{"title":"XR / AIGC 视频生成：问题谱系与逐篇拆解","date":"2026-10-15","publication":"GitHub Topics 集成版；本地与线上共享结构化内容","notes":["[推断] 本报告的“谱系”是按研究问题与操作定义整理的近邻图；仅 OMNI→VOID 等已核实关系可称明确技术继承，其余不冒称 citation genealogy。奠基代表是本脉络选取的早期支点，不宣称历史上最早。","[推断] “最新”指检索截至 2026-10-15、实际核实的近邻；不是对全网论文的穷尽排序。2026 年较多工作仅核实预印本状态，未核 venue 处保留待核。","[推断] 论文自己的 novelty/结果主张与本报告的证据强度判断分列；使用原作者报告的实验结果，未复现训练或运行模型。","[推断] 实验设计为本报告提案，尚未执行；未给出虚构 pilot 数字。真人研究样本量应预注册并依目标效应做 power analysis，不能把论文的人评人数复用为新实验的合理样本量。","[待核] L9 同题近期代表作未达到 3 篇；GPS 的 arXiv ID/DOI 未找到；GENEX、STREAM 的 human-study 总人数在本轮阅读中未确认；BLACK 正文与附录 human-baseline 口径不一致，已并列记录。未确认不等于论文未报告。","[论文] SPIRAL 正文为 v3，GENEX 正文为 v3，STATE 为 v6、VIS 为 v2；书目年份与正文版本分别记录。发表 venue 并不意味着此处读取的后续 arXiv 修订内容已全部经过该次同行评审。","[推断] 2026-10-15 补充阅读卡与原 38 篇全文拆解分层展示；发表状态与任务距离分开。XR-L 按最新要求仅研究 how，GT 存在于用户处；新颖性最多表述为“本论文池内未见”。"],"taxonomy_schema":{"tasks":["文本或图像生成视频","条件控制生成","编辑与修改","长视频与多镜头叙事","动作条件的预测与交互世界模型","3D/4D 一致生成","视频作为推理或回答","其他"],"contributions":["表示","条件注入","架构与生成范式","记忆与上下文","训练信号","推理时方法","数据","评测"],"control_signals":["文本","图像","相机","轨迹或光流","姿态或结构图","动作","音频","物理或模拟代理","身份参考"],"target_properties":["画质","短时连贯","实体一致","几何一致","环境记忆","因果或状态一致","物理合理","指令遵循","效率与延迟"],"generation_paradigms":["离线整段生成","因果自回归或流式","少步或实时"],"paper_types":["方法","基准","数据集","分析","综述"],"notes":["分类均为[推断]：按论文的主要任务与贡献归类，不按全部可演示应用重复贴主轴；类型按主要交付选一个。","图像信号包含参考图与源视频帧；姿态或结构图在此包含 mask、depth、box、layout 等空间条件。纯理解/QA 输入列入任务定义，不冒记为视频生成控制。","生成范式按视频的主合成过程判定；chunk/shot-level autoregressive 不代表单帧因果 attention，interactive 或 training-free 也不自动等于少步/实时。","非视频生成任务与不限定 renderer 的 agent 框架使用“不适用”并说明原因；没有直接对应的目标性质同样解释，不强行贴生成一致性标签。","目标性质描述论文优化/评测的目标，不意味着已经保证该性质；“必须保持”是任务约束整理，单独标为[推断]。","证据按已核实的实验标注；有定量实验的论文中的 demo 写作“定性展示”，只有全部证据均为定性才写“仅定性展示”。未知人评人数保留[待核]。"]},"directions":[{"code":"XR-C","name":"面向 XR 操作后果的因果视频生成与编辑","definition":"在现实操作教程中，生成漏步、换序或机制变化后的后果对照，同时保留不受干预影响的对象与步骤。","lineages":["L1","L2"],"history":"原始 procedure planning 问“从起点到目标做哪些动作”；SCHEMA 显式引入状态；STATE 把 counterfactual 用作表征监督，VANS 把输出改为下一事件视频。另一分支从对象后效去除扩展到 VOID 的动态删除，GIVE 的几何变化和 PHYS 的定时物理干预。两条分支尚不能仅凭名称合并成同一 causal generation 能力。","boundary":"已有较直接证据：程序数据集规划/表征结果，以及 engine 中 paired geometric/physical targets。较弱证据：真实复杂场景的“如果没做这一步”主要是偏好、judge 或定性；VANS 人评是答案质量，并非教学执行的随机对照。","core":"局部观察无法区分多个隐藏机制时，系统应何时给出分支后果、何时先获取证据，而不是生成唯一看似可信的教程？","why":"(a) 设计观察等价、干预后结果不同的机制，观察数据再多也不能在当前实例确定真机制，属于可识别性问题；(b) 上述参考基准未联合测试该歧义与校准拒绝/分支；(c) 可控制隐藏机制和证据揭示时刻直接检验。","experiment":{"case":"XR 装配或支撑物实验：外观相同的零件分别有隐藏锁扣／磁吸／纯支撑；源视频中都静止，移除支撑后的行为不同。先在引擎建立可复现实验，再用可切换机构录制受控真实视频。","operation":"2×2：隐藏机制可辨/不可辨 × 允许/不允许一次诊断性观察或轻触；固定初始条件、编辑时间、相机、随机种子，生成漏步/换序的结果。","controls":"同 backbone 的 VANS/文本条件、VOID 删除、PHYS 重建、显式多假设＋主动查询；oracle mechanism 为上界；另设同成本随机查询。","metrics":"对象轨迹与接触/步骤结果；未受影响对象状态误改率；不同机制下预测分布 calibration、错误确定性；查询成本；后续真人执行成功另作为教学验证。","falsifier":"如果在观察等价组中，单一假设方法仍可在预注册的隐藏机制随机化下稳定预测正确，需先查是否有数据泄漏；若多假设/诊断查询无收益，则当前核心假设或观测设计不成立。"},"neighbours":"PHYS 是物理重建直接竞争者；VOID 覆盖删除，GIVE 覆盖几何算子。DSCM 提供 abduction–action–prediction 原理，但依赖结构假设。差别不能写成“我们也考虑因果”，应是部分可识别下的后果分布与查询协议。","verdict":"与 AIGC-A 合并为一个方法方向；XR-C 是需要部分观察和操作证据的实例。只把电影对象换成工具不构成独立研究。","sources":["PROC","SCHEMA","STATE","VANS","DSCM","VOID","GIVE","PHYS"]},{"code":"XR-L","name":"面向 XR 教程的主动澄清：用户持有 GT，研究 how","definition":"输入不充分但目标或指代真值存在于用户认知中，且需要澄清已给定；XR agent 选择并执行澄清交互，获取用户答案并可靠更新教程。","lineages":["L3"],"history":"从条件询问与 dialogue actions，推进到第一人称多模态消歧、可回答性奖励与 AR 候选呈现。原来的主动探索分支仍保留在 AIGC-D 与旧范围档案。","boundary":"Agent 自己选择并发起问题、对象高亮、候选对照或确认界面；用户提供 GT。主动感知未知物理事实、何时提问和跨任务数据选择均不是当前主问题。","core":"固定澄清触发时刻后，agent 如何选择提问内容、呈现形式与候选对照，让用户低负担地给出可执行答案？","why":"关键自变量是澄清行动及其语义可区分性；模型置信度和触发策略不作为主贡献。需同时检验答案恢复和后续视频执行，不能止于问句质量。","experiment":{"case":"装配教程里用户知道目标部件、操作目的或所指对象，但初始表达不充分；系统已收到“需要澄清”信号。","operation":"固定触发时刻、可用传感信息和生成器，让 agent 选择开放问句、候选对照、高亮指认或确认组合；用户只回答意图／指代，不承担 agent 的策略选择。","controls":"固定模板、仅文本问答、候选列表、随机同成本交互、GPS-style DAG、Plug-and-Play Clarifier；oracle user answer。","metrics":"目标／指代恢复率、问题可回答率、用户时间与轮数、错误确认率、答案到教程约束的执行率；分别报告澄清失败和 renderer 失败。","falsifier":"若同成本固定候选列表达到相同效果，或 GT 已正确恢复但视频不执行，不能将收益归因于更好的澄清策略。"},"neighbours":"Plug-and-Play Clarifier、GPS、CLARITI 最接近 how；Uncertain Pointer 支撑界面呈现。前者部分流程仍要求用户调整位置，须区分 agent 发起交互与 agent 独立获得所有信息。","verdict":"XR-L 是 AIGC-D 中“用户持有 GT”的受限子问题；共享缺口表示和答案执行，但不与广义主动探索、最优停止或数据采集研究合并。","sources":["GPS","ACT","KNOW"]},{"code":"XR-S","name":"新证据下可纠错的 XR 教程视频世界模型","definition":"真实观察否定教程假设时，撤销错误状态及其依赖，只修改受影响的未来演示。","lineages":["L5","L6"],"history":"生成 memory 先服务于“别忘记旧画面”；WorldMem 强化 spatial recall，PermaVid 已明确处理编辑后旧 context 失效。BlackSwan 则研究证据如何推翻解释，VIS-Ground 把异质上下文依赖转换为续写约束。XR-S 的交集是证据纠错如何传到未来生成。","boundary":"已可靠展示局部历史保持、编辑后重访、组件检索和某些约束续写。尚不足以证明：带噪传感/教师证据否定隐藏状态时，所有依赖的未来动作都撤销、无关已确认事实不变。仅添加一段新 prompt 或刷新 memory 不足以界定这个目标。","core":"当生成假设与新观测冲突且影响跨越多个步骤时，能否识别“哪个推断错了”，并使 memory、计划与生成条件一致地局部失效？","why":"(a) 涉及证据来源、冲突裁决和依赖失效语义，增加容量并不定义何者应信；(b) 现有 memory consistency 评测不等于含噪反证的纠错精确率；(c) 可控制错误来源、证据可靠性及 downstream dependency GT。","experiment":{"case":"生成装配教程先推断“锁扣已扣上”；后续侧视图揭示未扣紧，需修改测试/承重步骤，却保持已装好的其它部件。","operation":"插入早期错误假设；在不同延迟揭示视觉/传感/教师证据；正交控制证据准确/误报和依赖链长度。","controls":"append-only history、整段重生成、PermaVid-style spatial invalidation、A2RD critic loop、VIS-Ground constraints；provenance＋dependency invalidation；oracle dependency graph。","metrics":"错误状态撤销 precision/recall、下游错误残留率、无关事实误改率、回到旧视角后的错误复现、未来教程约束满足；固定算力下的修正延迟。","falsifier":"若简单最新帧/最近 prompt 替换已达到同样依赖修正与无关保留，则复杂 provenance graph 没有必要；若噪声下新方法更易推翻正确事实，核心价值未成立。"},"neighbours":"PermaVid 必须列作强 baseline：已知 edit region 的 RGB/depth 失效已有实现。VIS-Ground 必须测试 source/history/intent 约束。候选差别是外部证据的可靠性与非空间依赖；在本次检索范围内未找到完整覆盖此协议的已验证工作。 特别注意：VIS 已有 evidence-linked predicates、pass/fail/unknown 和 targeted script revision，因此 provenance graph 本身不是可守住的新意；必须以旧推断失效后的依赖撤销和 memory/KV 残留为区别。","verdict":"与 AIGC-B 合并；就你的 abductive reasoning 背景而言，是自然衔接，但“可纠错 memory”这个宽标题已经不够新。","sources":["STREAM","WORLD","PERMA","A2RD","WEAVE","BLACK","VIS"]},{"code":"XR-E","name":"跨任务持续改善的 XR 教程生成","definition":"将多个学员的误操作与教师反馈转成可验证的持久教学改进，同时避免旧任务能力退化。","lineages":["L7"],"history":"Voyager 积累可执行技能；VideoAgent 用外部反馈修视频计划并收集再训练数据；SPIRAL 联合闭环和生成器后训练；CineForge 已从多故事轨迹提炼持久策略。MEMFAULT 表明持续归纳 memory 也会伤害原有能力。","boundary":"“系统能跨任务改进”已经有直接近邻；“改得越来越好”没有无条件证据。模拟任务成功、故事质量、critic 分数，都不直接测真实学员学习增益或保持旧知识。","core":"怎样辨别一次学员失败是教程错误、学员差异还是环境异常，并只把有跨用户证据的修复升级为持久策略？","why":"(a) 教程效果与能力/任务分配混淆，增加反馈量不能替代因果对照；(b) 现有视频 self-improvement papers 不直接提供教学归因和跨学员保留证据；(c) 可随机化反馈来源与策略部署，测试迁移及回归。","experiment":{"case":"多个装配任务与新手/熟练学员；一个错误可能由镜头没拍到锁扣、术语不清或学员粗心导致。","operation":"分离 current-session repair 与 persistent update；按任务/学员随机分派教程版本；控制反馈噪声，保留原始 episode。","controls":"无更新、每次都写 memory、episodic retrieval、CineForge-style stage patch、SPIRAL-style parameter update、证据门控策略；所有方法用同一反馈池。","metrics":"新学员未见任务执行成功/保持测验、旧任务 regression、最坏子群收益、反馈样本效率；跨用户划分而不是随机视频划分。","falsifier":"若盲测新用户无学习增益，只有 critic 分数提高，则不能称教学 self-improvement；若 episodic retrieval 已匹配全部收益，持久规则归纳或参数更新不必要。"},"neighbours":"ARify／AgentCoach 最接近教学界面；ACE／MCE／Darwinian Memory 支撑经验与策略更新；SILVR 连接执行反馈。VISTA 属单任务优化，CineForge 属跨故事策略，DGM 属自修改代码，距离不同。","verdict":"[推断] 可把跨任务教学策略迁移独立为候选问题，但须有留出任务的学员收益、旧任务回归和错误经验隔离证据；仅 self-evolution 命名或持续写记忆不足以立项。","sources":["VOY","VIDEOAG","SPIRAL","CINE","MEMFAULT"]},{"code":"AIGC-A","name":"事件干预后的选择性视频生成与编辑","definition":"固定同一历史与明确干预，只改变受影响的对象、事件和时间段，保留无关状态。","lineages":["L2","L1"],"history":"对象删除最初解决补背景及 effects；VOID 将目标推进到互动轨迹。GIVE 用几何 pre/post 状态扩展算子，PHYS 从观察重建后重新模拟；DSCM 提醒个体 counterfactual 还要求明确 abduction 和不变因素。STATE 的 counterfactual 是监督信号，VANS 的未来回答则是邻接任务。","boundary":"已解决若干受限算子的局部问题；“所有干预都只改该改的部分”尚无通用保证。synthetic paired GT 可测物理/几何；真实视频 preference 只能支持感知合理性。新 novelty 必须避开自动 mask、LLM reasoner、物理合理等已有宽表述。","core":"干预的影响范围不是输入已知 mask 时，如何同时保证因果后果的完整修改与非后代状态的保留，并在不可识别时承认多解？","why":"(a) 编辑边界应由机制决定，而非固定像素面积；未知机制会产生观察等价；(b) 现有分数常把“改对后果”与“少改背景”混为一个总分；(c) 受控机制提供 descendants/non-descendants 与时间边界 GT。","experiment":{"case":"游戏引擎中连锁机关、球体碰撞、支撑结构；同一源 episode 分别干预删除、属性、动作时刻，生成电影/游戏替代结局。","operation":"固定同一 external noise/seed 与历史；改变 intervention type、chain depth、远距离依赖、隐藏参数；区分 observed factual future 可用和只有 prefix 两种设定。","controls":"OMNI、VOID、GIVE、CWMDT、PHYS，各按其支持算子公平比较；共享 backbone 的 mask/structured state/trajectory control；oracle impact set。","metrics":"affected-state accuracy 与 unaffected-state error 分开；trajectory/contact error；影响集合 precision/recall；干预前一致性；跨算子泛化，不能只合成一个 VLM 总分。","falsifier":"如果已知 mask/轨迹 oracle 足以解释全部收益，问题在上游机制推断；如果标准编辑器在相同信息下达标，则候选选择性机制没有额外贡献。"},"neighbours":"直接竞争就是 VOID/GIVE/PHYS/CWMDT；DSCM 提供定义而非视频成绩。较可守住的差别是作用集合的可识别性、个体不变性与证据校准；在本次检索范围内未找到完整覆盖上述组合协议的工作。","verdict":"属于 controllable/physical video editing 的明确近邻，不能称偏离 video generation；但应与 XR-C 合并并选一类机制，不以“大一统因果编辑”作为首篇目标。","sources":["DSCM","OMNI","VOID","GIVE","PHYS","CWMDT","STATE","VANS"]},{"code":"AIGC-B","name":"外来反证驱动的视频生成记忆修订","definition":"延迟到达的外部证据推翻早期推断时，局部修订生成记忆和未来视频，保留其它已确认事实。","lineages":["L5","L6"],"history":"从 StreamingT2V 的第一段锚定到 WorldMem 的 pose/time memory，研究主要要求历史一致；PermaVid 已明确解决 edit-induced staleness。BlackSwan 提出 defeasible reasoning，VIS-Ground 联合多源上下文约束；WEAVE 说明内部 KV 也可以组件化选择。","boundary":"不能再把“记忆可更新”“只取相关角色”“编辑后保持更新”作为空白。现有证据更少覆盖互相矛盾来源、可靠性差异、被推翻的 latent assumption 与其非局部依赖。此处是本次检索的条件性判断，不是断言无人研究。","core":"如何使生成器使用的内部条件忠实反映“哪些事实仍有效”，而非把相关但已失效的记忆继续写入新视频？","why":"(a) relevance 与 validity 是不同语义，更多 history 不会自动定义撤销；(b) 高一致性甚至可能奖励重现错误；(c) 可操纵旧错误进入 prompt、frame bank 或 KV 的位置，分别检测其残留。","experiment":{"case":"电影/动漫人物早先被误认；游戏 NPC 随后提供可靠证据澄清身份；保留同一人的衣着、空间及无关经历，只撤销依赖错误身份的后续互动。","operation":"对照真实 world change、编辑 instruction、纠正旧 belief 三种事件；注入不同可信度、延迟、作用范围的证据。","controls":"append new prompt、清空全部 memory、PermaVid、WEAVE routing、VIS constraints；显式 provenance state 到 memory/KV invalidation 的映射。","metrics":"事实有效性准确率、错误重现、非受影响一致性、dependency repair recall、拒绝假证据率；相同生成预算与随机种子。","falsifier":"如果单纯 prompt 更新即可跨多轮消除错误而不损坏无关事实，则内部干预没有必要；内部 cache 修订若牺牲其它角色，不能只报纠错成功率。"},"neighbours":"PermaVid 是最需要回应的近邻；WEAVE 提供内部 memory routing 实现；BlackSwan 提供证据事件设计；VIS-Ground 提供约束 baseline。区别应落实为外部证据真实性裁决与依赖撤销，而非模块命名。 特别注意：VIS 已有 evidence-linked predicates、pass/fail/unknown 和 targeted script revision，因此 provenance graph 本身不是可守住的新意；必须以旧推断失效后的依赖撤销和 memory/KV 残留为区别。","verdict":"与 XR-S 是同一核心；推荐共同保留“证据来源—失效依赖—生成控制”的窄问题。新颖性风险高于仅看原库时的判断。","sources":["STREAM","WORLD","PERMA","WEAVE","A2RD","BLACK","VIS"]},{"code":"AIGC-C","name":"事件时间与信息揭示时间解耦的视频生成","definition":"保持世界事件不变，控制观众从画面何时得知关键事实，并保证早期留白与后续揭示相容。","lineages":["L8","L9"],"history":"计算叙事早已区分故事事件与读者获得的信息；Dramatis 模拟读者记忆与脱困计划，SUSP 直接实验 revelation timing。视频侧从 DirectorGPT、OneStory 到 VIS/MovieGrid 主要改善计划、跨镜头一致性和续写；这条实现链未自动接上观众知识的测量。","boundary":"“事件／揭示分离”不是全新问题。SUSP 的时机比较不显著，不能用它保证 suspense 收益。现有 video consistency 指标未直接检验角色知识、观众知识、真实世界状态的分离；本次未找到第三篇符合 L9 同题条件的近期代表作。","core":"是否能在生成器掌握完整世界状态的同时，使早期 pixels 对指定秘密不泄漏、后期又可一致解释；XR 自由视角下这种约束能否保持？","why":"(a) 原则上应限制观众可获得的信息，非靠更好图像质量；(b) 角色外观一致、镜头数和故事评分不测泄漏；(c) 固定世界事件、只改变 reveal schedule，可直接测观众知识获得时间。","experiment":{"case":"剧本杀短片：关键道具已被调包但到指定镜头才揭示；为同一事件图制作早揭示/晚揭示两版，随后允许 XR 观众换视角。","operation":"固定 actor actions/事件时间/随机种子，只改 camera、遮挡、画面线索、reveal schedule；另操纵是否给 renderer 完整秘密信息。","controls":"普通 storyboard prompt、OneStory、VIS-Ground constraints、显式 world/character/viewer state planner；2D 固定视角与 XR 可移动视角分别比较。","metrics":"分时点秘密问答与 confidence、提前泄漏率、预期 reveal 后识别率、后续回看一致性；人类读者为主，多个 VLM probes 为辅助；suspense/享受度单列，不能替代信息控制。","falsifier":"若被遮挡和 prompt-only 基线在自由视角下已同样控制信息，复杂方法无优势；若泄漏低只是画面完全不可理解，需用后期可识别率排除伪成功。"},"neighbours":"SUSP 已操作 revelation timing；Dramatis 已模型化 reader state；VIS 处理依赖约束但不是泄密 benchmark；OneStory/MovieGrid 是渲染近邻。候选新问题需限定为 video perceptual information constraints，而不是叙事学术语搬运。 VIS 也报告 dataset information-leakage audit；本提案应区分基准答案泄漏检查与分时点观众知识控制，不能声称该文完全不考虑 leakage。","verdict":"可独立于 A/B 保留，但论文重心应是视频中可检验的信息边界；如果只调整台词/镜头顺序，更多是现有 narrative planning 的应用。","sources":["DRAM","SUSP","DIRECTOR","ONE","VIS","GRID"]},{"code":"AIGC-D","name":"获取最少但足够的上下文与生成控制信号","definition":"在固定用户负担与生成预算下，选择能提高最终视频约束满足率的必要上下文或控制信号。","lineages":["L3","L10"],"history":"KnowNo/GPS/ACT 研究何时问什么；ControlNet 研究给定条件怎样控制；SketchDynamics 将草图解释与用户澄清结合，CogOmniControl 内部融合专业意图；Qwen/VideoGen-Agent 已将检索、工具、反馈用于生成。D 位于 acquisition 与 execution 的接缝。","boundary":"主动 agent、稀疏输入、专业意图和多条件控制都有近邻。没有证据允许把“用户少写了 prompt”直接当最小充分性，或把更强 renderer 的提升归于更好的询问。","core":"如何区分缺少事实、意图本来多解、控制信号不足和 renderer 执行失败，进而选择正确且成本最低的补充？","why":"(a) 信息不足与执行能力不足是不同因果因素，多搜资料不能修复不支持的控制；(b) 现有总体 benchmark 容易混淆工具升级与 policy；(c) 可固定或替换 oracle context/control/renderer 逐层诊断。","experiment":{"case":"动漫短剧生成：同样要求“角色绕过桌子并拿起红杯”，分别缺角色参考、遮挡后的布局、运动路径或时间要求；部分案例故意给齐条件但 renderer 不能执行。","operation":"候选 query 为 ask、reference、depth/sketch、trajectory、keyframe；操纵每种价格和可用 renderer，预算恒定。","controls":"全信号、固定 prompt、随机同成本、置信阈值、GPS-style rule query、VideoGen-Agent；同 backbone；再用 oracle context 与 oracle control 分别定位错误。","metrics":"视频约束满足—用户秒数/工具费/token/延迟曲线；minimal-subset deletion test、过度询问率、无效重试率；跨 renderer 迁移。","falsifier":"若一个固定 low-cost condition 即可解释全部收益，则自适应策略不必要；若 oracle 输入仍失败，不能将其记为 acquisition failure。"},"neighbours":"Qwen-Image-Agent 已明确定义 context gap；VideoGen-Agent 已训练共享工具策略；SketchDynamics 已做自由草图澄清；CogOmniControl 已做意图和内部生成对齐。更窄差别是可验证充分性和失败归因。","verdict":"与 XR-L 合并为 acquisition research；外部 control 与内部 injection 是执行层可替换模块，不应单凭内部/外部二分声称新方向。","sources":["KNOW","ACT","GPS","QWEN","VGEN","CONTROL","SKETCH","COG"]}],"lineages":{"L1":{"name":"视觉目标条件的 procedure planning → 状态变化监督 → 视频形式的下一事件回答","papers":["PROC","SCHEMA","STATE","VANS"],"relation":"直接任务脉络与邻接扩展。PROC→SCHEMA 延续 procedure planning；STATE 改为表征监督；VANS 改为 next-event video answer。后两者不是同一 benchmark 上的连续排名。","distance":"与 XR 教程接近，但“步骤规划／表征／下一事件答案”均不自动包含同一历史下的显式事件干预。"