":"R0076","pool:P077":"R0077","reasoning:NeurIPS-2026-1384":"R0077","pool:P078":"R0078","pool:P079":"R0079","pool:P080":"R0080","pool:P081":"R0081","pool:P082":"R0082","pool:P083":"R0083","pool:P084":"R0084","pool:P085":"R0085","pool:P086":"R0086","pool:P087":"R0087","pool:P088":"R0088","pool:P089":"R0089","pool:P090":"R0090","pool:P091":"R0091","lineage:SPIRAL":"R0091","reasoning:NeurIPS-2026-3987":"R0091","pool:P092":"R0092","pool:P093":"R0093","pool:P094":"R0094","pool:P095":"R0095","pool:P096":"R0096","pool:P097":"R0097","pool:P098":"R0098","pool:P099":"R0099","pool:P100":"R0100","pool:P101":"R0101","pool:P102":"R0102","pool:P103":"R0103","pool:P104":"R0104","pool:P105":"R0105","pool:P106":"R0106","pool:P107":"R0107","reasoning:CVPR-2026-3739":"R0107","pool:P108":"R0108","pool:P109":"R0109","pool:P110":"R0110","pool:P111":"R0111","pool:P112":"R0112","pool:P113":"R0113","pool:P114":"R0114","pool:P115":"R0115","pool:P116":"R0116","pool:P117":"R0117","pool:P118":"R0118","pool:P119":"R0119","pool:P120":"R0120","pool:P121":"R0121","pool:P122":"R0122","pool:P123":"R0123","pool:P124":"R0124","pool:P125":"R0125","lineage:COG":"R0125","pool:P126":"R0126","pool:P127":"R0127","pool:P128":"R0128","pool:P129":"R0129","pool:P130":"R0130","pool:P131":"R0131","pool:P132":"R0132","pool:P133":"R0133","pool:P134":"R0134","pool:P135":"R0135","pool:P136":"R0136","pool:P137":"R0137","pool:P138":"R0138","pool:P139":"R0139","pool:P140":"R0140","pool:P141":"R0141","pool:P142":"R0142","pool:P143":"R0143","pool:P144":"R0144","pool:P145":"R0145","pool:P146":"R0146","pool:P147":"R0147","pool:P148":"R0148","pool:P149":"R0149","pool:P150":"R0150","pool:P151":"R0151","pool:P152":"R0152","pool:P153":"R0153","pool:P154":"R0154","pool:P155":"R0155","pool:P156":"R0156","pool:P157":"R0157","pool:P158":"R0158","pool:P159":"R0159","pool:P160":"R0160","pool:P161":"R0161","pool:P162":"R0162","pool:P163":"R0163","pool:P164":"R0164","pool:P165":"R0165","pool:P166":"R0166","pool:P167":"R0167","pool:P168":"R0168","library:VG10":"R0168","pool:P169":"R0169","library:P081":"R0169","pool:P170":"R0170","pool:P171":"R0171","pool:P172":"R0172","pool:P173":"R0173","pool:P174":"R0174","pool:P175":"R0175","pool:P176":"R0176","pool:P177":"R0177","pool:P178":"R0178","pool:P179":"R0179","pool:P180":"R0180","pool:P181":"R0181","pool:P182":"R0182","pool:P183":"R0183","pool:P184":"R0184","pool:P185":"R0185","pool:P186":"R0186","pool:P187":"R0187","pool:P188":"R0188","pool:P189":"R0189","pool:P190":"R0190","pool:P191":"R0191","pool:P192":"R0192","pool:P193":"R0193","pool:P194":"R0194","pool:P195":"R0195","library:P012":"R0196","library:P021":"R0197","lineage:BLACK":"R0197","library:P022":"R0198","library:P024":"R0199","library:P033":"R0200","library:P036":"R0201","library:P059":"R0202","library:P061":"R0203","library:P143":"R0204","lineage:IVR":"R0204","library:P146":"R0205","library:P149":"R0206","lineage:ACT":"R0206","library:P157":"R0207","library:P071":"R0208","lineage:CWMDT":"R0208","library:P159":"R0209","lineage:GPS":"R0209","recent:GPS":"R0209","library:P097":"R0210","lineage:MEMFAULT":"R0210","library:P161":"R0211","library:P251":"R0212","library:P342":"R0213","recent:SPACE":"R0213","library:P372":"R0214","library:P214":"R0215","library:P259":"R0216","library:P379":"R0217","library:P390":"R0218","library:P391":"R0219","library:P112":"R0220","library:P117":"R0221","library:P125":"R0222","library:P133":"R0223","lineage:ONE":"R0223","recent:ONE":"R0223","library:VG17":"R0224","lineage:QWEN":"R0224","recent:QWEN":"R0224","library:VG31":"R0225","lineage:VANS":"R0225","reasoning:CVPR-2026-3642":"R0225","library:VG13":"R0226","lineage:VOID":"R0226","recent:VOID":"R0226","reasoning:ECCV-2026-0955":"R0226","library:VG46":"R0227","lineage:GIVE":"R0227","recent:GIVE":"R0227","library:VG16":"R0228","library:VG30":"R0229","library:VG63":"R0230","library:VG15":"R0231","lineage:A2RD":"R0231","library:VG60":"R0232","lineage:S3":"R0232","library:VG43":"R0233","library:VG37":"R0234","library:VG50":"R0235","library:VG02":"R0236","lineage:VOY":"R0236","library:VG04":"R0237","lineage:STATE":"R0237","library:VG09":"R0238","library:VG11":"R0239","library:VG19":"R0240","lineage:CINE":"R0240","recent:CINE":"R0240","library:VG20":"R0241","lineage:VGEN":"R0241","recent:VGEN":"R0241","library:VG22":"R0242","library:VG58":"R0243","library:VG61":"R0244","library:VG68":"R0245","lineage:PROC":"R0246","lineage:SCHEMA":"R0247","external:XR-REF-SCHEMA":"R0247","lineage:DSCM":"R0248","external:XR-REF-DSCM":"R0248","lineage:OMNI":"R0249","lineage:PHYS":"R0250","external:EXT-PHYS":"R0250","lineage:KNOW":"R0251","lineage:GENEX":"R0252","external:EXT-GENEX":"R0252","lineage:STREAM":"R0253","lineage:WEAVE":"R0254","lineage:VIDEOAG":"R0255","lineage:DIRECTOR":"R0256","lineage:VIS":"R0257","recent:VIS":"R0257","external:EXT-VIS":"R0257","lineage:GRID":"R0258","lineage:DRAM":"R0259","lineage:SUSP":"R0260","recent:SUSP":"R0260","lineage:CONTROL":"R0261","lineage:SKETCH":"R0262","recent:CLARIFIER":"R0263","recent:CLARITI":"R0264","recent:POINTER":"R0265","recent:IGCLAR":"R0266","recent:SC2":"R0267","recent:SPATIA":"R0268","recent:COME":"R0269","recent:CAUSAL":"R0270","recent:MULTI":"R0271","recent:ARIFY":"R0272","recent:COACH":"R0273","recent:ACE":"R0274","recent:MCE":"R0275","recent:DARWIN":"R0276","recent:SILVR":"R0277","recent:JARVIS":"R0278","recent:VISTA":"R0279","reasoning:CVPR-2026-3841":"R0279","recent:DGM":"R0280","recent:TUTOR":"R0281","external:EXT-TWIN":"R0282","external:EXT-LAYER":"R0283","external:EXT-VIMAX":"R0284","external:XR-REF-SARA":"R0285","external:XR-REF-CITRIS":"R0286","external:XR-REF-ABSTRACTION":"R0287","library:P001":"R0288","library:P002":"R0289","library:P004":"R0290","library:P005":"R0291","library:P006":"R0292","library:P007":"R0293","library:P008":"R0294","library:P009":"R0295","library:P010":"R0296","library:P011":"R0297","library:P013":"R0298","library:P015":"R0299","library:P016":"R0300","library:P017":"R0301","library:P018":"R0302","library:P019":"R0303","library:P020":"R0304","library:P023":"R0305","library:P027":"R0306","library:P028":"R0307","library:P029":"R0308","library:P030":"R0309","library:P031":"R0310","library:P032":"R0311","library:P034":"R0312","library:P035":"R0313","library:P037":"R0314","library:P038":"R0315","library:P039":"R0316","library:P040":"R0317","library:P041":"R0318","library:P042":"R0319","library:P043":"R0320","library:P044":"R0321","library:P046":"R0322","library:P047":"R0323","library:P048":"R0324