},"L2":{"name":"个体反事实的结构化定义，与对象后效删除／几何及物理编辑的两条实现分支","papers":["DSCM","OMNI","CWMDT","VOID","GIVE","PHYS"],"relation":"DSCM 是概念奠基参照，不声称视频方法继承了其 SCM 或定理。实现分支中 OMNI→VOID 有明确权重初始化和任务推进关系；CWMDT、GIVE、PHYS 分别选择结构化语义、几何指令、重建后物理模拟。","distance":"最直接对应 A；对 XR-C 还缺隐藏状态、主动操作及教学结果。geometry-dependent effects 与动态 causal consequences 必须分别评测。"},"L3":{"name":"不充分输入下的主动澄清／信息获取 → 生成工具策略","papers":["KNOW","ACT","GPS","QWEN","VGEN"],"relation":"方法问题近邻。KnowNo 的 calibrated help、ACT 的 dialogue action、GPS 的条件图、Qwen 的 context acquisition、VideoGen-Agent 的工具策略，分别实现“缺信息时怎么办”；并非同一理论保证逐篇传递。","distance":"自动获取更多 context 已有实作；跨 ask/look/search/demo/control 的最小充分预算及最终视频风险，需要单独定义。"},"L4":{"name":"不确定性下的主动视觉实验与空间观察","papers":["IVR","GENEX","S3"],"relation":"共享部分观察下的信息获取问题，但三者分别执行真实模拟试验、生成想象、处理 streaming observations。这里明确比较这三种证据来源，不将 imagined view 等同 sensed view。","distance":"直接帮助界定 XR-L 的 perception policy；终点是因果识别/决策/空间 QA，而不是教程生成成功。"},"L5":{"name":"生成历史的长期记忆 → 编辑后有效性更新 → 组件级检索","papers":["STREAM","WORLD","PERMA","A2RD","WEAVE"],"relation":"生成 memory 的直接技术邻域；STREAM 作为本表早期代表，不声称它是所有生成记忆的最早工作。WORLD 加入 pose/time；PERMA 显式使过时 memory 失效；A2RD 用 agent 检索修订；WEAVE 选择内部 KV。","distance":"“有可更新 memory”已经被覆盖；来自新证据的错误诊断、来源裁决、依赖传播和选择性撤销，不能用普通一致性代替。"},"L6":{"name":"部分视觉证据下的解释与 belief/context 修订","papers":["BLACK","GENEX","VIS"],"relation":"问题近邻而非已核实的引用链。BlackSwan 是新证据推翻解释的任务起点；GenEx 用想象更新 belief；VIS-Ground 将 source/history/intent 的依赖落实为续写约束。","distance":"三者既不共用证据来源，也不都做 memory invalidation。用它们界定 B/S 的缺口，比把所有 consistency papers 都叫 belief revision 更准确。"},"L7":{"name":"执行经验与外部反馈 → 持久技能／生成器／生产策略改进及回归","papers":["VOY","VIDEOAG","SPIRAL","CINE","MEMFAULT"],"relation":"共通的是反馈跨执行保留，但更新对象不同：代码库、生成参数、stage-local policy、文本记忆。MEMFAULT 是反证分支，用于检验持续更新是否会退化。","distance":"跨任务改进并非新空白；XR-E 必须提出关于学习者反馈归因及教学收益的额外问题。"},"L8":{"name":"多场景计划 → 多镜头历史条件化 → 可交互约束续写与长片组织","papers":["DIRECTOR","ONE","VIS","GRID"],"relation":"视频叙事生成的直接近邻：从 scene/layout plan 到 adaptive memory、contextual constraints 和 multi-grid。它们解决的主要是可实现性和一致性，而不是观众知识状态。","distance":"可作为 C 的 renderer/planner baseline；不能把镜头顺序、角色一致性直接当作 revelation timing。"},"L9":{"name":"读者可得信息、脱困计划与 suspense：计算模型 → LLM 叙事操纵","papers":["DRAM","SUSP"],"relation":"Dramatis 作为计算认知奠基代表；SUSP 将相近理论用于 LLM 生成，并直接操纵 information revelation timing。","distance":"在本次检索范围内，未找到第三篇 2024–2026 且同样直接检验信息揭示时机的论文；不以普通 long-story/consistency 工作填足。该脉络明确保留缺额。","missing":"未找到：第三篇满足同题与时间条件、已打开原文的代表作。当前 2 篇；目标 3–6 篇未完全达到。"},"L10":{"name":"外部空间条件 → 自由草图意图澄清 → 专业意图与生成模型内部对齐","papers":["CONTROL","SKETCH","COG"],"relation":"控制信号的执行与解释脉络。ControlNet 奠基于图像条件分支；SketchDynamics 是 HCI 意图交互；CogOmniControl 是视频 cognition/generation 对齐。跨模态迁移关系不等于共同 benchmark。","distance":"解决“给了控制后如何执行”较多；“何时值得请求哪个控制”与“最低需要多少”仍是另一个实验问题。"}},"papers":{"PROC":{"key":"PROC","title":"Procedure Planning in Instructional Videos","authors":["Chien-Yi Chang","De-An Huang","Danfei Xu","Ehsan Adeli","Li Fei-Fei","Juan Carlos Niebles"],"first_year":2019,"year":2020,"venue":"ECCV 2020","venue_source":"https://www.ecva.net/papers/eccv_2020/papers_ECCV/html/1340_ECCV_2020_paper.php","venue_proof":"正式 proceedings","arxiv":"1907.01172","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/1907.01172","full_source":"https://arxiv.org/html/1907.01172v3","version":"1907.01172v3","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/1907.01172v3","origin":"既有补充参考","origin_detail":"旧 topics 数据已有，未列入 papers.json 主表","question":"能否从 instructional videos 学得可规划的状态与动作空间，从起点、终点视觉观察推断中间操作？","case":"CrossTask instructional videos；procedure planning 与 walkthrough planning。","operation":"输入起始和目标视觉观察、规划长度；输出离散动作序列；改变起点、终点与规划跨度。","method":"Dual Dynamics Network 同时学习 forward dynamics 和 conjugate dynamics；对比 UPN 的连续梯度规划以及去掉任一 dynamics 的动作预测模型。","evidence":"CrossTask 上 success rate、动作准确率及序列覆盖指标；规划跨度比较、两个 dynamics 消融；另测 walkthrough 排序。","strength":"支持带动作监督的离线视频规划；序列标签匹配不能证明真实执行成功，也没有视频生成或因果识别实验。","advance":"把从真实教学视频学习的表征用于目标条件规划，而非只预测即将发生的动作。","premise":"训练有动作序列监督；评测提供起终点观察和任务动作集合。","assumption":"可识别的动作词表及观察到的状态足以定位规划问题；相同外观下不同隐藏机制会使该映射不唯一。","locator":"§3 Dual Dynamics Network；§4 Experiments；Tables 1–2","taxonomy":{"task":"其他","task_note":"教学视频中的 procedure planning；不输出视频","contribution_primary":"表示","contribution_secondary":["训练信号"],"control_signals":[],"target_properties":[],"property_note":"规划成功与动作序列正确性","generation_paradigm":"不适用","paradigm_note":"","paper_type":"方法","contribution_focus":"可规划的 state/action latent space；forward/conjugate dynamics 联合约束。","task_definition":{"inputs":"起始视觉片段、目标视觉片段与规划长度。","outputs":"连接起点和目标的离散操作序列。","invariants":"给定起终条件与动作语义；没有要求生成像素保持不变。","judgement":"CrossTask 序列成功率、动作准确率与覆盖/排序指标。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"CrossTask 上 success rate、动作准确率及序列覆盖指标；规划跨度比较、两个 dynamics 消融；另测 walkthrough 排序。","one_sentence_advance":"把从真实教学视频学习的表征用于目标条件规划，而非只预测即将发生的动作。","source":"https://arxiv.org/html/1907.01172v3","locator":"§3 Dual Dynamics Network；§4 Experiments；Tables 1–2"}},"SCHEMA":{"key":"SCHEMA","title":"SCHEMA: State CHangEs MAtter for Procedure Planning in Instructional Videos","authors":["Yulei Niu","Wenliang Guo","Long Chen","Xudong Lin","Shih-Fu Chang"],"first_year":2024,"year":2024,"venue":"ICLR 2024","venue_source":"https://arxiv.org/abs/2403.01599","venue_proof":"作者 arXiv comments 明示 accepted","arxiv":"2403.01599","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2403.01599","full_source":"https://arxiv.org/html/2403.01599v1","version":"2403.01599v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2403.01599v1","origin":"既有补充参考","origin_detail":"旧 topics 数据已有，未列入 papers.json 主表","question":"只建模动作顺序会忽略状态：能否用明确 state changes 改善 procedure planning？","case":"CrossTask、COIN、NIV；给定起终状态的步骤规划。","operation":"LLM 生成动作前后状态文本；视觉观察与状态文本对齐；预测中间状态和步骤。","method":"GPT-3.5 状态描述＋cross-modal contrastive alignment＋两个 decoder；与直接步骤序列建模及单 decoder、随机 memory、无 state alignment 等设置比较。","evidence":"三个 benchmark；CrossTask state alignment 及 mid-state prediction 消融，另测 Viterbi、描述类型与 memory。","strength":"支持状态描述对离线规划的增益；没有验证可见状态相同、隐藏机制不同的干预后果。","advance":"将动作表示为状态变化，并显式预测中间状态。","premise":"论文明确承认：看不见的状态变化难以识别；LLM 描述质量构成瓶颈；使用 closed vocabulary。","assumption":"语言先验补全的状态不能自动视为个体场景的观测事实；在 XR 中应区分 inferred 与 observed state。","locator":"§3；§4；Appendix Limitations；Tables 4、9–12","taxonomy":{"task":"其他","task_note":"状态变化驱动的步骤规划","contribution_primary":"表示","contribution_secondary":["训练信号","记忆与上下文"],"control_signals":[],"target_properties":[],"property_note":"步骤与中间状态预测正确性","generation_paradigm":"不适用","paradigm_note":"","paper_type":"方法","contribution_focus":"动作前后状态文本与视觉 state embeddings 对齐；显式预测 intermediate states。","task_definition":{"inputs":"起终视觉状态；训练时提供步骤与 LLM 状态描述。","outputs":"中间状态表示及步骤序列。","invariants":"目标状态、任务动作含义与可行顺序。","judgement":"CrossTask/COIN/NIV 的 procedure planning 指标。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"三个 benchmark；CrossTask state alignment 及 mid-state prediction 消融，另测 Viterbi、描述类型与 memory。","one_sentence_advance":"将动作表示为状态变化，并显式预测中间状态。","source":"https://arxiv.org/html/2403.01599v1","locator":"§3；§4；Appendix Limitations；Tables 4、9–12"}},"STATE":{"key":"STATE","title":"What Changed and What Could Have Changed? State-Change Counterfactuals for Procedure-Aware Video Representation Learning","authors":["Chi-Hsi Kung","Frangil Ramirez","Juhyung Ha","Yi-Ting Chen","David Crandall","Yi-Hsuan Tsai"],"first_year":2025,"year":2025,"venue":"ICCV 2025","venue_source":"https://arxiv.org/abs/2503.21055","venue_proof":"arXiv journal reference","arxiv":"2503.21055","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2503.21055","full_source":"https://arxiv.org/html/2503.21055v6","version":"2503.21055v6","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2503.21055v6","origin":"库内主表","origin_detail":"VG04","question":"能否通过 state-change 描述及假想失败结果，学到更有 procedure awareness 的视频表征？","case":"Ego4D 预训练；GTEA 等 temporal action segmentation、EgoPER error detection、EPIC-KITCHENS retrieval、Charades-Ego recognition 等。","operation":"改变监督文本与动作序列：before/after 描述、state-change counterfactual、missing-step 与 misordered-step counterfactual。","method":"LLM 生成正负状态描述，进行 clip-level 与 video-level 对齐；对比 narration-only／Weak HierVL 及去掉不同 counterfactual 的设置。","evidence":"多项下游 benchmark＋组件消融；描述质量由 22 名评审评价 300 对 SC/CF，另用 Gemini 评 1000 对；人评 CF plausibility 低于 factual SC。","strength":"支持 counterfactual 文本监督改善表征；不能据此认定模型能生成正确的 counterfactual pixels、动力学或可执行教程。","advance":"把动作的状态变化与假想错误结果引入 procedure-aware representation learning。","premise":"用 Llama 3.1 8B 由 narration 产生监督；论文承认 CF 有时为低概率情形；视频级 CF 包括删步与重排。","assumption":"描述的 plausible negative 能用于区分表征，但不必等于同一事件的真实反事实；训练价值与因果真实性须分开检验。","locator":"§3；§4 Tables 2–5；Appendix Text Generation、Human Evaluation","taxonomy":{"task":"其他","task_note":"视频表征学习；counterfactual 是监督","contribution_primary":"训练信号","contribution_secondary":["表示"],"control_signals":[],"target_properties":[],"property_note":"表征的 procedure awareness 与下游正确率","generation_paradigm":"不适用","paradigm_note":"","paper_type":"方法","contribution_focus":"state-change 正例与 counterfactual 负例用于 clip/video-level 表征对齐。","task_definition":{"inputs":"教学视频 clips、narration 与生成的前后状态/counterfactual 描述。","outputs":"procedure-aware video embeddings 与下游任务预测。","invariants":"原始训练视频不必被编辑；监督改变，不意味着生成反事实视频。","judgement":"分割、错误检测、检索、识别 benchmark 和描述质量人评。"},"evidence_types":["定量 benchmark","消融","user study"],"classification_label":"推断","evidence_detail":"多项下游 benchmark＋组件消融；描述质量由 22 名评审评价 300 对 SC/CF，另用 Gemini 评 1000 对；人评 CF plausibility 低于 factual SC。","one_sentence_advance":"把动作的状态变化与假想错误结果引入 procedure-aware representation learning。","source":"https://arxiv.org/html/2503.21055v6","locator":"§3；§4 Tables 2–5；Appendix Text Generation、Human Evaluation"}},"VANS":{"key":"VANS","title":"Video-as-Answer: Predict and Generate Next Video Event with Joint-GRPO","authors":["Junhao Cheng","Liang Hou","Xin Tao","Jing Liao"],"first_year":2025,"year":2026,"venue":"CVPR 2026","venue_source":"https://openaccess.thecvf.com/content/CVPR2026/html/Cheng_Video-as-Answer_Predict_and_Generate_Next_Video_Event_with_Joint-GRPO_CVPR_2026_paper.html","venue_proof":"正式 CVF proceedings","arxiv":"2511.16669","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2511.16669","full_source":"https://arxiv.org/html/2511.16669v2","version":"2511.16669v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2511.16669v2","origin":"库内主表","origin_detail":"VG31","question":"如何把 next-event prediction 的答案从文本扩展为符合输入情境的动态视频？","case":"Video-Next-Event Prediction；VANS-Data-100K；procedural 与 predictive benchmarks。","operation":"输入视频与 procedural/predictive question；VLM 输出可视化 caption，VDM 生成下一事件；改变联合优化策略及奖励。","method":"Joint-GRPO 以共享奖励分阶段对齐 VLM 与 VDM；对比 SFT、只训练 VLM/VDM、独立 GRPO 及 all-in-one 联训。","evidence":"文本指标、FVD、CLIP-V/T；奖励和训练策略消融；30 名评审，每人评 20 个案例（procedural/predictive 各 10）。","strength":"支持该数据分布下下一事件答案与视觉一致性改善；文本/视觉相似度及满意度不足以证明干预因果正确性或学员学习收益。","advance":"联合优化“推断怎样回答”与“能否忠实生成该答案”，减小简单串联的跨模态错配。","premise":"有专门构建的 VNEP 训练数据及奖励；任务为输入情境之后的事件回答。","assumption":"下一事件参考及奖励具有足够代表性；现实分支可能多解，单一参考不等于可识别的反事实。","locator":"§3 Joint-GRPO；§4 Tables 1–3；Appendix Human Evaluation Table 5","taxonomy":{"task":"视频作为推理或回答","task_note":"","contribution_primary":"训练信号","contribution_secondary":["数据"],"control_signals":["文本","图像"],"target_properties":["实体一致","因果或状态一致","指令遵循"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"一次回答生成一个未来片段；不是逐 token streaming 的能力主张。","paper_type":"方法","contribution_focus":"Joint-GRPO 对齐 VLM caption 与 VDM video answer。","task_definition":{"inputs":"输入视频、procedural 或 predictive question。","outputs":"下一事件的文本描述及动态视频答案。","invariants":"输入情境中的对象/外观与问题约束；答案须接续已有情境。","judgement":"文本正确性指标、FVD/CLIP-V/T 与答案质量人评。"},"evidence_types":["定量 benchmark","消融","user study"],"classification_label":"推断","evidence_detail":"文本指标、FVD、CLIP-V/T；奖励和训练策略消融；30 名评审，每人评 20 个案例（procedural/predictive 各 10）。","one_sentence_advance":"联合优化“推断怎样回答”与“能否忠实生成该答案”，减小简单串联的跨模态错配。","source":"https://arxiv.org/html/2511.16669v2","locator":"§3 Joint-GRPO；§4 Tables 1–3；Appendix Human Evaluation Table 5"}},"DSCM":{"key":"DSCM","title":"Deep Structural Causal Models for Tractable Counterfactual Inference","authors":["Nick Pawlowski","Daniel C. Castro","Ben Glocker"],"first_year":2020,"year":2020,"venue":"NeurIPS 2020","venue_source":"https://arxiv.org/abs/2006.06485","venue_proof":"作者 arXiv comments 明示 accepted","arxiv":"2006.06485","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2006.06485","full_source":"https://arxiv.org/html/2006.06485v2","version":"2006.06485v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2006.06485v2","origin":"既有补充参考","origin_detail":"旧 topics 数据已有，未列入 papers.json 主表","question":"如何让深度模型支持 association、intervention 与个体 counterfactual inference，并可推断 exogenous noise？","case":"合成 Morpho-MNIST 与 UK Biobank brain MRI；不是视频生成实验。","operation":"干预 SCM 中的属性变量；abduction 保留该样本的外生变量，再替换机制并预测。","method":"用 normalizing flows 或 amortized variational inference 参数化 structural mechanisms；相较仅条件生成，明确实现外生噪声后验与 abduction–action–prediction。","evidence":"合成数据可核查干预属性及反事实；MRI 观察与可视化实验；结构与推断公式。","strength":"给出“同一个体反事实”可执行建模方式；不是从任意 observational video 唯一识别因果图或真实反事实的定理。","advance":"将高维影像纳入可进行外生噪声推断的 SCM。","premise":"acyclic structural assignments；独立外生噪声；指定的 causal graph；flow 对噪声可逆，或采用近似后验。","assumption":"图与机制的正确性需要外部依据；多个模型同样拟合观察时，增加像素拟合能力不能消除反事实不可识别性。","locator":"§2 SCM；§3 Deep mechanisms；§4 Experiments；Appendices B–C","taxonomy":{"task":"其他","task_note":"图像上的结构因果反事实推断","contribution_primary":"表示","contribution_secondary":["架构与生成范式"],"control_signals":[],"target_properties":[],"property_note":"SCM 下的个体反事实一致性；非视频性质","generation_paradigm":"不适用","paradigm_note":"","paper_type":"方法","contribution_focus":"SCM structural mechanisms 与 exogenous-noise posterior，使个体反事实可计算。","task_definition":{"inputs":"观测影像/属性、指定 causal graph 与属性干预。","outputs":"干预后属性及个体 counterfactual image。","invariants":"同一个体 abduction 得到的外生因素；替换指定机制而非重抽全部样本。","judgement":"合成属性/反事实实验及 MRI 影像结果；不是视频 benchmark。"