","library:P049":"R0325","library:P050":"R0326","library:P051":"R0327","library:P052":"R0328","library:P053":"R0329","library:P054":"R0330","library:P055":"R0331","library:P056":"R0332","library:P057":"R0333","library:P058":"R0334","library:P060":"R0335","library:P062":"R0336","library:P067":"R0337","library:P079":"R0338","library:P094":"R0339","library:P063":"R0340","library:P141":"R0341","library:P142":"R0342","library:P064":"R0343","library:P065":"R0344","library:P144":"R0345","library:P066":"R0346","library:P145":"R0347","library:P162":"R0348","library:P163":"R0349","library:P147":"R0350","library:P148":"R0351","library:P150":"R0352","library:P151":"R0353","reasoning:ICLR-2025-2989":"R0353","library:P152":"R0354","library:P153":"R0355","library:P164":"R0356","library:P068":"R0357","library:P069":"R0358","library:P070":"R0359","library:P072":"R0360","library:P154":"R0361","library:P073":"R0362","library:P074":"R0363","library:P075":"R0364","library:P076":"R0365","library:P077":"R0366","library:P155":"R0367","library:P156":"R0368","library:P078":"R0369","library:P080":"R0370","library:P158":"R0371","library:P083":"R0372","library:P084":"R0373","library:P085":"R0374","library:P086":"R0375","library:P087":"R0376","library:P088":"R0377","library:P089":"R0378","library:P090":"R0379","library:P091":"R0380","library:P092":"R0381","library:P093":"R0382","library:P160":"R0383","library:P095":"R0384","library:P096":"R0385","library:P098":"R0386","library:P099":"R0387","library:P100":"R0388","library:P101":"R0389","library:P103":"R0390","library:P165":"R0391","library:P168":"R0392","library:P175":"R0393","library:P176":"R0394","library:P177":"R0395","library:P179":"R0396","library:P182":"R0397","library:P184":"R0398","library:P185":"R0399","library:P186":"R0400","library:P188":"R0401","library:P189":"R0402","library:P192":"R0403","library:P193":"R0404","library:P201":"R0405","library:P202":"R0406","library:P205":"R0407","library:P210":"R0408","library:P215":"R0409","library:P218":"R0410","library:P220":"R0411","library:P221":"R0412","library:P227":"R0413","library:P229":"R0414","library:P238":"R0415","library:P244":"R0416","library:P249":"R0417","library:P250":"R0418","library:P257":"R0419","library:P263":"R0420","library:P266":"R0421","library:P267":"R0422","library:P272":"R0423","library:P280":"R0424","library:P285":"R0425","library:P289":"R0426","library:P292":"R0427","library:P302":"R0428","library:P308":"R0429","library:P309":"R0430","library:P316":"R0431","library:P328":"R0432","library:P330":"R0433","library:P333":"R0434","library:P347":"R0435","library:P354":"R0436","library:P356":"R0437","library:P357":"R0438","library:P360":"R0439","library:P362":"R0440","library:P376":"R0441","library:P389":"R0442","library:P394":"R0443","library:P396":"R0444","reasoning:NeurIPS-2026-4979":"R0444","library:P398":"R0445","library:P400":"R0446","library:P169":"R0447","library:P171":"R0448","library:P172":"R0449","library:P174":"R0450","library:P191":"R0451","library:P196":"R0452","library:P199":"R0453","library:P203":"R0454","library:P204":"R0455","library:P213":"R0456","library:P222":"R0457","library:P225":"R0458","library:P226":"R0459","library:P230":"R0460","library:P239":"R0461","library:P245":"R0462