},"evidence_types":["定量 benchmark","定性展示"],"classification_label":"推断","evidence_detail":"合成数据可核查干预属性及反事实；MRI 观察与可视化实验；结构与推断公式。","one_sentence_advance":"将高维影像纳入可进行外生噪声推断的 SCM。","source":"https://arxiv.org/html/2006.06485v2","locator":"§2 SCM；§3 Deep mechanisms；§4 Experiments；Appendices B–C"}},"OMNI":{"key":"OMNI","title":"Generative Omnimatte: Learning to Decompose Video into Layers","authors":["Yao-Chih Lee","Erika Lu","Sarah Rumbley","Michal Geyer","Jia-Bin Huang","Tali Dekel","Forrester Cole"],"first_year":2024,"year":2025,"venue":"CVPR 2025","venue_source":"https://gen-omnimatte.github.io/","venue_proof":"作者项目页与 CVF proceedings","arxiv":"2411.16683","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2411.16683","full_source":"https://arxiv.org/html/2411.16683v2","version":"2411.16683v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2411.16683v2","origin":"本次库外新找到","origin_detail":"对照旧 topics 主表与补充 JSON，未发现相同 ID/完整题名","question":"如何把自然视频分解为对象及其 shadows/reflections 等关联效果的独立层，并补全动态遮挡区域？","case":"casually captured videos；10 个合成场景的 background-removal benchmark；real-video qualitative examples。","operation":"输入视频与 object masks；trimask 指定 remove、preserve、uncertain effects 三类区域。","method":"微调视频 inpainting prior 得到 Casper，先生成 clean background/solo videos，再分解层；相较 reconstructive Omnimatte 无需静态背景、相机 pose 或 depth。","evidence":"10 场景 PSNR/LPIPS；数据和 mask conditioning 消融；真实视频定性。论文提醒小 benchmark 未充分体现 trimask 收益。","strength":"支持对象与可见关联后效分离；不能外推至删掉支撑物后的新运动，也不能把自动 effect association 当作 VOID 独有贡献。","advance":"引入生成先验以补全动态遮挡，摆脱前作背景/重建约束。","premise":"需目标 masks；训练数据未覆盖 physical deformation；原文显示弯曲物体等效果去除失败，动态背景可能被误关联。","assumption":"学到的 effects association 不等于 causal descendants；对不该改的动态背景需要独立保存约束。","locator":"§3 Fig.4 trimask；§4 Table 1；§5 Discussion and Limitations","taxonomy":{"task":"编辑与修改","task_note":"","contribution_primary":"训练信号","contribution_secondary":["条件注入","数据"],"control_signals":["图像","姿态或结构图"],"target_properties":["画质","实体一致","短时连贯"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"","paper_type":"方法","contribution_focus":"微调 Casper object-effect removal；trimask 区分删除、保留与不确定后效。","task_definition":{"inputs":"源视频、对象 masks；结构图指 trimask。","outputs":"clean background、solo videos 与 RGBA object/effect layers。","invariants":"保留对象与不应被删除的内容；去除目标及关联 effects。","judgement":"10 个合成场景 PSNR/LPIPS、mask/data 消融与真实分层结果。"},"evidence_types":["定量 benchmark","消融","定性展示"],"classification_label":"推断","evidence_detail":"10 场景 PSNR/LPIPS；数据和 mask conditioning 消融；真实视频定性。论文提醒小 benchmark 未充分体现 trimask 收益。","one_sentence_advance":"引入生成先验以补全动态遮挡，摆脱前作背景/重建约束。","source":"https://arxiv.org/html/2411.16683v2","locator":"§3 Fig.4 trimask；§4 Table 1；§5 Discussion and Limitations"}},"CWMDT":{"key":"CWMDT","title":"Counterfactual World Models via Digital Twin-conditioned Video Diffusion","authors":["Yiqing Shen","Aiza Maksutova","Chenjia Li","Mathias Unberath"],"first_year":2025,"year":2025,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2511.17481","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2511.17481","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2511.17481","full_source":"https://arxiv.org/html/2511.17481v1","version":"2511.17481v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2511.17481v1","origin":"库内主表","origin_detail":"P071","question":"如何在 world model 中把干预作为显式输入，而非仅从 factual pixels 做 forward simulation？","case":"RVEBench、FiVE；另在 CausalVQA debug split 测试生成辅助推理。","operation":"把观测转换为对象与关系的 structured-text digital twin；干预场景属性并传播到后续时刻。","method":"perception→LLM intervention reasoning→conditioned video diffusion；消融 digital twin、reasoning mode、LLM size 和 modified initial frame。","evidence":"两个编辑 benchmark 的 CLIP、GroundingDINO、LLM judge 等指标与消融；CausalVQA 额外实验的人类基线为 5 名标注者。","strength":"支持显式结构作为编辑控制的实证价值；名称 digital twin 不自动意味着精确物理模拟，指标也不是个体动力学真值。","advance":"在结构化对象/关系表示中实施干预，再将结果送入视频生成。","premise":"依赖 perception 构建 twin、LLM 推理、修改后的初帧与 diffusion 条件化。","assumption":"遗漏的隐藏关系无法通过文本图自动恢复；若结构错误，后续生成可能稳定地实现错误因果故事。","locator":"§3 Framework；§4 Tables 2–3；CausalVQA additional evaluation","taxonomy":{"task":"编辑与修改","task_note":"","contribution_primary":"表示","contribution_secondary":["推理时方法"],"control_signals":["文本","图像","姿态或结构图"],"target_properties":["因果或状态一致","指令遵循","短时连贯"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"按干预请求生成序列；不是已核实的实时 action loop。","paper_type":"方法","contribution_focus":"structured-text digital twin 表示对象/关系，先传播干预再条件化生成。","task_definition":{"inputs":"观测视频/场景与 counterfactual intervention。","outputs":"干预条件下的后续视觉序列。","invariants":"未被干预及其后果改变的场景事实；保持与源观察的关联。","judgement":"RVEBench/FiVE 的语义、grounding、judge 等指标，另测 CausalVQA。"},"evidence_types":["定量 benchmark","消融","user study"],"classification_label":"推断","evidence_detail":"两个编辑 benchmark 的 CLIP、GroundingDINO、LLM judge 等指标与消融；CausalVQA 额外实验的人类基线为 5 名标注者。","one_sentence_advance":"在结构化对象/关系表示中实施干预，再将结果送入视频生成。","source":"https://arxiv.org/html/2511.17481v1","locator":"§3 Framework；§4 Tables 2–3；CausalVQA additional evaluation"}},"VOID":{"key":"VOID","title":"VOID: Video Object and Interaction Deletion","authors":["Saman Motamed","William Harvey","Benjamin Klein","Luc Van Gool","Zhuoning Yuan","Ta-Ying Cheng"],"first_year":2026,"year":2026,"venue":"ECCV 2026","venue_source":"https://void-model.github.io/","venue_proof":"2026-10-15 作者项目页明确标注 ECCV 2026；正文版本仍沿用原拆解记录。","arxiv":"2604.02296","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2604.02296","full_source":"https://arxiv.org/html/2604.02296v1","version":"2604.02296v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2604.02296v1","origin":"库内主表","origin_detail":"VG13","question":"删除与场景发生物理交互的对象后，如何同时修正下游互动，保留未受影响内容？","case":"Kubric 刚体与 HUMOTO 人体交互构造 paired removal data；75 条真实视频评估。","operation":"用户点选删除对象；VLM 扩展 interaction-aware quadmask；VDM 生成新轨迹，可选第二遍用 flow-warped noise 稳定形状。","method":"从 Generative Omnimatte 的 CogVideoX 权重初始化，进行 paired counterfactual training＋quadmask＋两遍生成；相较前作 trimask 和 photometric-effects 数据，引入动态互动后果监督。","evidence":"合成指标、数据/mask/VLM/refiner 消融；真实视频无 GT：25 人每人 5 案例，VOID 得 64.8% 选择率；另有多 VLM judge。","strength":"支持所测删除场景下更受偏好且更合理的后果；没有证明 VLM 精确解出复杂轨迹，也没有真实视频个体 counterfactual GT。","advance":"把 object removal 从背景和阴影修补推进到需要改写其它物体后续运动的删除。","premise":"操作以删除为主；由 VLM 决定需放开修改的区域；训练配对由模拟/动作数据构造；真实评测依赖偏好与 judge。","assumption":"有效性依赖影响区域召回、生成 prior 与训练后果类型共同作用；“自动推 mask”不足以区别既有 effect-aware 流程。未见物体被漏标时，保真约束可能保留错误后果。","locator":"§3 Fig.3、dataset/quadmask/refiner；§4 Tables 1–4；Appendix mask reasoner Table 5","taxonomy":{"task":"编辑与修改","task_note":"","contribution_primary":"训练信号","contribution_secondary":["条件注入","数据","推理时方法"],"control_signals":["图像","姿态或结构图","轨迹或光流"],"target_properties":["因果或状态一致","物理合理","实体一致","短时连贯"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"","paper_type":"方法","contribution_focus":"paired removal supervision＋quadmask；第二遍用 flow-warped noise 稳定新运动。","task_definition":{"inputs":"源视频与点选删除目标；quadmask 指明影响区，第二遍使用预测光流。","outputs":"删除目标并重写下游互动的视频。","invariants":"未受影响的对象、背景和动态；删除目标导致的后果允许变化。","judgement":"合成评测与消融；真实数据用人类偏好和 VLM judge，没有真实 counterfactual GT。"},"evidence_types":["定量 benchmark","消融","user study","定性展示"],"classification_label":"推断","evidence_detail":"合成指标、数据/mask/VLM/refiner 消融；真实视频无 GT：25 人每人 5 案例，VOID 得 64.8% 选择率；另有多 VLM judge。","one_sentence_advance":"把 object removal 从背景和阴影修补推进到需要改写其它物体后续运动的删除。","source":"https://arxiv.org/html/2604.02296v1","locator":"§3 Fig.3、dataset/quadmask/refiner；§4 Tables 1–4；Appendix mask reasoner Table 5"}},"GIVE":{"key":"GIVE","title":"Geometry-Instructed Video Editing","authors":["Chirui Chang","Xiaoyang Lyu","Yi-Hua Huang","Haoru Tan","Shizhen Zhao","Yikang Ding","Jianmin Bao","Xin Tao","Pengfei Wan","Xiaojuan Qi"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2606.24225","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2606.24225","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2606.24225","full_source":"https://arxiv.org/html/2606.24225v1","version":"2606.24225v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2606.24225v1","origin":"库内主表","origin_detail":"VG46","question":"如何无歧义地指定跨视角、跨时间的对象几何修改，并一致更新 shadows/reflections？","case":"Unreal Engine GIVE-Data/GIVE-Bench；translate、rotate、scale、duplicate、remove；真实视频转移。","operation":"编辑前/后 depth-box 与 orientation-box 两路几何指令；固定非编辑因素生成配对监督。","method":"统一 object-state transition 条件输入 diffusion editor；与 text/mask/track 控制及依赖逐视频重建的方法比较。","evidence":"engine-rendered paired benchmark、算子与条件消融；30 条 Pexels 视频有 copy-paste 构造的近似 removal reference；其余真实几何结果主要定性。","strength":"支持几何算子执行及几何依赖外观后效；copy-paste reference 不能充当真实干预后的动力学真值。","advance":"用轻量 pre/post geometry 编码替代易失效的逐视频完整几何重建，并统一多个算子。","premise":"训练/基准的受控配对来自 engine；需几何指令；作者展示新暴露的运动手部细节失败，多路 tokens 增加计算。","assumption":"预先给定对象目标几何不等于由物理机制求出所有 downstream trajectories；不能直接作为任意事件干预的解法。","locator":"§3 Geometry instruction/data；§4；Supplement Table S3、Fig.S7","taxonomy":{"task":"编辑与修改","task_note":"","contribution_primary":"条件注入","contribution_secondary":["表示","数据"],"control_signals":["图像","姿态或结构图"],"target_properties":["几何一致","实体一致","短时连贯","指令遵循"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"","paper_type":"方法","contribution_focus":"pre/post depth-box＋orientation-box 两路 geometry instruction；统一多种几何算子。","task_definition":{"inputs":"源视频、目标对象编辑前/后的几何 box streams。","outputs":"平移/旋转/缩放/复制/删除后的编辑视频。","invariants":"非目标内容、时间连贯和未改场景；几何相关阴影/反射须随编辑变化。","judgement":"GIVE-Bench 配对参考、算子/条件消融；30 条真实 removal 的近似参考。"},"evidence_types":["定量 benchmark","消融","定性展示"],"classification_label":"推断","evidence_detail":"engine-rendered paired benchmark、算子与条件消融；30 条 Pexels 视频有 copy-paste 构造的近似 removal reference；其余真实几何结果主要定性。","one_sentence_advance":"用轻量 pre/post geometry 编码替代易失效的逐视频完整几何重建，并统一多个算子。","source":"https://arxiv.org/html/2606.24225v1","locator":"§3 Geometry instruction/data；§4；Supplement Table S3、Fig.S7"}},"PHYS":{"key":"PHYS","title":"VideoPhysEdit: Physical Counterfactual Video Editing via Rigid-Body Physical Scene Reconstruction","authors":["Conghan Yue","Yuanjie Chen","Yue Han","Ya Gao","Yunyan Xiao","WeiYao Zhang","Zhineng Chen"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2609.35134","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2609.35134","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2609.35134","full_source":"https://arxiv.org/html/2609.35134v1","version":"2609.35134v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2609.35134v1","origin":"既有补充参考","origin_detail":"旧 topics 数据已有，未列入 papers.json 主表","question":"源视频在指定帧发生物理编辑后，如何生成其导致的新运动与交互？","case":"PCVE-RigidBench：20 个源场景、129 个任务；真实视频额外定性；Add/Delete/Set 物理编辑。","operation":"给定源视频、执行帧、物理编辑；重建物体状态/物理参数，干预后模拟轨迹并用于生成。","method":"训练外 physical scene reconstruction＋rigid-body simulation＋trajectory-guided generation；相较直接编辑器使物理推理显式化。","evidence":"paired synthetic GT、PES/轨迹误差及消融。完整 129 任务 PES 0.376；附录另列成功产出视频的 116 任务 PES 0.418，不能混用分母。117 任务在首帧干预，仅 12 个在中途干预；真实视频为定性。","strength":"在此刚体基准上有比偏好更直接的轨迹证据；但生成覆盖率与较小的中途干预子集限制结论，不能把首帧结果当作所有定时编辑已充分验证。不是一般现实物理识别保证。","advance":"把定时物理属性/对象编辑与 downstream trajectory evaluation 接起来。","premise":"明确仅静态相机、平面支撑、sphere/box 几何；不覆盖 articulated/actively controlled humans or robots；阈值受尺度影响。","assumption":"可拟合原视频的物理参数不必唯一；隐藏质量/摩擦若未被激发，准确复现 factual motion 仍可给出不同 intervention outcome。","locator":"§3 Physical Scene Reconstruction；§4 PCVE-RigidBench；Appendix comparison with VOID、Limitations","taxonomy":{"task":"编辑与修改","task_note":"","contribution_primary":"推理时方法","contribution_secondary":["表示","评测"],"control_signals":["文本","图像","轨迹或光流","物理或模拟代理"],"target_properties":["因果或状态一致","物理合理","几何一致"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"","paper_type":"方法","contribution_focus":"重建可模拟的刚体场景，执行物理干预后用新轨迹指导生成。","task_definition":{"inputs":"源视频、物理编辑与执行帧；可包含数值或空间修改。","outputs":"干预后运动/交互改变的 counterfactual video。","invariants":"干预前历史及未受影响内容；执行指定 Add/Delete/Set。","judgement":"PCVE-RigidBench 的物理 GT、轨迹误差/PES；真实视频仅定性。"},"evidence_types":["定量 benchmark","消融","定性展示"],"classification_label":"推断","evidence_detail":"paired synthetic GT、PES/轨迹误差及消融。完整 129 任务 PES 0.376；附录另列成功产出视频的 116 任务 PES 0.418，不能混用分母。117 任务在首帧干预，仅 12 个在中途干预；真实视频为定性。","one_sentence_advance":"把定时物理属性/对象编辑与 downstream trajectory evaluation 接起来。","source":"https://arxiv.org/html/2609.35134v1","locator":"§3 Physical Scene Reconstruction；§4 PCVE-RigidBench；Appendix comparison with VOID、Limitations"}},"KNOW":{"key":"KNOW","title":"Robots That Ask For Help: Uncertainty Alignment for Large Language Model Planners","authors":["Allen Z. Ren","Anushri Dixit","Alexandra Bodrova","Sumeet Singh","Stephen Tu","Noah Brown","Peng Xu","Leila Takayama","Fei Xia","Jake Varley","Zhenjia Xu","Dorsa Sadigh","Andy Zeng","Anirudha Majumdar"],"first_year":2023,"year":2023,"venue":"CoRL 2023","venue_source":"https://arxiv.org/abs/2307.01928","venue_proof":"作者 arXiv comments：CoRL 2023 Oral","arxiv":"2307.01928","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2307.01928","full_source":"https://arxiv.org/html/2307.01928v2","version":"2307.01928v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2307.01928v2","origin":"本次库外新找到","origin_detail":"对照旧 topics 主表与补充 JSON，未发现相同 ID/完整题名","question":"LLM robot planner 何时知道自己不知道，并以尽量少的求助达到指定成功覆盖率？","case":"模拟及真实 tabletop rearrangement、bimanual、mobile manipulation；多种语言歧义。","operation":"从候选动作构造 conformal prediction set；非单元素集合向人求助；控制目标覆盖率。","method":"KnowNo 用 calibration set 校准 LLM 分数，不微调模型；对比 Simple Set、prompting 等求助策略。","evidence":"conformal coverage 推导＋模拟/硬件实验；论文报告硬件移动操作低层成功率约 86%，明确区分 plan 与 task success。","strength":"可支持条件成立时的统计覆盖与较少求助；不是单次实例零风险、任意分布漂移或最少感知成本保证。","advance":"将可校准不确定性与 ask-for-help 连接，而非只用语言置信表达。","premise":"正文假设 calibration/test scenarios 从同一分布 i.i.d. 抽样；可靠的人类帮助和低层控制；部分任务已知物体位置并用脚本动作。","assumption":"未列入候选的真实状态、错误人类反馈及执行风险需独立建模；XR 的“足够信息”必须关联后续生成/执行结果。","locator":"§3 Uncertainty Alignment；§4；Appendix Low-level control success rate","taxonomy":{"task":"其他","task_note":"机器人任务规划与校准求助","contribution_primary":"推理时方法","contribution_secondary":[],"control_signals":[],"target_properties":[],"property_note":"计划覆盖率、任务成功与求助成本","generation_paradigm":"不适用","paradigm_note":"","paper_type":"方法","contribution_focus":"conformal prediction set 校准：行动候选不唯一时求助。","task_definition":{"inputs":"语言任务、环境状态、候选动作与 calibration examples。","outputs":"执行的动作或向用户询问的候选集合。","invariants":"用户真实目标及 calibration/test 分布条件；不涉及像素保真。","judgement":"计划/执行成功率、求助负担及 coverage；理论保证的条件单列于原拆解。"},"evidence_types":["定量 benchmark"],"classification_label":"推断","evidence_detail":"conformal coverage 推导＋模拟/硬件实验；论文报告硬件移动操作低层成功率约 86%，明确区分 plan 与 task success。","one_sentence_advance":"将可校准不确定性与 ask-for-help 连接，而非只用语言置信表达。","source":"https://arxiv.org/html/2307.01928v2","locator":"§3 Uncertainty Alignment；§4；Appendix Low-level control success rate"}},"ACT":{"key":"ACT","title":"Learning to Clarify: Multi-turn Conversations with Action-Based Contrastive Self-Training","authors":["Maximillian Chen","Ruoxi Sun","Tomas Pfister","Sercan Ö. Arık"],"first_year":2024,"year":2025,"venue":"ICLR 2025","venue_source":"https://arxiv.org/abs/2406.00222","venue_proof":"作者 arXiv comments","arxiv":"2406.00222","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2406.00222","full_source":"https://arxiv.org/html/2406.00222v2","version":"2406.00222v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2406.00222v2","origin":"库内主表","origin_detail":"P149","question":"在对话数据有限时，如何让 LLM 学会先澄清而不是擅自猜测用户意图？","case":"PACIFIC tabular QA、machine reading comprehension、AmbigSQL。","operation":"改变隐式 dialogue action（如澄清/回答），比较其后续对话达成信息目标的效果。","method":"Action-Based Contrastive Self-Training：quasi-online DPO 对比不同行动及模拟后续轨迹；相较 SFT/标准 DPO 不只做同类回答偏好。","evidence":"多任务少样本实验、action/turn/trajectory 指标、消融与失败轨迹；非 XR 人机教学实验。","strength":"支持澄清策略学习；不提供对真实用户误答、主动视角获取或视频生成成本的保证。","advance":"以行动层面的反事实对比改善多轮 dialogue policy。","premise":"利用任务目标答案与对话数据构造/评价候选轨迹；后续互动含模拟成分。","assumption":"若 simulator 的反馈分布不代表真实学员，策略收益不能直接转移；语言答对也不等于视频操作正确。","locator":"§3 ACT；Experiments；Appendix Algorithms、Tables A28–A29","taxonomy":{"task":"其他","task_note":"多轮澄清对话策略","contribution_primary":"训练信号","contribution_secondary":[],"control_signals":[],"target_properties":[],"property_note":"对话决策与目标答案正确性","generation_paradigm":"不适用","paradigm_note":"","paper_type":"方法","contribution_focus":"以不同 dialogue actions 及其后续结果构造对比，进行 quasi-online DPO。","task_definition":{"inputs":"用户请求、表格/文档/SQL schema 与对话历史。","outputs":"澄清问题或最终回答。","invariants":"用户的信息目标与 grounded facts；澄清应解决歧义。","judgement":"action、turn、trajectory 指标及最终任务答案。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"多任务少样本实验、action/turn/trajectory 指标、消融与失败轨迹；非 XR 人机教学实验。","one_sentence_advance":"以行动层面的反事实对比改善多轮 dialogue policy。","source":"https://arxiv.org/html/2406.00222v2","locator":"§3 ACT；Experiments；Appendix Algorithms、Tables A28–A29"}},"GPS":{"key":"GPS","title":"GPS: Graph-guided Proactive Information Seeking in Large Language Models","authors":["Ruiqing Li","Yifeng Xu","Xinke Jiang","Zhibang Yang","Xinyu Ma","Yue Fang","Junfeng Zhao","Yasha Wang","Xu Chu"],"first_year":2026,"year":2026,"venue":"ICLR 2026","venue_source":"https://proceedings.iclr.cc/paper_files/paper/2026/hash/1f49b7fbd7c7afc52c1db0d4ed1a338d-Abstract-Conference.html","venue_proof":"正式 proceedings","arxiv":"","doi":[],"identifier_note":"[待核] arXiv ID／DOI 未找到；保留正式 proceedings 入口","source":"https://proceedings.iclr.cc/paper_files/paper/2026/hash/1f49b7fbd7c7afc52c1db0d4ed1a338d-Abstract-Conference.html","full_source":"https://proceedings.iclr.cc/paper_files/paper/2026/file/1f49b7fbd7c7afc52c1db0d4ed1a338d-Paper-Conference.pdf","version":"正式 proceedings PDF","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://proceedings.iclr.cc/paper_files/paper/2026/file/1f49b7fbd7c7afc52c1db0d4ed1a338d-Paper-Conference.pdf","origin":"库内主表","origin_detail":"P159","question":"RAG 查询缺少用户条件时，如何依据文档中的条件规则主动获取必要信息？","case":"Synthetic、ConditionalQA、ShARC 三个 QA benchmarks；规则型 RAG。","operation":"将文档规则转为 conditional DAG；依据用户回答动态剪枝，选择下一澄清问题。","method":"DAG reasoning＋traversal clarification；conditional-path synthetic data 与 accuracy/efficiency 混合 reward 的 RL；对比无显式条件结构的澄清。","evidence":"三个 benchmark 的成功率/clarification efficiency、消融；理论证明 Boolean 条件的 DAG 表达完备性。","strength":"表示完备不等于 LLM 抽取正确、检索文档完备或现实世界观测充分；也不等于一般最优 information gain。","advance":"把开放式提问收敛为依据条件依赖的查询与剪枝。","premise":"有限取值条件变量；答案由检索文档中的逻辑条件决定；采用 User-Simulator LLM；理论分析针对所定义 DAG。","assumption":"开放 XR 中“哪些变量重要”本身可能未知，因此不能假定已有完整条件图；先发现结构与在已知结构中问值是不同问题。","locator":"ICLR proceedings PDF pp.3–6；§5；Appendix theoretical analysis","taxonomy":{"task":"其他","task_note":"规则型 RAG 主动信息获取","contribution_primary":"表示","contribution_secondary":["训练信号","推理时方法"],"control_signals":[],"target_properties":[],"property_note":"答案成功率与澄清效率","generation_paradigm":"不适用","paradigm_note":"","paper_type":"方法","contribution_focus":"conditional DAG 表示缺失条件与依赖，随回答遍历/剪枝。","task_definition":{"inputs":"不充分的用户问题、检索规则文档、已有回答。","outputs":"逐步澄清问题与最终规则型答案。","invariants":"检索文档中的逻辑条件和用户已确认信息。","judgement":"Synthetic/ConditionalQA/ShARC 成功率与 Weighted Clarification Turns；理论性质见拆解。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"三个 benchmark 的成功率/clarification efficiency、消融；理论证明 Boolean 条件的 DAG 表达完备性。","one_sentence_advance":"把开放式提问收敛为依据条件依赖的查询与剪枝。","source":"https://proceedings.iclr.cc/paper_files/paper/2026/file/1f49b7fbd7c7afc52c1db0d4ed1a338d-Paper-Conference.pdf","locator":"ICLR proceedings PDF pp.3–6；§5；Appendix theoretical analysis"}},"QWEN":{"key":"QWEN","title":"Qwen-Image-Agent: Bridging the Context Gap in Real-World Image Generation","authors":["Zekai Zhang","Jiahao Li","Jie Zhang","Kaiyuan Gao","Kun Yan","Lihan Jiang","Ningyuan Tang","Shengming Yin","Tianhe Wu","Xiaoyue Chen","Xiao Xu","Yan Shu","Yanran Zhang","Yixian Xu","Yuxiang Chen","Zhendong Wang","Zihao Liu","Zikai Zhou","Huishuai Zhang","Dongyan Zhao","Chenfei Wu"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2606.26907","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2606.26907","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2606.26907","full_source":"https://arxiv.org/html/2606.26907v2","version":"2606.26907v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2606.26907v2","origin":"库内主表","origin_detail":"VG17","question":"真实 image generation 请求缺省、隐含或依赖最新信息时，如何弥合 user context 与 generation context 的差距？","case":"图像生成；IA-Bench 的 Plan/Reason/Search/Memory，Mindbench 与 WISE-Verified。","operation":"agent 在 plan、reason、search、memory、feedback 中组织所需上下文；改变 context 组成和 renderer。","method":"Context-Aware Planning＋Context Grounding，再生成/反馈；消融各类 context 与图像 backbone。","evidence":"三个 benchmark 与 context/backbone 消融；去掉对应 context 会降低相应能力维度。","strength":"支持获取/整合上下文改善图像生成；没有给出最小充分集合保证，也不是视频或 XR 主动感知结果。","advance":"把用户输入视为部分上下文，并显式构建生成所需条件。","premise":"依赖外部检索、记忆与图像生成/编辑工具；论文指出 renderer 本身仍会计数/参考跟随失败。","assumption":"正确 context 与 renderer 能力要分开归因；工具调用更丰富不必然意味着交互成本最优。","locator":"§3 Context-centric framework；Experiments Table 4；Conclusion","taxonomy":{"task":"其他","task_note":"图像生成 agent；不是视频生成","contribution_primary":"推理时方法","contribution_secondary":["记忆与上下文","评测"],"control_signals":["文本","图像","身份参考"],"target_properties":["指令遵循"],"property_note":"","generation_paradigm":"不适用","paradigm_note":"该生成范式维度针对视频，图像任务不套用。","paper_type":"方法","contribution_focus":"agent 规划并从 reasoning/search/memory/feedback 构建 generation context。","task_definition":{"inputs":"可能缺省或需最新知识的图像请求与可获取参考。","outputs":"补齐上下文后的生成/编辑图像。","invariants":"用户意图、已核实知识与参考约束；不是长视频状态保持。","judgement":"IA-Bench、Mindbench、WISE-Verified 及 context/renderer 消融。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"三个 benchmark 与 context/backbone 消融；去掉对应 context 会降低相应能力维度。","one_sentence_advance":"把用户输入视为部分上下文，并显式构建生成所需条件。","source":"https://arxiv.org/html/2606.26907v2","locator":"§3 Context-centric framework；Experiments Table 4；Conclusion"}},"VGEN":{"key":"VGEN","title":"VideoGen-Agent: Reinforcing Video Generation Agents","authors":["Binxu Li","Haoyi Duan","Yuhui Zhang","Yaohui Zhang","Zihao Lin","Kaituo Feng","Suozhi Huang","Xiangyi Li","Yu Li","Chunyuan Li","Shilong Liu","Mengdi Wang"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2609.24997","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2609.24997","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2609.24997","full_source":"https://arxiv.org/html/2609.24997v3","version":"2609.24997v3","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2609.24997v3","origin":"库内主表","origin_detail":"VG20","question":"能否训练共享 agent policy 使用外部工具，满足单一视频生成器难以实现的知识、身份、物理与事件顺序要求？","case":"VABench：600 prompts，六类各 100；多种 augmentation/generation/verification tools。","operation":"agent 依据 prompt 与中间观测选工具；改变 SFT/RL、reward components 与工具集。","method":"teacher trajectories SFT＋multitask agentic RL；相较 standalone generator 学习多轮工具编排与验证。","evidence":"同一基础工具集 56.5→75.6，升级工具 86.1；human preference 每项比较 100 对×4 人；另 judge agreement 600 对×3 人。","strength":"有 agent 训练与工具升级的分离比较；部分训练/测试共享 evaluator rubric，不能把分数直接等同独立物理真值或最少信息证明。","advance":"跨六类任务学习共享工具策略，并可受益于升级工具。","premise":"固定可调用工具接口与类别 reward；物理检查的 image-plane momentum 指标明确假设 fixed camera，不是校准的 3D 守恒检验。","assumption":"若评测只奖励生成质量，策略可能用更多工具换分；要论证 D 需报告质量—成本曲线并测试必要信息被移除后的退化。","locator":"§3 Training；§4 VABench/Human Evaluation；Appendix D、physics rubric","taxonomy":{"task":"条件控制生成","task_note":"","contribution_primary":"训练信号","contribution_secondary":["推理时方法","评测"],"control_signals":["文本","图像","身份参考","物理或模拟代理"],"target_properties":["指令遵循","实体一致","物理合理","因果或状态一致"],"property_note":"","generation_paradigm":"不适用","paradigm_note":"不适用统一范式：agent 编排多种 renderer；原文未把所有工具限定为同一种离线/因果/少步架构。","paper_type":"方法","contribution_focus":"SFT＋agentic RL 学习工具选择；主落点是 agent policy，不是新视频 backbone。","task_definition":{"inputs":"复杂视频 prompt；可调用增强、生成、验证工具及中间结果。","outputs":"通过多轮工具调用生成的视频。","invariants":"各任务的身份、物理、空间和事件顺序约束。","judgement":"VABench 六类 rubric、训练/工具集对照与人类偏好。"},"evidence_types":["定量 benchmark","消融","user study"],"classification_label":"推断","evidence_detail":"同一基础工具集 56.5→75.6，升级工具 86.1；human preference 每项比较 100 对×4 人；另 judge agreement 600 对×3 人。","one_sentence_advance":"跨六类任务学习共享工具策略，并可受益于升级工具。","source":"https://arxiv.org/html/2609.24997v3","locator":"§3 Training；§4 VABench/Human Evaluation；Appendix D、physics rubric"}},"IVR":{"key":"IVR","title":"Interactive Visual Reasoning under Uncertainty","authors":["Manjie Xu","Guangyuan Jiang","Wei Liang","Chi Zhang","Yixin Zhu"],"first_year":2022,"year":2023,"venue":"NeurIPS 2023 Datasets and Benchmarks","venue_source":"https://arxiv.org/abs/2206.09203","venue_proof":"作者 arXiv comments 明示 accepted；不是 2022 venue","arxiv":"2206.09203","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2206.09203","full_source":"https://arxiv.org/html/2206.09203v2","version":"2206.09203v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2206.09203v2","origin":"库内主表","origin_detail":"P143","question":"面对含糊的 action–effect 关系，agent 能否设计实验识别对象的因果作用？","case":"Interactive Visual Reasoning / IVRE；CLEVR 风格 Blicket detector 环境。","operation":"给定不充分 context，主动选择物体组合做试验，更新 Blicket belief；有最多 trial 数及效率惩罚。","method":"构建交互评测平台并比较 RL、启发式、LLM 与人类；区别于被动看完整样本再答题。","evidence":"visual/symbolic 输入控制；算法比较；54 名参与者各完成一个随机 episode，最多 10 步。","strength":"可隔离主动实验与视觉感知难度；有限规则环境的因果发现证据不能外推到自然视频生成。","advance":"把 uncertainty resolution 作为闭环可操作任务，而非只读既有观测。","premise":"生成器规定对象与 detector 规则；可安全、重复执行候选 trial；目标和反馈可评价。","assumption":"现实 XR 的试验可能改变环境且不可逆；需要把信息价值与操作后果成本一同定义。","locator":"Task/environment；Experiments Human study；Appendices environment and human examples","taxonomy":{"task":"其他","task_note":"主动视觉因果实验环境","contribution_primary":"评测","contribution_secondary":["数据"],"control_signals":[],"target_properties":[],"property_note":"因果识别准确率与试验代价","generation_paradigm":"不适用","paradigm_note":"","paper_type":"基准","contribution_focus":"可主动试验的 Blicket 环境，分别测视觉感知与因果识别。","task_definition":{"inputs":"不充分的物体—detector 观察与可用实验动作。","outputs":"下一试验与对象 Blicketness belief。","invariants":"当前 episode 的隐藏 detector 机制；试验要识别它而非改写规则。","judgement":"belief 正确性、试验效率、visual/symbolic 对照与人类 baseline。"