","library:P246":"R0463","library:P253":"R0464","library:P261":"R0465","library:P268":"R0466","library:P283":"R0467","library:P284":"R0468","library:P288":"R0469","library:P291":"R0470","library:P325":"R0471","library:P344":"R0472","library:P345":"R0473","library:P383":"R0474","library:P385":"R0475","library:P393":"R0476","library:P200":"R0477","library:P212":"R0478","library:P232":"R0479","library:P233":"R0480","library:P236":"R0481","library:P248":"R0482","library:P254":"R0483","library:P275":"R0484","library:P277":"R0485","library:P282":"R0486","library:P293":"R0487","library:P298":"R0488","library:P312":"R0489","library:P313":"R0490","library:P318":"R0491","library:P324":"R0492","library:P336":"R0493","library:P340":"R0494","library:P346":"R0495","library:P363":"R0496","library:P392":"R0497","library:P170":"R0498","library:P173":"R0499","library:P181":"R0500","library:P187":"R0501","library:P402":"R0502","library:P190":"R0503","library:P195":"R0504","library:P403":"R0505","library:P197":"R0506","library:P404":"R0507","library:P198":"R0508","library:P206":"R0509","library:P207":"R0510","library:P209":"R0511","library:P211":"R0512","library:P216":"R0513","library:P219":"R0514","library:P234":"R0515","library:P235":"R0516","library:P240":"R0517","library:P242":"R0518","library:P243":"R0519","library:P247":"R0520","library:P252":"R0521","library:P255":"R0522","library:P258":"R0523","library:P260":"R0524","library:P269":"R0525","library:P270":"R0526","library:P271":"R0527","library:P274":"R0528","library:P278":"R0529","library:P279":"R0530","library:P281":"R0531","library:P286":"R0532","library:P290":"R0533","library:P296":"R0534","library:P297":"R0535","library:P300":"R0536","library:P301":"R0537","library:P305":"R0538","library:P306":"R0539","library:P314":"R0540","library:P317":"R0541","library:P319":"R0542","library:P321":"R0543","library:P327":"R0544","library:P331":"R0545","library:P335":"R0546","library:P341":"R0547","library:P348":"R0548","library:P355":"R0549","library:P359":"R0550","library:P361":"R0551","library:P365":"R0552","library:P367":"R0553","library:P368":"R0554","library:P369":"R0555","library:P370":"R0556","library:P371":"R0557","library:P373":"R0558","library:P377":"R0559","library:P380":"R0560","library:P388":"R0561","library:P401":"R0562","library:P167":"R0563","library:P178":"R0564","library:P180":"R0565","library:P183":"R0566","library:P194":"R0567","library:P208":"R0568","library:P217":"R0569","library:P223":"R0570","library:P224":"R0571","library:P228":"R0572","library:P231":"R0573","library:P237":"R0574","library:P241":"R0575","library:P256":"R0576","library:P262":"R0577","reasoning:CVPR-2025-2277":"R0577","library:P264":"R0578","library:P265":"R0579","library:P273":"R0580","library:P276":"R0581","library:P294":"R0582","library:P299":"R0583","library:P303":"R0584","library:P304":"R0585","library:P310":"R0586","library:P311":"R0587","library:P320":"R0588","library:P322":"R0589","library:P323":"R0590","library:P326":"R0591","library:P329":"R0592","library:P332":"R0593","library:P334":"R0594","reasoning:ICML-2026-1668":"R0594","library:P337":"R0595","library:P338":"R0596","library:P339":"R0597","library:P343":"R0598","library:P350":"R0599","library:P351":"R0600","