},"evidence_types":["定量 benchmark","user study"],"classification_label":"推断","evidence_detail":"visual/symbolic 输入控制；算法比较；54 名参与者各完成一个随机 episode，最多 10 步。","one_sentence_advance":"把 uncertainty resolution 作为闭环可操作任务，而非只读既有观测。","source":"https://arxiv.org/html/2206.09203v2","locator":"Task/environment；Experiments Human study；Appendices environment and human examples"}},"GENEX":{"key":"GENEX","title":"Generative World Explorer","authors":["Taiming Lu","Tianmin Shu","Alan Yuille","Daniel Khashabi","Jieneng Chen"],"first_year":2024,"year":2025,"venue":"ICLR 2025","venue_source":"https://proceedings.iclr.cc/paper_files/paper/2025/file/8204f54f6cd8f0db1765a274e94ceaac-Paper-Conference.pdf","venue_proof":"正式 proceedings","arxiv":"2411.11844","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2411.11844","full_source":"https://arxiv.org/html/2411.11844v3","version":"2411.11844v3","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2411.11844v3","origin":"既有补充参考","origin_detail":"旧 topics 数据已有，未列入 papers.json 主表","question":"部分观察下，能否用 imagined observations 更新 belief 并帮助 embodied planning？","case":"GenEx-DB synthetic urban scenes、GenEx-EQA；360° 环境探索和决策；真实街景/室内额外泛化。","operation":"由当前 egocentric view 生成想象中的探索观察，交给决策模型；改变探索轨迹与可用视角。","method":"训练 generative world explorer，将生成的全景观察用于 belief/decision；对比仅文本/当前视角等输入条件。","evidence":"生成质量、loop consistency、探索与规划评测；人类评估说明见附录，但[待核] 本次未核实总人数。","strength":"支持所测任务里想象有辅助价值；imagined evidence 不是实际测量，不能证明未知事实被观测确认。","advance":"将可连续探索的生成视觉用于部分可见场景的决策辅助。","premise":"生成模型由合成城市数据训练；决策可以使用生成观察；本文读取正文为 v3，摘要页与正文的表述存在版本差异。","assumption":"若把 imagined observation 与 sensed observation 混存，会循环强化先验；XR 纠错必须保留来源与置信度。","locator":"正文 v3 Abstract、Experiments Findings；Appendix Human Evaluation","taxonomy":{"task":"动作条件的预测与交互世界模型","task_note":"","contribution_primary":"架构与生成范式","contribution_secondary":["表示","数据","评测"],"control_signals":["图像","动作"],"target_properties":["几何一致","环境记忆","因果或状态一致"],"property_note":"","generation_paradigm":"因果自回归或流式","paradigm_note":"按探索动作连续生成；此标签不意味着已达到实时速度。","paper_type":"方法","contribution_focus":"panoramic generative explorer 预测想象中的观察，用于 belief/decision 更新。","task_definition":{"inputs":"当前 egocentric view、预期移动/探索方向。","outputs":"未来全景观察及基于想象的决策。","invariants":"探索中的场景结构与回环一致性；想象不是已证实的外部事实。","judgement":"生成质量、loop consistency、GenEx-EQA decision accuracy 与 human comparisons。"},"evidence_types":["定量 benchmark","user study"],"classification_label":"推断","evidence_detail":"生成质量、loop consistency、探索与规划评测；人类评估说明见附录，但[待核] 本次未核实总人数。","one_sentence_advance":"将可连续探索的生成视觉用于部分可见场景的决策辅助。","source":"https://arxiv.org/html/2411.11844v3","locator":"正文 v3 Abstract、Experiments Findings；Appendix Human Evaluation"}},"S3":{"key":"S3","title":"See, Remember, Explore: A Benchmark and Baselines for Streaming Spatial Reasoning","authors":["Yuxi Wei","Wei Huang","Qirui Chen","Lu Hou","Xiaojuan Qi"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2603.23864","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2603.23864","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2603.23864","full_source":"https://arxiv.org/html/2603.23864v1","version":"2603.23864v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2603.23864v1","origin":"库内主表","origin_detail":"VG60","question":"如何在只用当前时刻之前的观察、有限计算下持续进行空间推理，并主动获取缺失视野？","case":"S3-Bench：10K+ scenes、26K+ trajectories；模拟可控探索与真实 streaming videos。","operation":"timestamp-grounded QA；memory folding；显式 move/rotate/scan 探索动作。","method":"AMF-VLM 压缩结构化 memory 并学习 active exploration；对比用相同训练数据的模型及模块消融。","evidence":"S3-Eval simulated/real splits、其他空间 benchmark transfer；同数据比较报告 8.8%/13.3% 改善；详细模块消融。","strength":"直接支持 streaming spatial reasoning 的问题定位；真实流视频部分不自动等于在线真人 XR 探索的随机试验，更未测教程生成。","advance":"把离线后验空间 QA 改为 streaming、过去信息限定、可主动探索的任务。","premise":"标注来自模拟 GT 或真实数据 3D annotation 的投影/遮挡计算；输入和问题有时间边界。","assumption":"提升空间 QA 不一定提高最终教程；有些观测减少几何不确定性，却不改变操作或生成决策。","locator":"§3 S3-Bench；§4 AMF-VLM；§5；Appendix visibility/QA generation、Tables 13–16","taxonomy":{"task":"其他","task_note":"流式空间推理与主动感知；不生成视频","contribution_primary":"评测","contribution_secondary":["数据","记忆与上下文","训练信号"],"control_signals":[],"target_properties":[],"property_note":"空间 QA 正确率、记忆与主动感知效能","generation_paradigm":"不适用","paradigm_note":"streaming 指输入/推理方式，不是视频生成范式。","paper_type":"基准","contribution_focus":"streaming spatial QA＋active exploration；AMF-VLM 提供 memory folding baseline。","task_definition":{"inputs":"截至查询时间的视觉流、空间问题及探索接口。","outputs":"空间答案及 move/rotate/scan 等取证动作。","invariants":"不能使用查询时间之后的观察；仍有效的空间证据应被保留。","judgement":"S3-Eval simulated/real accuracy、同数据 baseline 与模块消融。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"S3-Eval simulated/real splits、其他空间 benchmark transfer；同数据比较报告 8.8%/13.3% 改善；详细模块消融。","one_sentence_advance":"把离线后验空间 QA 改为 streaming、过去信息限定、可主动探索的任务。","source":"https://arxiv.org/html/2603.23864v1","locator":"§3 S3-Bench；§4 AMF-VLM；§5；Appendix visibility/QA generation、Tables 13–16"}},"BLACK":{"key":"BLACK","title":"Black Swan: Abductive and Defeasible Video Reasoning in Unpredictable Events","authors":["Aditya Chinchure","Sahithya Ravi","Raymond Ng","Vered Shwartz","Boyang Li","Leonid Sigal"],"first_year":2024,"year":2025,"venue":"CVPR 2025","venue_source":"https://arxiv.org/abs/2412.05725","venue_proof":"作者 arXiv comments","arxiv":"2412.05725","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2412.05725","full_source":"https://arxiv.org/html/2412.05725v2","version":"2412.05725v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2412.05725v2","origin":"库内主表","origin_detail":"P021","question":"VLM 能否解释意外事件，并在新视觉证据出现后推翻先前假设？","case":"BlackSwanSuite：1655 videos；Forecaster、Detective、Reporter，MCQ/generative/yes-no tasks。","operation":"人为限制可见片段，随后提供新信息；要求 abductive 与 defeasible responses。","method":"构造不可预期事件 benchmark，评测多种 VLM；相较常见可预测事件 QA 强调隐藏事件与证据改变解释。","evidence":"15,469 问的多模型评测。Appendix D 写明筛选后 32 名生成解释评审；F.1 的判别题 human upper bound 为 2 名实验室学生，MCQ 各 100 题、Y/N 各 150 题，取较高分。正文 §5.2 简述为一位专家每 variant 150 题：[待核] 两处口径不一致。","strength":"支持被测模型存在可撤销推理缺口；最高约 32 点差距针对该 benchmark 及人类上界设置，不是总体人群差距。人类基线取两人较高值、正文/附录样本口径不一致，应降低外推强度。不是视频生成纠错测试。","advance":"将 visual evidence arrival 与 hypothesis revision 显式纳入视频推理评估。","premise":"来自 Oops! test videos；作者承认潜在训练接触与开放式生成评审主观性。","assumption":"从能改文字答案到能使生成 memory、latent context 和未来 pixels 同步纠错，仍需另设实验。","locator":"§3 BlackSwanSuite；Experiments；Appendix Evaluation Metrics、Data Source","taxonomy":{"task":"其他","task_note":"视频理解与可撤销推理；不是视频形式的回答","contribution_primary":"评测","contribution_secondary":["数据"],"control_signals":[],"target_properties":[],"property_note":"解释正确性与证据响应能力","generation_paradigm":"不适用","paradigm_note":"","paper_type":"基准","contribution_focus":"用逐步揭示的视频证据测试 abductive/defeasible reasoning。","task_definition":{"inputs":"受限的视频片段、假设或问题；后续可增加新视觉证据。","outputs":"事件解释、MCQ 选择或假设是否成立的判断。","invariants":"已见真实证据不能被改写；解释可因新证据而撤回。","judgement":"MCQ/Y-N accuracy、生成解释评分与人类上界；人评口径冲突见原证据。"},"evidence_types":["定量 benchmark","user study"],"classification_label":"推断","evidence_detail":"15,469 问的多模型评测。Appendix D 写明筛选后 32 名生成解释评审；F.1 的判别题 human upper bound 为 2 名实验室学生，MCQ 各 100 题、Y/N 各 150 题，取较高分。正文 §5.2 简述为一位专家每 variant 150 题：[待核] 两处口径不一致。","one_sentence_advance":"将 visual evidence arrival 与 hypothesis revision 显式纳入视频推理评估。","source":"https://arxiv.org/html/2412.05725v2","locator":"§3 BlackSwanSuite；Experiments；Appendix Evaluation Metrics、Data Source"}},"STREAM":{"key":"STREAM","title":"StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text","authors":["Roberto Henschel","Levon Khachatryan","Hayk Poghosyan","Daniil Hayrapetyan","Vahram Tadevosyan","Zhangyang Wang","Shant Navasardyan","Humphrey Shi"],"first_year":2024,"year":2025,"venue":"CVPR 2025","venue_source":"https://openaccess.thecvf.com/content/CVPR2025/html/Henschel_StreamingT2V_Consistent_Dynamic_and_Extendable_Long_Video_Generation_from_Text_CVPR_2025_paper.html","venue_proof":"正式 CVF proceedings","arxiv":"2403.14773","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2403.14773","full_source":"https://arxiv.org/html/2403.14773v2","version":"2403.14773v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2403.14773v2","origin":"本次库外新找到","origin_detail":"对照旧 topics 主表与补充 JSON，未发现相同 ID/完整题名","question":"如何让短视频 diffusion 延长为连贯长视频，同时避免 hard cuts 和运动停滞？","case":"50 个动作/对象/场景 prompts；80 到 1200+ frames 的生成。","operation":"当前 chunk 接收前一 chunk 特征和第一 chunk appearance；改变 memory 模块、增强策略与长度。","method":"CAM short-term memory＋APM first-chunk appearance＋randomized blending enhancer；对比 autoregressive I2V、SparseControl、FreeNoise 等。","evidence":"一致性、motion/CLIP 等自动指标及消融；pairwise user study 报告偏好，[待核] 本次未核实人数。","strength":"支持长视频延展与运动保持；保持初始外观是目标，不是识别初始记忆错误并撤回。","advance":"联合短期衔接与长期外观锚定，缓解长生成停滞。","premise":"第一 chunk 提供长期外观；上一 chunk 提供局部连续条件。","assumption":"当第一 chunk 本身错误或过时，强锚定可能固化错误；需从“保持一致”转向“与有效证据一致”。","locator":"§3–4 memory/enhancement；§5；Appendix user study Fig.6","taxonomy":{"task":"长视频与多镜头叙事","task_note":"","contribution_primary":"记忆与上下文","contribution_secondary":["架构与生成范式","推理时方法"],"control_signals":["文本","图像"],"target_properties":["画质","短时连贯","实体一致"],"property_note":"","generation_paradigm":"因果自回归或流式","paradigm_note":"chunk-level autoregressive；不等于单帧 causal attention。","paper_type":"方法","contribution_focus":"前一 chunk 的 CAM＋首段 appearance anchor；blending 支持长程增强。","task_definition":{"inputs":"文本 prompt 与已生成 chunks。","outputs":"连贯且保持运动的长视频。","invariants":"初始场景/对象外观与相邻 chunk 衔接。","judgement":"50 prompts 上 motion/consistency/CLIP 指标、消融与 pairwise 人评。"},"evidence_types":["定量 benchmark","消融","user study"],"classification_label":"推断","evidence_detail":"一致性、motion/CLIP 等自动指标及消融；pairwise user study 报告偏好，[待核] 本次未核实人数。","one_sentence_advance":"联合短期衔接与长期外观锚定，缓解长生成停滞。","source":"https://arxiv.org/html/2403.14773v2","locator":"§3–4 memory/enhancement；§5；Appendix user study Fig.6"}},"WORLD":{"key":"WORLD","title":"WorldMem: Long-term Consistent World Simulation with Memory","authors":["Zeqi Xiao","Yushi Lan","Yifan Zhou","Wenqi Ouyang","Shuai Yang","Yanhong Zeng","Xingang Pan"],"first_year":2025,"year":2025,"venue":"NeurIPS 2025","venue_source":"https://papers.neurips.cc/paper_files/paper/2025/hash/470629a47e2d65ce0606c40055df5d26-Abstract-Conference.html","venue_proof":"正式 proceedings；优先于个别作者页面的 ICCV 标签","arxiv":"2504.12369","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2504.12369","full_source":"https://arxiv.org/html/2504.12369v3","version":"2504.12369v3","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2504.12369v3","origin":"库内主表","origin_detail":"P135","question":"有限 context window 下，如何重访早先视角时仍保持世界一致，并表现随时间的变化？","case":"Minecraft 与真实场景实验；长距离视角重访、动态变化。","operation":"memory units 储存 frame、pose、timestamp；根据状态检索并 attention 条件化。","method":"WorldMem memory attention＋pose/time embeddings；对比 full-sequence、Diffusion Forcing 及不同 pose/retrieval 设置。","evidence":"重访 reconstruction/LPIPS/rFID 等指标；memory 长度、采样策略和编码消融；长度从 8 到 16 并非继续改善。","strength":"支持显式 memory 的 spatial recall；重访自己生成的历史一致，不等于历史推断符合外部事实。","advance":"以带状态的长期 memory 超越固定生成窗口。","premise":"依赖 camera/state metadata；原文承认遮挡时 overlap retrieval 不充分、交互不够真实多样、memory 线性增长。","assumption":"保存的 frame 被默认可作为有效参考；新反证需先决定哪段 memory 失效，单纯检索更好无法完成这个判断。","locator":"§3；§4 Tables 1–2；Limitations；Appendix Table 7","taxonomy":{"task":"3D/4D 一致生成","task_note":"","contribution_primary":"记忆与上下文","contribution_secondary":["条件注入"],"control_signals":["图像","相机","动作"],"target_properties":["几何一致","环境记忆","因果或状态一致"],"property_note":"","generation_paradigm":"因果自回归或流式","paradigm_note":"按世界状态向前生成并读取过去 memory。","paper_type":"方法","contribution_focus":"带 pose/time 的 memory frames，经 memory attention 检索旧视角状态。","task_definition":{"inputs":"当前视觉/动作条件、相机状态及长期 frame-state memory。","outputs":"世界推进和视角重访时的后续视频。","invariants":"相同位置/相应时间的场景状态；世界可按动作/时间演化。","judgement":"重访 reconstruction/LPIPS/rFID；pose encoding、检索与 memory 长度消融。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"重访 reconstruction/LPIPS/rFID 等指标；memory 长度、采样策略和编码消融；长度从 8 到 16 并非继续改善。","one_sentence_advance":"以带状态的长期 memory 超越固定生成窗口。","source":"https://arxiv.org/html/2504.12369v3","locator":"§3；§4 Tables 1–2；Limitations；Appendix Table 7"}},"PERMA":{"key":"PERMA","title":"PermaVid: Consistent Video Generation Across Edits via Disentangled Context Memory","authors":["Shuai Yang","Bingjie Gao","Ziwei Liu","Jiaqi Wang","Dahua Lin","Tong Wu"],"first_year":2026,"year":2026,"venue":"NeurIPS 2026（已录用；作者列表确认）","venue_source":"https://liuziwei7.github.io/publications.html","venue_proof":"2026-10-15 Ziwei Liu 作者发表列表将该论文列于 NeurIPS 2026；未将 arXiv 版本冒记为正式 proceedings 版本。","arxiv":"2606.16449","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2606.16449","full_source":"https://arxiv.org/html/2606.16449v2","version":"2606.16449v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2606.16449v2","origin":"本次库外新找到","origin_detail":"对照旧 topics 主表与补充 JSON，未发现相同 ID/完整题名","question":"编辑使历史 context 过时后，如何在后续生成与视角重访中保持更新后的世界？","case":"UE-Mem；global style/lighting/season edits 与 local insertion/removal/replacement。","operation":"分离 RGB/depth memory；全局编辑清空旧 RGB 并推进 semantic version；局部编辑使与 Ωe 重叠的 memory units 失效。","method":"edit-aware update/retrieval＋mixed-modality conditioning；相较 RGB-only history 保留仍有效的 geometry，过滤过时 appearance。","evidence":"跨编辑重访 PSNR/SSIM/LPIPS、VBench 与 disentanglement 消融；存在局部/全局编辑的定量证据。","strength":"直接覆盖“编辑后旧记忆失效并选择性更新”的宽泛主张；未等同含噪外部证据下的诊断、来源冲突和隐藏机制纠错。","advance":"将 memory validity 与编辑类型、空间影响范围联动。","premise":"全局 appearance 编辑通常保留 geometry；局部有编辑区域 Ωe 与 pose footprint；几何变化时才同时使 depth memory 失效。","assumption":"若修改范围未知、证据不可靠或影响沿因果依赖跨区域传播，空间 footprint 规则可能不足；这才是 B/S 可检验的区别。","locator":"§3.1–3.3，Eqs.1–3；§4 Tables 1、Fig.6","taxonomy":{"task":"3D/4D 一致生成","task_note":"","contribution_primary":"记忆与上下文","contribution_secondary":["条件注入","数据"],"control_signals":["图像","相机","姿态或结构图"],"target_properties":["几何一致","环境记忆","实体一致"],"property_note":"","generation_paradigm":"因果自回归或流式","paradigm_note":"主任务是跨编辑的持续世界生成；局部编辑本身不是唯一输出。","paper_type":"方法","contribution_focus":"RGB/depth 双 memory；按全局/局部 edit invalidation 保留有效参考。","task_definition":{"inputs":"编辑后的观察、camera trajectory、RGB/depth memory 与编辑范围。","outputs":"跨时间/视角仍遵守更新状态的后续视频。","invariants":"未失效的几何/内容；被编辑内容应保持新版本而非退回旧版。","judgement":"编辑后重访 PSNR/SSIM/LPIPS、VBench 与 memory 消融。