library:P352":"R0601","library:P353":"R0602","library:P358":"R0603","library:P366":"R0604","library:P374":"R0605","library:P375":"R0606","library:P381":"R0607","library:P382":"R0608","library:P386":"R0609","library:P395":"R0610","library:P397":"R0611","library:P399":"R0612","library:P104":"R0613","library:P105":"R0614","library:P106":"R0615","library:P107":"R0616","library:P108":"R0617","library:P110":"R0618","library:P111":"R0619","library:P113":"R0620","library:P114":"R0621","library:P115":"R0622","library:P116":"R0623","library:P118":"R0624","library:P119":"R0625","library:P120":"R0626","library:P121":"R0627","library:P122":"R0628","library:P124":"R0629","reasoning:NeurIPS-2024-2391":"R0629","library:P126":"R0630","library:P127":"R0631","library:P128":"R0632","library:P129":"R0633","library:P131":"R0634","library:P132":"R0635","library:P136":"R0636","library:P137":"R0637","reasoning:ICCV-2025-1556":"R0637","library:P138":"R0638","library:P139":"R0639","library:VG18":"R0640","library:VG01":"R0641","library:VG14":"R0642","library:VG03":"R0643","library:VG05":"R0644","library:VG06":"R0645","library:VG07":"R0646","reasoning:CVPR-2026-2789":"R0646","library:VG08":"R0647","library:VG12":"R0648","library:VG21":"R0649","library:VG23":"R0650","library:VG24":"R0651","library:VG25":"R0652","library:VG26":"R0653","library:VG27":"R0654","library:VG28":"R0655","library:VG29":"R0656","library:VG32":"R0657","library:VG33":"R0658","library:VG34":"R0659","library:VG35":"R0660","library:VG36":"R0661","library:VG38":"R0662","library:VG39":"R0663","library:VG40":"R0664","library:VG41":"R0665","library:VG42":"R0666","library:VG44":"R0667","library:VG45":"R0668","library:VG47":"R0669","library:VG48":"R0670","library:VG49":"R0671","library:VG51":"R0672","library:VG52":"R0673","library:VG53":"R0674","library:VG54":"R0675","library:VG55":"R0676","library:VG56":"R0677","library:VG57":"R0678","library:VG59":"R0679","library:VG62":"R0680","library:VG65":"R0681","library:VG67":"R0682","reasoning:CVPR-2026-1584":"R0683","reasoning:CVPR-2026-1722":"R0684","reasoning:CVPR-2026-2024":"R0685","reasoning:CVPR-2026-2326":"R0686","reasoning:CVPR-2026-2560":"R0687","reasoning:CVPR-2026-2680":"R0688","reasoning:CVPR-2026-3054":"R0689","reasoning:CVPR-2026-3115":"R0690","reasoning:CVPR-2026-3787":"R0691","reasoning:CVPR-2026-3919":"R0692","reasoning:CVPR-2026-4022":"R0693","reasoning:ICLR-2026-1627":"R0694","reasoning:ICLR-2026-2063":"R0695","reasoning:ICLR-2026-2954":"R0696","reasoning:ICLR-2026-3626":"R0697","reasoning:ICLR-2026-4101":"R0698","reasoning:ICLR-2026-4273":"R0699","reasoning:ICLR-2026-4471":"R0700","reasoning:ICLR-2026-4937":"R0701","reasoning:NeurIPS-2026-0460":"R0702","reasoning:NeurIPS-2026-0902":"R0703","reasoning:NeurIPS-2026-2519":"R0704","reasoning:NeurIPS-2026-2747":"R0705","reasoning:NeurIPS-2026-3007":"R0706","reasoning:NeurIPS-2026-3426":"R0707","reasoning:NeurIPS-2026-3727":"R0708","reasoning:NeurIPS-2026-4117":"R0709","reasoning:NeurIPS-2026-4145":"R0710","reasoning:NeurIPS-2026-4719":"R0711","reasoning:ECCV-2026-0039":"R0712","reasoning:ECCV-2026-0106":"R0713","reasoning:ECCV-2026-0260":"R0714","reasoning:ECCV-2026-0267":"R0715","reasoning:ECCV-2026-0278":"R0716","reasoning:ECCV-2026-0321":"R0717