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"跨编辑重访 PSNR/SSIM/LPIPS、VBench 与 disentanglement 消融；存在局部/全局编辑的定量证据。","one_sentence_advance":"将 memory validity 与编辑类型、空间影响范围联动。","source":"https://arxiv.org/html/2606.16449v2","locator":"§3.1–3.3，Eqs.1–3；§4 Tables 1、Fig.6"}},"WEAVE":{"key":"WEAVE","title":"Weave Forcing: Compositional Memory Routing for Interactive Long Video Generation","authors":["Ziyi Wang","Junchi Yao","Heqian Qiu","Wenbo Shi","Chengjiu Wang","Jinyang He","Binkai Hong","Hongliang Li"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2610.03510","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2610.03510","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2610.03510","full_source":"https://arxiv.org/html/2610.03510v1","version":"2610.03510v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2610.03510v1","origin":"本次库外新找到","origin_detail":"对照旧 topics 主表与补充 JSON，未发现相同 ID/完整题名","question":"交互式新镜头需组合不同历史镜头的人物与背景时，如何避免取入无关内容？","case":"interactive long video generation；跨 shot subject/background consistency。","operation":"LLM 把 prompt 分成 character/background slots，选择历史 references；semantic masks 控制历史 KV tokens。","method":"training-free semantic routing＋masked memory weaving＋coverage-adaptive RoPE；对比全历史拼接、weighted fusion。","evidence":"跨/镜头内一致性、文本对齐与质量指标；cache fusion 消融，inter-shot 指标更强但不是所有 intra-shot 指标领先。","strength":"支持组件级历史重用及内部 KV 控制；“选择哪些参考”并非“判断哪个参考事实是错的”。","advance":"从 whole-prompt retrieval 转为按语义组成选择和融合历史。","premise":"需要历史中存在可用 reference；按 coverage 调整 RoPE 与 retention。","assumption":"旧记忆包含错误时，semantic relevance 仍可很高；需把 validity/provenance 加入 routing 的决策变量。","locator":"§3 Semantic slot routing、Masked memory weaving、RoPE；§4 Tables 1–2","taxonomy":{"task":"长视频与多镜头叙事","task_note":"","contribution_primary":"记忆与上下文","contribution_secondary":["推理时方法","条件注入"],"control_signals":["文本","图像","身份参考"],"target_properties":["实体一致","环境记忆","短时连贯"],"property_note":"","generation_paradigm":"因果自回归或流式","paradigm_note":"逐镜头使用历史 cache；training-free 不等于少步或实时。","paper_type":"方法","contribution_focus":"semantic slots 选择历史 KV；masked weaving 与 adaptive RoPE 控制组合参考。","task_definition":{"inputs":"新镜头 prompt、不同历史镜头中的人物/背景 references。","outputs":"组合指定历史内容的新镜头与长视频。","invariants":"选定人物身份、背景语义与相邻片段连贯；无关历史不应混入。","judgement":"inter/intra-shot consistency、text alignment 与 cache-fusion 消融。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"跨/镜头内一致性、文本对齐与质量指标；cache fusion 消融，inter-shot 指标更强但不是所有 intra-shot 指标领先。","one_sentence_advance":"从 whole-prompt retrieval 转为按语义组成选择和融合历史。","source":"https://arxiv.org/html/2610.03510v1","locator":"§3 Semantic slot routing、Masked memory weaving、RoPE；§4 Tables 1–2"}},"A2RD":{"key":"A2RD","title":"A$^2$RD: Agentic Autoregressive Diffusion for Long Video Consistency","authors":["Do Xuan Long","Yale Song","Min-Yen Kan","Tomas Pfister","Long T. Le"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2605.06924","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2605.06924","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2605.06924","full_source":"https://arxiv.org/html/2605.06924v1","version":"2605.06924v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2605.06924v1","origin":"库内主表","origin_detail":"VG15","question":"长视频如何避免 semantic drift 与 narrative collapse，维持出现—消失—再出现实体的一致性？","case":"LVBench-C 与 public long-video benchmarks；1–10 分钟视频。","operation":"每段 retrieve–synthesize–refine–update；adaptive generation mode；frame/video 两层 critic。","method":"training-free agentic autoregressive diffusion，将 creative synthesis 与 consistency enforcement 分离；消融 memory、adaptive mode、hierarchical refinement。","evidence":"自动指标/MLLM judge、消融；人评 7 名，40 条 VBench-Long single-scene 样本，每条至少 2 人；未报告人评 agreement。","strength":"支持 agent loop 改善生成一致性；单场景人评不能覆盖所有 nonlinear narrative，更不等于外部反证 revision。","advance":"将多模态 memory 与逐段双层自纠错统一成闭环。","premise":"依赖参考图、剧情状态与 critics；本次生成错误主要由内部评价环纠正。","assumption":"critic 可能和 generator 共享常识偏差；需要独立观测来区分“不一致”与“原设定错误”。","locator":"§3 Retrieve–Synthesize–Refine–Update；Experiments；Appendix Human Study、limitations","taxonomy":{"task":"长视频与多镜头叙事","task_note":"","contribution_primary":"推理时方法","contribution_secondary":["记忆与上下文","评测"],"control_signals":["文本","图像","身份参考"],"target_properties":["实体一致","环境记忆","因果或状态一致","短时连贯"],"property_note":"","generation_paradigm":"因果自回归或流式","paradigm_note":"segment-level autoregressive；内部单片段可以 extrapolation/interpolation。","paper_type":"方法","contribution_focus":"Retrieve–Synthesize–Refine–Update；frame/video 两级校验与重采样。","task_definition":{"inputs":"剧情/场景描述、参考图与已生成片段的 multimodal memory。","outputs":"逐段修订后的一至十分钟视频。","invariants":"人物/物体/环境在重现时与有效状态一致；剧情允许规定的状态变化。","judgement":"LVBench-C/public benchmarks、模块消融及 7 人 human study。"},"evidence_types":["定量 benchmark","消融","user study"],"classification_label":"推断","evidence_detail":"自动指标/MLLM judge、消融；人评 7 名，40 条 VBench-Long single-scene 样本，每条至少 2 人；未报告人评 agreement。","one_sentence_advance":"将多模态 memory 与逐段双层自纠错统一成闭环。","source":"https://arxiv.org/html/2605.06924v1","locator":"§3 Retrieve–Synthesize–Refine–Update；Experiments；Appendix Human Study、limitations"}},"VOY":{"key":"VOY","title":"Voyager: An Open-Ended Embodied Agent with Large Language Models","authors":["Guanzhi Wang","Yuqi Xie","Yunfan Jiang","Ajay Mandlekar","Chaowei Xiao","Yuke Zhu","Linxi Fan","Anima Anandkumar"],"first_year":2023,"year":2024,"venue":"TMLR 2024","venue_source":"https://openreview.net/pdf?id=ehfRiF0R3a","venue_proof":"正式论文页首 Published in TMLR 03/2024","arxiv":"2305.16291","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2305.16291","full_source":"https://arxiv.org/html/2305.16291v2","version":"2305.16291v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2305.16291v2","origin":"库内主表","origin_detail":"VG02","question":"LLM agent 能否开放式探索，积累可复用技能，并在新世界完成未见任务？","case":"Minecraft；探索、科技树、零样本新任务。","operation":"automatic curriculum 选任务；环境/执行错误反馈改代码；成功代码进入可检索 skill library。","method":"GPT-4 iterative prompting＋executable skill memory；相较只保留语言反思，存可调用程序；不更新 LLM 参数。","evidence":"物品多样性、探索距离、科技树速度与组件消融；新任务每项 3 次试验。","strength":"支持环境内可执行技能的累积与复用；不支持教程生成的持续提升或真实学员学习改善。","advance":"把执行成功的程序作为可组合长期经验。","premise":"Mineflayer 提供低层控制与结构化反馈；GPT-4 通过 API 调用。","assumption":"代码可复现执行不等于教学策略的可迁移有效性；学员改变可能造成结果变化而非策略改善。","locator":"§2 Curriculum/Skill Library/Iterative Prompting；Experiments Tables 1–2","taxonomy":{"task":"其他","task_note":"开放式 embodied agent 学习","contribution_primary":"记忆与上下文","contribution_secondary":["推理时方法"],"control_signals":[],"target_properties":[],"property_note":"技能积累、任务成功与迁移","generation_paradigm":"不适用","paradigm_note":"","paper_type":"方法","contribution_focus":"持久 executable skill library；环境反馈改程序并复用。","task_definition":{"inputs":"Minecraft 环境状态、任务/curriculum 与已学技能。","outputs":"可执行代码、动作及新技能。","invariants":"游戏规则和已验证技能语义；不是视频外观保持。","judgement":"物品/探索/科技树和新世界未见任务成功率。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"物品多样性、探索距离、科技树速度与组件消融；新任务每项 3 次试验。","one_sentence_advance":"把执行成功的程序作为可组合长期经验。","source":"https://arxiv.org/html/2305.16291v2","locator":"§2 Curriculum/Skill Library/Iterative Prompting；Experiments Tables 1–2"}},"VIDEOAG":{"key":"VIDEOAG","title":"VideoAgent: Self-Improving Video Generation","authors":["Achint Soni","Sreyas Venkataraman","Abhranil Chandra","Sebastian Fischmeister","Percy Liang","Bo Dai","Sherry Yang"],"first_year":2024,"year":2024,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2410.10076","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2410.10076","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2410.10076","full_source":"https://arxiv.org/html/2410.10076v3","version":"2410.10076v3","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2410.10076v3","origin":"本次库外新找到","origin_detail":"对照旧 topics 主表与补充 JSON，未发现相同 ID/完整题名","question":"如何利用外部反馈修正 hallucinated video plans，并通过环境交互进一步改善生成器？","case":"MetaWorld manipulation、iTHOR navigation；真实机器人视频额外展示。","operation":"生成视频计划先由反馈自条件修正；执行后可采集环境数据再训练。","method":"self-conditioning consistency 将反馈与当前计划用于改进生成；与直接执行 AVDC 类 video plan 比较。","evidence":"模拟下游 task success、生成指标与消融；iTHOR 为 4 rooms×3 objects×20 seeds，只测无 online access 的修正；真实机器人主要定性。","strength":"已覆盖“视频计划接受环境反馈并进一步训练”的宽泛问题；不能把所有设置都说成已验证在线自进化。","advance":"把外部反馈接入生成的 inference refinement 与数据积累。","premise":"可以从视频提取控制并在模拟环境执行；不同实验允许的 online access 不同。","assumption":"执行反馈同时受控制器影响；若不拆分 planner/generator/controller，不能把任务成功增益全归因于视频因果理解。","locator":"§3 Self-conditioning；§4 MetaWorld/iTHOR；Appendix real-world qualitative evaluation","taxonomy":{"task":"动作条件的预测与交互世界模型","task_note":"","contribution_primary":"推理时方法","contribution_secondary":["训练信号"],"control_signals":["文本","图像"],"target_properties":["因果或状态一致","物理合理","指令遵循"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"按轮生成/修订一段 video plan；闭环执行不等于生成器本身逐帧因果流式。","paper_type":"方法","contribution_focus":"self-conditioning consistency 修订 video plans；执行数据可进一步训练生成器。","task_definition":{"inputs":"当前观察、任务指令与生成/执行反馈。","outputs":"修订的视频计划及从计划导出的控制。","invariants":"任务目标与当前场景；计划必须可被下游控制器执行。","judgement":"MetaWorld/iTHOR task success、生成指标与消融；真实机器人视频定性。"},"evidence_types":["定量 benchmark","消融","定性展示"],"classification_label":"推断","evidence_detail":"模拟下游 task success、生成指标与消融；iTHOR 为 4 rooms×3 objects×20 seeds，只测无 online access 的修正；真实机器人主要定性。","one_sentence_advance":"把外部反馈接入生成的 inference refinement 与数据积累。","source":"https://arxiv.org/html/2410.10076v3","locator":"§3 Self-conditioning；§4 MetaWorld/iTHOR；Appendix real-world qualitative evaluation"}},"SPIRAL":{"key":"SPIRAL","title":"SPIRAL: Self-Evolving Action-Conditioned Video Generation via Reflective Planning Agents","authors":["Yu Yang","Yue Liao","Jianbiao Mei","Baisen Wang","Xuemeng Yang","Licheng Wen","Jiangning Zhang","Xiangtai Li","Liang Lv","Hanlin Chen","Botian Shi","Yong Liu","Shuicheng Yan","Gim Hee Lee"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2603.08403","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2603.08403","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2603.08403","full_source":"https://arxiv.org/html/2603.08403v3","version":"2603.08403v3","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2603.08403v3","origin":"本次库外新找到","origin_detail":"对照旧 topics 主表与补充 JSON，未发现相同 ID/完整题名","question":"长时动作条件生成怎样避免漏步、幻觉运动与 temporal drift，并以闭环反馈改进模型？","case":"ActVideoGen-Dataset/Bench；EgoPlan-Bench、VideoGen-RewardBench、VBench。","operation":"PlanAgent 分解 sub-actions，VideoGenerator 按 memory 生成，CriticAgent 反馈重试；用 actions/rewards 做 GRPO post-training。","method":"think–act–reflect 闭环＋generator 参数后训练；区别于一次生成或仅推理时修补。","evidence":"多 backbone benchmark、组件/重试消融；规划与 critic 也分别评测；原文 OOD planning 并非所有 baseline 都领先。","strength":"支持闭环与后训练的联合价值；不是跨真实学员长期教学改进证据。不能把每轮局部纠错等同跨任务累积。","advance":"将 action planning、segment correction 与后训练统一。","premise":"需 task-specific 数据和 critic reward；正文读取 v3，题名/数据集名不同于早期版本。","assumption":"critic 的 action-quality reward 若偏离教学结果，参数改进仍可能只提高可见执行形式；需独立任务指标。","locator":"正文 v3 §3；Experiments；Appendix Tables 10–11","taxonomy":{"task":"动作条件的预测与交互世界模型","task_note":"","contribution_primary":"架构与生成范式","contribution_secondary":["训练信号","记忆与上下文","数据","评测"],"control_signals":["文本","图像","动作"],"target_properties":["因果或状态一致","短时连贯","实体一致","指令遵循"],"property_note":"","generation_paradigm":"因果自回归或流式","paradigm_note":"原文明确按 action step 分段 streaming；主张不等于低延迟实时。","paper_type":"方法","contribution_focus":"plan–generate–critic 逐动作闭环；memory 保持状态，GRPO 后训练 generator。","task_definition":{"inputs":"高层任务、分解后的 sub-actions 与历史视觉/latent memory。","outputs":"依动作逐段生成并可重试的视频。","invariants":"操作次序、持续动作执行与场景状态连贯。","judgement":"ActVideoGen-Bench、规划/critic benchmarks、VBench 与消融。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"多 backbone benchmark、组件/重试消融；规划与 critic 也分别评测；原文 OOD planning 并非所有 baseline 都领先。","one_sentence_advance":"将 action planning、segment correction 与后训练统一。","source":"https://arxiv.org/html/2603.08403v3","locator":"正文 v3 §3；Experiments；Appendix Tables 10–11"}},"CINE":{"key":"CINE","title":"CineForge: Self-Improving Agents for Long-Horizon Video Generation","authors":["Junxiang Liu","Lin Wang","Haiyu Shi","Hongxu Ma","Xiaoyu Yang","Chunjie Chen","Xiaoxiao Xu","Kaiqiao Zhan","Boao Wang","Shuizhou Shi","Tianyun Zhu","Jie Li","Jiangtong Li"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2608.29621","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2608.29621","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2608.29621","full_source":"https://arxiv.org/html/2608.29621v1","version":"2608.29621v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2608.29621v1","origin":"库内主表","origin_detail":"VG19","question":"能否从多个故事生产的重复失败中获得持久、针对阶段的策略改进？","case":"CineScope-Data 100 scripts；AnimeShooter、ViStoryBench；长故事视频生产。","operation":"记录 typed production trajectories；case→pattern→policy patches；通过 replay 和 paired tests 决定 persistent update。","method":"CineForge-Produce 与 CineForge-Evolve 分离；区别于只修当前视频的 test-time loop；修改 stage-local policy。","evidence":"跨故事与 evolution rounds；CineScope 4.024→4.380；metric 校准由 10 名评审评 180 视频，每条至少 5 评分；有受条件限制的统计 lemma。","strength":"直接覆盖“跨故事累积改进视频生产 agent”；理论不覆盖 correlated opportunities，评分提升也不是教学效能验证。","advance":"让生产轨迹形成可审核的跨任务策略更新，而非仅单次返工。","premise":"outcome metric 在演化阶段冻结且不参与 discovery/admission；统计 lemma 依赖条件独立/预声明机会等条件。","assumption":"真实学员反馈受用户能力和任务分配混淆；需要证据归因、保留旧任务和跨用户迁移，不能只把 story 改名 tutorial。","locator":"§3 Produce/Evolve；§5；Appendix human alignment、CPPE admission lemma","taxonomy":{"task":"长视频与多镜头叙事","task_note":"","contribution_primary":"推理时方法","contribution_secondary":["表示","记忆与上下文","评测"],"control_signals":["文本","图像","身份参考"],"target_properties":["因果或状态一致","实体一致","指令遵循","效率与延迟"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"整故事离线计划与分镜渲染；跨故事 policy 演化不是 token 自回归。","paper_type":"方法","contribution_focus":"typed production trajectory→跨故事 stage-local policy patch；不是 backbone 参数自训练。","task_definition":{"inputs":"完整 story/script、生产策略与既有生产案例。","outputs":"完成的长故事视频及经验证的持久 production-policy 更新。","invariants":"故事的事件/角色/空间约束；更新需满足回归 guards。","judgement":"CineScope/public benchmarks、跨故事演化对照与 metric–human alignment。"