","reasoning:ECCV-2026-0382":"R0718","reasoning:ECCV-2026-0574":"R0719","reasoning:ECCV-2026-0677":"R0720","reasoning:ECCV-2026-1141":"R0721","reasoning:ECCV-2026-1189":"R0722","reasoning:ECCV-2026-1203":"R0723","reasoning:ECCV-2026-1501":"R0724","reasoning:ECCV-2026-1512":"R0725","reasoning:ECCV-2026-1597":"R0726","reasoning:ECCV-2026-1602":"R0727","reasoning:ECCV-2026-1910":"R0728","reasoning:ECCV-2026-1966":"R0729","reasoning:ECCV-2026-2217":"R0730","reasoning:ICML-2026-1741":"R0731","reasoning:ICML-2026-1784":"R0732","reasoning:ICML-2026-1911":"R0733","reasoning:ICML-2026-3051":"R0734","reasoning:ICML-2026-4190":"R0735","reasoning:ICML-2026-5339":"R0736","reasoning:ICML-2026-6035":"R0737","reasoning:ECCV-2024-1178":"R0738","reasoning:ICLR-2024-1470":"R0739","reasoning:ICLR-2024-1627":"R0740","reasoning:ICLR-2025-2467":"R0741","reasoning:ICLR-2025-3077":"R0742","reasoning:ICLR-2025-3138":"R0743","reasoning:ICML-2024-1859":"R0744","reasoning:ICML-2024-2015":"R0745","reasoning:ICML-2025-0897":"R0746","reasoning:NeurIPS-2024-0388":"R0747","reasoning:NeurIPS-2024-1367":"R0748","reasoning:NeurIPS-2025-2377":"R0749","reasoning:NeurIPS-2025-2664":"R0750","reasoning:NeurIPS-2025-4954":"R0751","reasoning:NeurIPS-2025-5490":"R0752","reasoning:CVPR-2024-0659":"R0753","reasoning:CVPR-2024-1231":"R0754","reasoning:CVPR-2025-0149":"R0755","reasoning:CVPR-2025-1103":"R0756","reasoning:ICCV-2025-0423":"R0757","reasoning:ICCV-2025-0426":"R0758","reasoning:ICCV-2025-1440":"R0759","reasoning:ICCV-2025-1517":"R0760","reasoning:ICCV-2025-1565":"R0761","reasoning:ICCV-2025-1714":"R0762","reasoning:ICCV-2025-1897":"R0763","reasoning:AAAI-2025-1293":"R0764","reasoning:AAAI-2025-2219":"R0765","reasoning:AAAI-2026-3029":"R0766","reasoning:AAAI-2026-3148":"R0767","reasoning:AAAI-2026-3449":"R0768","reasoning:AAAI-2026-3563":"R0769","reasoning:MM-2025-0208":"R0770","reasoning:SIGGRAPH-2026-0075":"R0771","reasoning:SIGGRAPH-Asia-2024-0116":"R0772","reasoning:SIGGRAPH-Asia-2025-0111":"R0773","reasoning:SIGGRAPH-Asia-2025-0132":"R0774"},"technical_index":{"arxiv-1808.06601":"R0001","arxiv-2302.01329":"R0002","arxiv-2306.02018":"R0003","arxiv-2311.16933":"R0004","arxiv-2312.03641":"R0005","arxiv-2404.02101":"R0006","arxiv-2409.18964":"R0007","arxiv-2409.16160":"R0008","arxiv-2407.08136":"R0009","arxiv-2412.03603":"R0010","arxiv-2503.07598":"R0011","arxiv-2502.03207":"R0012","arxiv-2502.01061":"R0013","arxiv-2503.24379":"R0014","arxiv-2503.03751":"R0015","arxiv-2503.21781":"R0016","arxiv-2602.09600":"R0017","arxiv-2603.29092":"R0018","arxiv-2212.11565":"R0019","arxiv-2311.17982":"R0020","arxiv-2406.14477":"R0021","arxiv-2405.18750":"R0022","arxiv-2407.08737":"R0023","arxiv-2412.15689":"R0024","arxiv-2506.08009":"R0025","arxiv-2505.23656":"R0026","arxiv-2509.08826":"R0027","arxiv-2501.13918":"R0028","arxiv-2511.21541":"R0029","arxiv-2506.01689":"R0030","arxiv-2602.02214":"R0031","arxiv-2604.19234":"R0032","arxiv-2603.21872":"R0033","arxiv-2603.27866":"R0034","arxiv-2605.11723":"R0035","arxiv-2605.15458":"R0036","arxiv-2005.12126":"R0037","arxiv-2310.06114":"R0038","arxiv-2309.17080":"R0039","arxiv-2402.15391":"R0040","arxiv-2408.14837":"R0041","arxiv-2405.12399":"R0042","arxiv-2503.18938":"R0043","arxiv-2504.12369":"R