},"evidence_types":["定量 benchmark","消融","user study"],"classification_label":"推断","evidence_detail":"跨故事与 evolution rounds；CineScope 4.024→4.380；metric 校准由 10 名评审评 180 视频，每条至少 5 评分；有受条件限制的统计 lemma。","one_sentence_advance":"让生产轨迹形成可审核的跨任务策略更新，而非仅单次返工。","source":"https://arxiv.org/html/2608.29621v1","locator":"§3 Produce/Evolve；§5；Appendix human alignment、CPPE admission lemma"}},"MEMFAULT":{"key":"MEMFAULT","title":"Useful Memories Become Faulty When Continuously Updated by LLMs","authors":["Dylan Zhang","Yanshan Lin","Zhengkun Wu","Yihang Sun","Bingxuan Li","Dianqi Li","Hao Peng"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2605.12978","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2605.12978","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2605.12978","full_source":"https://arxiv.org/html/2605.12978v2","version":"2605.12978v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2605.12978v2","origin":"库内主表","origin_detail":"P097","question":"LLM 持续将经验压缩成抽象记忆时，为什么有用的经历会变成有害 memory？","case":"ALFWorld、ScienceWorld、WebShop、AppWorld、Mind2Web、ARC-AGI Stream。","operation":"同样轨迹采用不同 consolidation schedule；可 Retain/Delete/Consolidate；对照 episodic-only 与 no-memory。","method":"受控比较 raw episodes、一次汇总与增量改写；把退化定位到 consolidation，而不是只比较最终 agent 总分。","evidence":"多 benchmark/time-stream 实验；论文报告特定 ARC 子集上 54% 原可解题在 GT consolidation 后失败；配有 schedule/solver 控制。","strength":"说明特定模型与任务中记忆更新能造成回归；不能概括为所有 memory 都有害，也不能推定同一比例适用于视频。","advance":"将“持续记忆必然改善”的假设变成可反驳的受控问题。","premise":"研究特定 LLM、prompt、task stream 与 memory budget；ARC 子集并非总体 ARC 成功率。","assumption":"教学经验应保留原始证据、版本与回归集合；抽象规则的可读性不能替代可迁移有效性。","locator":"§3–5 experiments；ARC-AGI Stream；Appendix controlled consolidation comparison","taxonomy":{"task":"其他","task_note":"LLM agent memory 更新的失败分析","contribution_primary":"评测","contribution_secondary":["记忆与上下文"],"control_signals":[],"target_properties":[],"property_note":"更新收益、旧任务回归与记忆效用","generation_paradigm":"不适用","paradigm_note":"","paper_type":"分析","contribution_focus":"受控替换 consolidation schedule；用 episodic-only 对照定位 memory regression。","task_definition":{"inputs":"相同任务经验、解题轨迹、memory update schedule 与 solver。","outputs":"不同记忆策略下的任务表现/回归分析。","invariants":"对照中的经验来源、solver 等条件；刻意改变 memory 更新方式。","judgement":"多 task streams 上准确率/成功率、保留/删除/归纳对照。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"多 benchmark/time-stream 实验；论文报告特定 ARC 子集上 54% 原可解题在 GT consolidation 后失败；配有 schedule/solver 控制。","one_sentence_advance":"将“持续记忆必然改善”的假设变成可反驳的受控问题。","source":"https://arxiv.org/html/2605.12978v2","locator":"§3–5 experiments；ARC-AGI Stream；Appendix controlled consolidation comparison"}},"DIRECTOR":{"key":"DIRECTOR","title":"VideoDirectorGPT: Consistent Multi-scene Video Generation via LLM-Guided Planning","authors":["Han Lin","Abhay Zala","Jaemin Cho","Mohit Bansal"],"first_year":2023,"year":2024,"venue":"COLM 2024","venue_source":"https://j-min.io/publication/videodirectorgpt_colm2024/","venue_proof":"作者出版页与论文首页","arxiv":"2309.15091","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2309.15091","full_source":"https://arxiv.org/html/2309.15091v2","version":"2309.15091v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2309.15091v2","origin":"本次库外新找到","origin_detail":"对照旧 topics 主表与补充 JSON，未发现相同 ID/完整题名","question":"LLM 能否规划视频内容及布局，使多场景生成保持实体与空间一致？","case":"VPEval Skill-based prompts、ActionBench-Direction、multi-scene generation。","operation":"文本扩为 scene descriptions、entity layouts、backgrounds、consistency groups；控制 layout strength。","method":"GPT-4 Video Planner＋Layout2Vid；以 image-level annotations 训练布局控制，对比 ModelScopeT2V、I2V 等。","evidence":"布局/方向准确率、生成质量、多场景结果和 gated attention/布局控制消融。","strength":"支持 planning-to-rendering 控制；没有度量观众何时知道秘密，也不证明事件世界和信息揭示独立。","advance":"在多场景中显式维护实体布局和一致性分组。","premise":"依赖 pretrained T2V backbone；论文承认训练分布、画质和 English corpus 限制。","assumption":"scene plan 写了保密要求，pixels 仍可泄漏；语义计划满足与感知层信息控制需两种评价。","locator":"§3；§4 Tables 1–2；Appendix limitations、gated attention ablation","taxonomy":{"task":"长视频与多镜头叙事","task_note":"","contribution_primary":"表示","contribution_secondary":["条件注入"],"control_signals":["文本","姿态或结构图","身份参考"],"target_properties":["实体一致","短时连贯","指令遵循"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"先全局规划再渲染场景；scene 顺序不自动构成 causal video decoder。","paper_type":"方法","contribution_focus":"显式 scene plan、entity layouts 与 consistency groups 指导 Layout2Vid。","task_definition":{"inputs":"文本请求，可选用户图像参考。","outputs":"有场景布局/实体一致性的多场景视频。","invariants":"人物/对象 identity、跨场景 consistency groups 与指定空间动作。","judgement":"布局/方向准确率、多场景结果与 attention/layout 消融。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"布局/方向准确率、生成质量、多场景结果和 gated attention/布局控制消融。","one_sentence_advance":"在多场景中显式维护实体布局和一致性分组。","source":"https://arxiv.org/html/2309.15091v2","locator":"§3；§4 Tables 1–2；Appendix limitations、gated attention ablation"}},"ONE":{"key":"ONE","title":"OneStory: Coherent Multi-Shot Video Generation with Adaptive Memory","authors":["Zhaochong An","Menglin Jia","Haonan Qiu","Zijian Zhou","Xiaoke Huang","Zhiheng Liu","Weiming Ren","Kumara Kahatapitiya","Ding Liu","Sen He","Chenyang Zhang","Tao Xiang","Fanny Yang","Serge Belongie","Tian Xie"],"first_year":2025,"year":2026,"venue":"CVPR 2026","venue_source":"https://openaccess.thecvf.com/content/CVPR2026/html/An_OneStory_Coherent_Multi-Shot_Video_Generation_with_Adaptive_Memory_CVPR_2026_paper.html","venue_proof":"正式 CVF proceedings","arxiv":"2512.07802","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2512.07802","full_source":"https://arxiv.org/html/2512.07802v1","version":"2512.07802v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2512.07802v1","origin":"库内主表","origin_detail":"P133","question":"多镜头叙事怎样压缩长期历史，同时保留新镜头需要的实体与环境上下文？","case":"60K multi-shot dataset；T2MSV/I2MSV benchmarks；reappearance/composition 等模式。","operation":"从先前镜头选 informative frames，importance-guided patchification 后作为 next-shot 条件。","method":"Frame Selection＋Adaptive Conditioner，将 MSV 改写为 autoregressive next-shot generation；相较单 keyframe/有限历史窗口引入 compact global memory。","evidence":"T2MSV/I2MSV 定量与 selection/conditioner/context budget 消融；一 latent-frame 等量 context 已带来改善。","strength":"支持跨镜头外观/叙事一致性及高效 context；未提供观众知识状态或提前泄密指标。","advance":"将多镜头历史的选择与空间 token 分配联合用于生成。","premise":"以 pretrained I2V（Wan2.1）微调；参考来自历史镜头，caption 带 referential relations。","assumption":"模型用于角色辨识的内部完整信息不应全部暴露给观众；需区别“生成器知道”与“画面允许显示”。","locator":"§3；§4 Tables 1–2；Appendix benchmark patterns","taxonomy":{"task":"长视频与多镜头叙事","task_note":"","contribution_primary":"记忆与上下文","contribution_secondary":["条件注入","数据"],"control_signals":["文本","图像","身份参考"],"target_properties":["实体一致","环境记忆","短时连贯","效率与延迟"],"property_note":"","generation_paradigm":"因果自回归或流式","paradigm_note":"shot-level autoregressive；单镜头仍调用 pretrained I2V。","paper_type":"方法","contribution_focus":"历史 frame selection＋importance-guided patchification，压缩 next-shot context。","task_definition":{"inputs":"当前 shot caption、历史镜头与可选初始图像。","outputs":"自回归生成的下一镜头和多镜头视频。","invariants":"跨镜头再出现的人物/背景与 referential relationships。","judgement":"T2MSV/I2MSV consistency/quality；selection、conditioner 和 context-budget 消融。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"T2MSV/I2MSV 定量与 selection/conditioner/context budget 消融；一 latent-frame 等量 context 已带来改善。","one_sentence_advance":"将多镜头历史的选择与空间 token 分配联合用于生成。","source":"https://arxiv.org/html/2512.07802v1","locator":"§3；§4 Tables 1–2；Appendix benchmark patterns"}},"VIS":{"key":"VIS","title":"VIS-Ground: Video Interactive Storytelling with Contextual Grounding","authors":["Bingxuan Li","Yiwen Song","Xueqing Wu","Yanzhou Pan","Yang Li","Kuang Su","Jingyun Liu","Sebastian Ko","Huan Zhang","Tong Zhang","Nanyun Peng","Tomas Pfister","Yale Song"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2610.09326","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2610.09326","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2610.09326","full_source":"https://arxiv.org/html/2610.09326v2","version":"2610.09326v2","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2610.09326v2","origin":"既有补充参考","origin_detail":"旧 topics 数据已有，未列入 papers.json 主表","question":"观众在已生成视频中提出新要求时，如何联合 source、rendered history 和 intent 得到可执行约束？","case":"narrative/knowledge grounded interactive continuation；三个 video backbones；小规模教育 pilot。","operation":"输入异质 source、已播 prefix 与新 request；抽取 state/dependencies，投影 candidate-specific constraints 后生成/验证/修订。","method":"Structured Context Abstraction→Generation Constraints Induction→Constrained Video Generation；约束保留 predicates/evidence 链接，输出 pass/fail/unknown 并指导定点修订；baseline 为共享 renderer 下的适配版。","evidence":"VIS-Bench 从 250 passages 构造 1004 interventions、2008 planned instruction–style instances；三个 backbone 的 composite 提升；pilot 招 9 人，learning gain 仅 7 人，experience ratings 4 人。数据打包前审计 information leakage。","strength":"已覆盖可追溯约束和 unknown evidence 的宽主张；benchmark 审计泄漏不等于测试观众秘密揭示时间。比较的是适配 baseline，pilot 不足以确立通用教学效果。","advance":"显式建模只有 source、history、intent 联合考虑才出现的依赖。","premise":"依赖 source 与 prefix 可用及 dependency extraction；原文承认抽取和 rendering faithful realization 尚有困难。","assumption":"任务要求 continuation 不违背 source 和 established history；当外部反证说明旧 history 所依据的推断本来错误时，需要决定哪些依赖撤销，而不仅验证新脚本兼容性。是否可直接扩展到此协议仍待实测。","locator":"§3；§4；Appendix E baseline adaptation、F pilot denominators","taxonomy":{"task":"长视频与多镜头叙事","task_note":"","contribution_primary":"表示","contribution_secondary":["推理时方法","评测"],"control_signals":["文本","图像","身份参考"],"target_properties":["因果或状态一致","实体一致","指令遵循"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"论文评估从 prefix 生成 48 秒 continuation；交互续写不是低延迟 streaming 证明。","paper_type":"方法","contribution_focus":"source/prefix/intent→evidence-linked dependencies→可执行 generation constraints。","task_definition":{"inputs":"外部 narrative/knowledge source、已播视频 prefix、新 viewer request 与 references。","outputs":"满足多源约束的多镜头 continuation。","invariants":"source 支持的事实及已建立 history；只在相容范围内响应新请求。","judgement":"source faithfulness、continuity、interaction fulfillment 与 visual metrics；pilot 单列。"},"evidence_types":["定量 benchmark","user study"],"classification_label":"推断","evidence_detail":"VIS-Bench 从 250 passages 构造 1004 interventions、2008 planned instruction–style instances；三个 backbone 的 composite 提升；pilot 招 9 人，learning gain 仅 7 人，experience ratings 4 人。数据打包前审计 information leakage。","one_sentence_advance":"显式建模只有 source、history、intent 联合考虑才出现的依赖。","source":"https://arxiv.org/html/2610.09326v2","locator":"§3；§4；Appendix E baseline adaptation、F pilot denominators"}},"GRID":{"key":"GRID","title":"Multi-Grid Post-Training for Long-Form Multi-Shot Video Generation","authors":["Jiawei Mao","Haoqin Tu","Hardy Chen","Yuhan Wang","Keyang Xu","Jieru Mei","Hongliang Fei","Ruogu Fang","Wei Shao","Cihang Xie","Yuyin Zhou"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2609.06373","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2609.06373","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2609.06373","full_source":"https://arxiv.org/html/2609.06373v1","version":"2609.06373v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2609.06373v1","origin":"既有补充参考","origin_detail":"旧 topics 数据已有，未列入 papers.json 主表","question":"单时间轴 packing 难表现完整镜头集合时，怎样在同预算下生成更长的多镜头叙事？","case":"MovieGrid；由 1000 长视频生成 54K grid videos；89 stories、五种视觉类别的 benchmark。","operation":"把时间有序 chunks 放入空间 grid，保留随机 subset 作为 clean context；控制 grid embedding/boundary loss。","method":"Multi-Grid Post-Training＋Noise-Free Random-Grid Training；与 Temporal Packing 及多镜头模型比较。","evidence":"同 token budget、1616-frame 条件下 shots 数为 Temporal Packing 的 6.05 倍；intra/inter-shot 一致性和模块消融。","strength":"支持表示/训练方式改善长镜头覆盖；镜头数和身份一致性并不测故事信息是否提前泄露。","advance":"通过空间网格分摊时间轴负担并允许 chunks 全局交换信息。","premise":"训练以固定间隔 chunk 组织，不要求 chunk 与物理 shot 一一对应；使用自动视觉指标。","assumption":"全局交换信息有利于一致性，却可能增加未来事实对早期画面的泄漏；这种风险需要显式信息边界测试。","locator":"§3；§4 Table 1；Appendix A story-level benchmark","taxonomy":{"task":"长视频与多镜头叙事","task_note":"","contribution_primary":"表示","contribution_secondary":["架构与生成范式","训练信号","数据"],"control_signals":["文本","图像"],"target_properties":["实体一致","短时连贯","指令遵循","效率与延迟"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"主要比较 joint multi-grid generation；支持多轮扩展不改变本表主范式。","paper_type":"方法","contribution_focus":"把有序 video chunks 空间 grid 化联合去噪；random clean-grid context 与 boundary loss。","task_definition":{"inputs":"character-aware story prompt 与可保留的 clean grid chunks。","outputs":"跨网格联合生成后还原顺序的长视频。","invariants":"事件顺序、跨 chunk 角色身份与网格布局边界。","judgement":"相同 token budget 下镜头覆盖、intra/inter-shot consistency、画质和消融。