0044","arxiv-2510.10125":"R0045","arxiv-2508.13009":"R0046","arxiv-2512.16924":"R0047","arxiv-2509.24527":"R0048","arxiv-2609.08230":"R0049","arxiv-2606.09803":"R0050","arxiv-2607.02517":"R0051","arxiv-2609.34470":"R0052","arxiv-2609.34606":"R0053","arxiv-2609.03952":"R0054","arxiv-2412.07772":"R0055","arxiv-2412.05263":"R0056","arxiv-2407.14505":"R0057","arxiv-2601.01720":"R0058","arxiv-2512.07469":"R0059","arxiv-2603.06408":"R0060","arxiv-2503.11647":"R0061","arxiv-2506.18903":"R0062","arxiv-2504.00983":"R0063","arxiv-2408.06072":"R0064","arxiv-2407.02371":"R0065","arxiv-2410.05954":"R0066","arxiv-2506.04216":"R0067","arxiv-2602.08032":"R0068","arxiv-2509.22622":"R0069","arxiv-2502.06734":"R0070","arxiv-2501.19252":"R0071","arxiv-2506.03517":"R0072","arxiv-2605.10434":"R0073","arxiv-2606.16449":"R0074","arxiv-2605.15199":"R0075","arxiv-2601.01528":"R0076","arxiv-2602.05986":"R0077","arxiv-2603.12146":"R0078","arxiv-2603.06048":"R0079","arxiv-2601.05239":"R0080","arxiv-2606.09150":"R0081","arxiv-2606.04775":"R0082","arxiv-2606.04737":"R0083","arxiv-2605.15196":"R0084","arxiv-2605.15190":"R0085","arxiv-2605.15042":"R0086","arxiv-2605.14269":"R0087","arxiv-2605.12938":"R0088","arxiv-2605.08712":"R0089","arxiv-2604.28169":"R0090","arxiv-2603.08403":"R0091","arxiv-2602.08277":"R0092","arxiv-2601.10553":"R0093","arxiv-2601.04068":"R0094","arxiv-2601.04033":"R0095","arxiv-2606.17730":"R0096","arxiv-2606.07326":"R0097","arxiv-2603.28757":"R0098","arxiv-2602.04876":"R0099","arxiv-2601.05138":"R0100","arxiv-2601.00393":"R0101","arxiv-2602.22208":"R0102","arxiv-2603.16666":"R0103","arxiv-2602.23956":"R0104","arxiv-2603.02802":"R0105","arxiv-2602.15031":"R0106","arxiv-2603.09094":"R0107","arxiv-2610.00812":"R0108","arxiv-2604.23789":"R0109","arxiv-2602.00607":"R0110","arxiv-2602.23969":"R0111","arxiv-2605.14843":"R0112","arxiv-2605.00873":"R0113","arxiv-2603.16271":"R0114","arxiv-2604.24764":"R0115","arxiv-2601.05729":"R0116","arxiv-2609.38146":"R0117","arxiv-2605.14382":"R0118","arxiv-2605.14274":"R0119","arxiv-2605.10543":"R0120","arxiv-2605.07503":"R0121","arxiv-2604.04142":"R0122","arxiv-2604.28078":"R0123","arxiv-2605.22996":"R0124","arxiv-2605.19995":"R0125","arxiv-2604.19679":"R0126","arxiv-2604.01761":"R0127","arxiv-2603.27520":"R0128","arxiv-2603.26599":"R0129","arxiv-2603.25931":"R0130","arxiv-2603.04379":"R0131","arxiv-2603.17461":"R0132","arxiv-2603.12257":"R0133","arxiv-2603.17808":"R0134","arxiv-2603.00159":"R0135","arxiv-2603.03646":"R0136","arxiv-2603.19672":"R0137","arxiv-2603.11421":"R0138","arxiv-2603.15478":"R0139","arxiv-2602.09609":"R0140","arxiv-2602.08820":"R0141","arxiv-2602.01624":"R0142","arxiv-2601.16933":"R0143","arxiv-2601.11087":"R0144","arxiv-2601.07287":"R0145","arxiv-2601.01352":"R0146","arxiv-2603.21547":"R0147","arxiv-2605.30090":"R0148","arxiv-2605.23271":"R0149","arxiv-2605.03849":"R0150","arxiv-2604.07026":"R0151","arxiv-2606.01164":"R0152","arxiv-2609.38123":"R0153","arxiv-2609.35560":"R0154","arxiv-2609.34722":"R0155","arxiv-2609.25652":"R0156","arxiv-2609.22641":"R0157","arxiv-2609.10540":"R0158","arxiv-2609.03563":"R0159","arxiv-2608.29910":"R0160","arxiv-2603.17652":"R0161","arxiv-2606.20083":"R0162","arxiv-2606.13376":"R0163","arxiv-2606.10135":"R0164","arxiv-2606.09507":"R0165","arxiv-2603.17117":"R0166","arxiv-2603.07145":"R0167","arxiv-2603.05449":"R0168","arxiv-2602.10104":"R0169","arxiv-2602.09022":"R0170","arxiv-2602.08025":"R0171","arxiv-2602.02393":"R0172","arxiv-2608.24680":"R0173","arxiv-2607.14076":"R0174","arxiv-2605.15178":"R0175","arxiv-2604.02330":"R0176","arxiv-2603.03482":"R0177","arxiv-2607.14005":"R0178","arxiv-2608.13489":"R0179","arxiv-2606.23296":"R0180","arxiv-2606.18960":"R0181","arxiv-2606.09457":"R0182","arxiv-2601.21998":"R0183","arxiv-2608.27328":"R0184","arxiv-2604.01641":"R0185","arxiv-2603.25716":"R0186","arxiv-2603.06679":"R0187","arxiv-2608.23565":"R0188","arxiv-2608.14022":"R0189","arxiv-2607.11836":"R0190","arxiv-2608.13552":"R0191","arxiv-2602.08971":"R0192","arxiv-2605.25874":"R0193","arxiv-2606.31734":"R0194","arxiv-2606.25473":"R0195","library-P012":"R0196","library-P021":"R0197","arxiv-2507.17131":"R0198","arxiv-2512.24695":"R0199","arxiv-2603.28767":"R0200","library-P036":"R0201","library-P059":"R0202","library-P061":"R0203","arxiv-2206.09203":"R0204","library-P146":"R0205","library-P149":"R0206","arxiv-2412.06438":"R0207","arxiv-2511.17481":"R0208","library-P159":"R0209","arxiv-2605.12978":"R0210","library-P161":"R0211","arxiv-2410.08464":"R0212","arxiv-2602.07055":"R0213","arxiv-2605.18746":"R0214","arxiv-2402.03162":"R0215","arxiv-2411.15262":"R0216","arxiv-2606.03951":"R0217","arxiv-2606.31392":"R0218","arxiv-2606.32026":"R0219","library-P112":"R0220","arxiv-2506.14965":"R0221","arxiv-2506.14404":"R0222","arxiv-2512.07802":"R0223","arxiv-2606.26907":"R0224","arxiv-2511.16669":"R0225","arxiv-2604.02296":"R0226","arxiv-2606.24225":"R0227","arxiv-2605.27589":"R0228","arxiv-2609.28466":"R0229","arxiv-2512.04519":"R0230","arxiv-2605.06924":"R0231","arxiv-2603.23864":"R0232","arxiv-2608.20699":"R0233","arxiv-2606.02564":"R0234","arxiv-2608.03974":"R0235","arxiv-2305.16291":"R0236","arxiv-2503.21055":"R0237","arxiv-2601.04778":"R0238","arxiv-2603.12310":"R0239","arxiv-2608.29621":"R0240","arxiv-2609.24997":"R0241","arxiv-2609.34309":"R0242","arxiv-2605.20342":"R0243","library-VG61":"R0244","arxiv-2608.08469":"R0245","arxiv-1907.01172":"R0246","arxiv-2403.01599":"R0247","arxiv-2006.06485":"R0248","arxiv-2411.16683":"R0249","arxiv-2609.35134":"R0250","arxiv-2307.01928":"R0251","arxiv-2411.11844":"R0252","arxiv-2403.14773":"R0253","arxiv-2610.03510":"R0254","arxiv-2410.10076":"R0255","arxiv-2309.15091":"R0256","arxiv-2610.09326":"R0257","arxiv-2609.06373":"R0258","lineage-DRAM":"R0259","arxiv-2402.17119":"R0260","arxiv-2302.05543":"R0261","arxiv-2601.20622":"R0262","arxiv-2511.08971":"R0263","recent-CLARITI":"R0264","arxiv-2602.13433":"R0265","recent-IGCLAR":"R0266","recent-SC2":"R0267","arxiv-2512.15716":"R0268","recent-COME":"R0269","arxiv-2509.24798":"R0270","recent-MULTI":"R0271","recent-ARIFY":"R0272","recent-COACH":"R0273","arxiv-2510.04618":"R0274","arxiv-2601.21557":"R0275","arxiv-2601.22528":"R0276","arxiv-2506.06658":"R0277","arxiv-2511.23002":"R0278","arxiv-2510.15831":"R0279","arxiv-2505.22954":"R0280","arxiv-2609.24955":"R0281","arxiv-2608.08982":"R0282","arxiv-2608.28460":"R0283","arxiv-2606.07649":"R0284","arxiv-2605.07800":"R0285","arxiv-2202.03169":"R0286","arxiv-2106.02997":"R0287"}}