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"同 token budget、1616-frame 条件下 shots 数为 Temporal Packing 的 6.05 倍；intra/inter-shot 一致性和模块消融。","one_sentence_advance":"通过空间网格分摊时间轴负担并允许 chunks 全局交换信息。","source":"https://arxiv.org/html/2609.06373v1","locator":"§3；§4 Table 1；Appendix A story-level benchmark"}},"DRAM":{"key":"DRAM","title":"Dramatis: A Computational Model of Suspense","authors":["Brian O’Neill","Mark Riedl"],"first_year":2014,"year":2014,"venue":"AAAI 2014","venue_source":"https://ojs.aaai.org/index.php/AAAI/article/view/8836","venue_proof":"正式 proceedings / DOI","arxiv":"","doi":["10.1609/aaai.v28i1.8836"],"identifier_note":"","source":"https://ojs.aaai.org/index.php/AAAI/article/view/8836","full_source":"https://ojs.aaai.org/index.php/AAAI/article/view/8836/8695","version":"正式 proceedings PDF","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://ojs.aaai.org/index.php/AAAI/article/view/8836/8695","origin":"本次库外新找到","origin_detail":"对照旧 topics 主表与补充 JSON，未发现相同 ID/完整题名","question":"能否从叙事结构计算接近人类读者体验的 suspense？","case":"Casino Royale、Rear Window、Harry Potter 场景的三对文本故事。","operation":"删改/重排事件、提前引入解决方案；改变 MEI-P memory model 与 planning goal 选择。","method":"以 STRIPS time-slices、reader memory activation 和最可能 escape plan 的成本计算 suspense；不同于穷举全部成功/失败路径。","evidence":"32 名参与者评三对故事；模型曲线面积与人类故事排序比较；memory 与 goal 构造消融。","strength":"支持三个受控故事上的认知模型对应关系；不是自动视频生成系统，也不是所有观众的心理定律。","advance":"把读者的记忆可及性和可行脱困计划显式用于 suspense 计算。","premise":"符号动作、preconditions/effects 与 scripts 可用；假设更易从记忆提取的行动被视为更可能成功。","assumption":"真实观众可自行看向未规划线索；符号 time-slice 无法独自保证 XR 画面的信息边界。","locator":"AAAI PDF pp.2–5：Dramatis、Evaluations 1–3","taxonomy":{"task":"其他","task_note":"计算叙事中的 suspense 建模","contribution_primary":"表示","contribution_secondary":["评测"],"control_signals":[],"target_properties":[],"property_note":"认知模型与人类体验的对应","generation_paradigm":"不适用","paradigm_note":"","paper_type":"方法","contribution_focus":"STRIPS time-slices＋reader memory activation＋escape-plan cost 表示 suspense。","task_definition":{"inputs":"符号叙事、角色/动作 effects 与典型情境 scripts。","outputs":"随故事推进的 suspense curve。","invariants":"输入叙事事实；建模读者状态，而非生成或编辑画面。","judgement":"三对故事的 curve-area 排序与 32 人主观 suspense、memory/goal 消融。"},"evidence_types":["user study","消融"],"classification_label":"推断","evidence_detail":"32 名参与者评三对故事；模型曲线面积与人类故事排序比较；memory 与 goal 构造消融。","one_sentence_advance":"把读者的记忆可及性和可行脱困计划显式用于 suspense 计算。","source":"https://ojs.aaai.org/index.php/AAAI/article/view/8836/8695","locator":"AAAI PDF pp.2–5：Dramatis、Evaluations 1–3"}},"SUSP":{"key":"SUSP","title":"Creating Suspenseful Stories: Iterative Planning with Large Language Models","authors":["Kaige Xie","Mark Riedl"],"first_year":2024,"year":2024,"venue":"EACL 2024","venue_source":"https://aclanthology.org/2024.eacl-long.147/","venue_proof":"正式 ACL Anthology","arxiv":"2402.17119","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2402.17119","full_source":"https://arxiv.org/html/2402.17119v1","version":"2402.17119v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2402.17119v1","origin":"本次库外新找到","origin_detail":"对照旧 topics 主表与补充 JSON，未发现相同 ID/完整题名","question":"如何用 LLM 生成具有 suspense 的故事，而不仅依赖一句“写悬疑故事”？","case":"English thriller text generation；Prolific human evaluation。","operation":"建立 protagonist/goal/dire outcome；迭代生成行动与失败原因；改变 clue setup 和 information revelation timing。","method":"理论驱动 iterative prompting：计划脱困后设置阻碍，最后展开章节；对比直接 prompt、Llama 2、Re3 及规划/展开消融。","evidence":"主比较 90 人评 30 对故事；revelation timing 另招 90 人评 30 对，事前 42.1% 对事后 36.3%，差异不显著；clue setup 效果显著。","strength":"信息揭示时间已被明确操纵；不能把不显著结果写成“提前揭示更有效”，也不能把文本实验外推为视频信息控制已解决。","advance":"把 suspense 理论落实为 LLM planning，并区分线索、揭示时机与 empathy 的影响。","premise":"仅英语；理论基于 Western storytelling tradition；人评关注读者体验。","assumption":"文本由明确语句揭示，视频还会通过背景、视角和反射泄漏；C 的增量须落在这些可测的感知信息约束。","locator":"§3；§4.1–4.4，Information revelation timing；Limitations","taxonomy":{"task":"其他","task_note":"LLM 文本悬疑故事生成","contribution_primary":"推理时方法","contribution_secondary":[],"control_signals":["文本"],"target_properties":[],"property_note":"suspense、享受度与叙事逻辑；不强映射为视频性质","generation_paradigm":"不适用","paradigm_note":"","paper_type":"方法","contribution_focus":"迭代规划 protagonist action/阻碍，再展开章节；操纵线索与信息揭示时机。","task_definition":{"inputs":"protagonist、目标、失败后果等故事设定。","outputs":"悬疑故事文本与迭代大纲。","invariants":"角色目标与故事背景；实验比较控制共同背景。","judgement":"人评 suspense/novelty/enjoyment/logical sense；揭示时机与 clue setup 对照。"},"evidence_types":["user study","消融"],"classification_label":"推断","evidence_detail":"主比较 90 人评 30 对故事；revelation timing 另招 90 人评 30 对，事前 42.1% 对事后 36.3%，差异不显著；clue setup 效果显著。","one_sentence_advance":"把 suspense 理论落实为 LLM planning，并区分线索、揭示时机与 empathy 的影响。","source":"https://arxiv.org/html/2402.17119v1","locator":"§3；§4.1–4.4，Information revelation timing；Limitations"}},"CONTROL":{"key":"CONTROL","title":"Adding Conditional Control to Text-to-Image Diffusion Models","authors":["Lvmin Zhang","Anyi Rao","Maneesh Agrawala"],"first_year":2023,"year":2023,"venue":"ICCV 2023","venue_source":"https://openaccess.thecvf.com/content/ICCV2023/papers/Zhang_Adding_Conditional_Control_to_Text-to-Image_Diffusion_Models_ICCV_2023_paper.pdf","venue_proof":"正式 CVF proceedings","arxiv":"2302.05543","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2302.05543","full_source":"https://arxiv.org/html/2302.05543v3","version":"2302.05543v3","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2302.05543v3","origin":"本次库外新找到","origin_detail":"对照旧 topics 主表与补充 JSON，未发现相同 ID/完整题名","question":"如何给大型 pretrained image diffusion 加入空间控制，而不破坏原有生成能力？","case":"Stable Diffusion；edge、depth、segmentation、human pose 等图像条件。","operation":"单/多 spatial conditions，可有/无 text prompt；控制信号与可训练分支。","method":"冻结 backbone、复制可训练 encoder，用 zero convolutions 连接；对比 sketch-guided diffusion、PITI、ControlNet-lite。","evidence":"多类条件实验、组件消融；user study 邀 12 人比较 20 sketches×5 methods。","strength":"支持 spatial conditioning 的架构方法；不决定何时索取条件、哪种条件最小充分，更不是视频因果控制证明。","advance":"以稳定初始化的可训练分支接入多种空间约束。","premise":"有 paired condition/image training examples；论文实验为图像。","assumption":"控制带宽高不等于信息必要性；D 需要依据目标选择条件，而不是固定让用户提供全部信号。","locator":"§3 ControlNet；§4 User Study、ablation","taxonomy":{"task":"其他","task_note":"图像条件生成；非视频 ControlNet","contribution_primary":"条件注入","contribution_secondary":["架构与生成范式"],"control_signals":["文本","图像","姿态或结构图"],"target_properties":["画质","指令遵循"],"property_note":"","generation_paradigm":"不适用","paradigm_note":"图像任务，视频生成范式不适用。","paper_type":"方法","contribution_focus":"冻结 diffusion backbone，用 zero-conv 接入可训练空间条件分支。","task_definition":{"inputs":"文本及 edge/depth/segmentation/pose 等条件图。","outputs":"符合空间条件的图像。","invariants":"给定空间布局/姿态约束；不要求保留某个源视频的全部像素。","judgement":"多条件实验、模块消融和 12 人草图生成偏好。"},"evidence_types":["定量 benchmark","消融","user study"],"classification_label":"推断","evidence_detail":"多类条件实验、组件消融；user study 邀 12 人比较 20 sketches×5 methods。","one_sentence_advance":"以稳定初始化的可训练分支接入多种空间约束。","source":"https://arxiv.org/html/2302.05543v3","locator":"§3 ControlNet；§4 User Study、ablation"}},"SKETCH":{"key":"SKETCH","title":"SketchDynamics: Exploring Free-Form Sketches for Dynamic Intent Expression in Animation Generation","authors":["Boyu Li","Lin-Ping Yuan","Zeyu Wang","Hongbo Fu"],"first_year":2026,"year":2026,"venue":"CHI 2026","venue_source":"https://yuanlinping.top/assets/paper/2026_SketchDynamics.pdf","venue_proof":"作者保存的正式论文，DOI 10.1145/3772318.3791071；已补核 conditional 标签","arxiv":"2601.20622","doi":["10.1145/3772318.3791071"],"identifier_note":"","source":"https://arxiv.org/abs/2601.20622","full_source":"https://arxiv.org/html/2601.20622v1","version":"2601.20622v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2601.20622v1","origin":"本次库外新找到","origin_detail":"对照旧 topics 主表与补充 JSON，未发现相同 ID/完整题名","question":"自由草图如何表达 dynamic intent，并通过人与 VLM 的共同澄清变成动画？","case":"Sketch storyboard→motion graphics；三阶段研究共 24 人；Manim 2D vector animations。","operation":"用户画自由草图、回应 interpretation clarification、迭代修改；少量 3D/video applicability 展示。","method":"VLM 将草图解释为可执行动画代码，界面允许澄清与 refinement；区别于固定 command brushes 的强约束符号。","evidence":"24 人三阶段设计研究，定性观察/问卷；主体为短 storyboard 和数十秒 motion graphics。","strength":"支持自由草图中的歧义及协同解释价值；不能据此声称对长视频所有控制量达到最少用户输入。","advance":"将草图歧义视作可共同解释的设计空间，而非强制消除为固定命令。","premise":"参与者多为有相关背景的大学生；渲染需秒级；受 Manim 表达范围限制。","assumption":"草图表达了什么、用户本来想要什么、生成器执行了什么必须分别记录；否则低保真易被错判为意图缺失。","locator":"System；Three-stage study；§Limitations；正式 CHI DOI 补核","taxonomy":{"task":"条件控制生成","task_note":"","contribution_primary":"推理时方法","contribution_secondary":["表示"],"control_signals":["文本","图像","姿态或结构图"],"target_properties":["指令遵循"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"解释并生成代码后离线渲染数十秒动画；不是少步视频 diffusion。","paper_type":"方法","contribution_focus":"草图→VLM 意图解释/澄清→Manim 动画代码；不是新增 diffusion backbone。","task_definition":{"inputs":"自由 storyboard sketches、文字补充、澄清回答及素材。","outputs":"motion graphics 动画，可迭代修改。","invariants":"用户确认的动态意图；草图不要求逐像素照搬。","judgement":"三阶段 24 人研究的任务观察、问卷和访谈；跨 3D/video 主要为展示。"},"evidence_types":["user study","定性展示"],"classification_label":"推断","evidence_detail":"24 人三阶段设计研究，定性观察/问卷；主体为短 storyboard 和数十秒 motion graphics。","one_sentence_advance":"将草图歧义视作可共同解释的设计空间，而非强制消除为固定命令。","source":"https://arxiv.org/html/2601.20622v1","locator":"System；Three-stage study；§Limitations；正式 CHI DOI 补核"}},"COG":{"key":"COG","title":"CogOmniControl: Reasoning-Driven Controllable Video Generation via Creative Intent Cognition","authors":["Hongji Yang","Songlian Li","Yucheng Zhou","Xiaotong Zhao","Alan Zhao","Chengzhong Xu","Jianbing Shen"],"first_year":2026,"year":2026,"venue":"arXiv 预印本；[待核] 本次未核实正式 venue","venue_source":"https://arxiv.org/abs/2605.19995","venue_proof":"预印本已核；不据此断言尚未录用","arxiv":"2605.19995","doi":[],"identifier_note":"","source":"https://arxiv.org/abs/2605.19995","full_source":"https://arxiv.org/html/2605.19995v1","version":"2605.19995v1","read_date":"2026-10-14","read_scope":"已打开摘要；正文重点核对方法、实验、消融和前提，非逐字全文复核","local_evidence":"https://arxiv.org/html/2605.19995v1","origin":"既有补充参考","origin_detail":"旧 topics 数据已有，未列入 papers.json 主表","question":"抽象、稀疏、专业制作条件下，如何识别 creative intent 并忠实生成视频？","case":"anime production storyboards/clay renders；CogReasonBench、CogControlBench（200 representative high-resolution samples）。","operation":"输入草图/制作条件；CogVLM 转成 dense reasoning 与语义特征，和 pixel conditions 一同控制生成。","method":"专业数据训练 CogVLM＋CogOmniDiT in-context fusion＋RL alignment＋adaptive evaluator Best-of-N；对比通用 VLM/adapter 控制。","evidence":"两个 benchmark、reasoning/控制/selection 组件比较；数据来自真实制作流程并经人工语义对齐。","strength":"支持专业意图理解与内部融合；“把 sparse cue 解释得更密”不证明补出的信息来自真实用户或达到最小充分性。","advance":"把专门 creative-intent cognition 与生成 backbone 对齐，而非只附加原始条件。","premise":"专业训练/评价分布；reasoning data 含教师生成并经人工核查；依赖 aligned condition-target pairs。","assumption":"一个稀疏草图可有多个合法意图，强先验可能替用户作决定；应把需要澄清的多解性与可执行控制区分。","locator":"§3 CogVLM/CogOmniDiT；§4 Benchmark construction、Experiments","taxonomy":{"task":"条件控制生成","task_note":"","contribution_primary":"条件注入","contribution_secondary":["训练信号","表示","推理时方法","评测"],"control_signals":["文本","图像","姿态或结构图"],"target_properties":["指令遵循","实体一致","短时连贯","画质"],"property_note":"","generation_paradigm":"离线整段生成","paradigm_note":"按制作请求生成视频片段；Best-of-N 是采样策略，不是实时范式。","paper_type":"方法","contribution_focus":"CogOmniDiT 融合 pixel conditions 与 CogVLM semantic features；RL 对齐并 Best-of-N。","task_definition":{"inputs":"专业 storyboard sketches/clay renders、文字等稀疏制作条件。","outputs":"遵循 creative intent 的视频。","invariants":"条件规定的构图、动作与专业意图；不必复制输入草图的渲染外观。","judgement":"CogReasonBench/CogControlBench 与 reasoning/控制/selection 消融。"},"evidence_types":["定量 benchmark","消融"],"classification_label":"推断","evidence_detail":"两个 benchmark、reasoning/控制/selection 组件比较；数据来自真实制作流程并经人工语义对齐。","one_sentence_advance":"把专门 creative-intent cognition 与生成 backbone 对齐，而非只附加原始条件。","source":"https://arxiv.org/html/2605.19995v1","locator":"§3 CogVLM/CogOmniDiT；§4 Benchmark construction、Experiments"}}},"merges":[{"group":"XR-C + AIGC-A","core":"部分可识别机制下的选择性后果生成","decision":"合并方法研究；XR 为受限观察和执行验证场景。","reason":"输入都包含历史与干预；主要评价都是后代改对、非后代保留。只有引入真实操作/部分观察才产生额外约束。","sources":["VOID","GIVE","PHYS","DSCM"]},{"group":"XR-S + AIGC-B","core":"证据冲突下的依赖撤销与生成记忆有效性","decision":"合并；优先收窄外部反证与来源裁决。","reason":"同一 history/evidence→memory/plan/video update 链；PermaVid 已覆盖宽泛 edit-aware update，必须测不同问题。","sources":["PERMA","BLACK","VIS","WEAVE"]},{"group":"XR-L + AIGC-D","core":"用户 GT 澄清是充分信息获取的一个受限子问题","decision":"共享接口，不直接合并任务；XR-L 固定触发并只研究 how。","reason":"D 包含搜索、观察和生成控制获取；L 的答案应由用户已知意图／指代提供，实验需排除未知物理机制探索。","sources":["KNOW","GPS","S3","VGEN","COG"]},{"group":"XR-E","core":"跨用户教学反馈的归因、迁移与回归控制","decision":"保留独立候选，以跨任务教学迁移和回归评测为门槛。","reason":"ACE／MCE 等已有持续改进机制；需证明教学反馈的归因与迁移，而非把单任务修补称为自进化。","sources":["CINE","VIDEOAG","MEMFAULT"]},{"group":"AIGC-C","core":"世界状态固定时的观众信息边界","decision":"可独立，但需 video/XR 感知证据，不能仅改剧情。","reason":"改变观众知道什么不等于改变世界；其变量与 A 的 intervention、B 的纠错不同。","sources":["DRAM","SUSP","ONE","VIS"]}],"control_map":[["外部语义状态/约束","CWMDT、VIS-Ground","先改 structured state / dependencies / constraints，再编排渲染。","CWMDT VIS"],["外部 mask 与几何","Generative Omnimatte、VOID、GIVE","指定允许修改区域，或 pre/post geometry；VOID 的新轨迹主要由 VDM 生成。","OMNI VOID GIVE"],["外部物理模拟","VideoPhysEdit","重建后执行物理干预，输出 trajectories 指导生成；条件受刚体场景限制。","PHYS"],["模型内部条件分支/特征","ControlNet、CogOmniControl","可训练分支或 pixel/semantic feature fusion；这是条件执行，不自动构成内部 causal variable。","CONTROL COG"],["模型内部 memory/KV","WorldMem、PermaVid、Weave Forcing","attention、validity-filtered memory、semantic KV masking/RoPE；已存在内部控制，未自动解决证据可信度。","WORLD PERMA WEAVE"],["模型参数/agent policy 更新","VANS、SPIRAL、VideoGen-Agent、CineForge","分别优化跨模态生成、生成器参数、工具策略、生产阶段策略；更新对象不能混淆。","VANS SPIRAL VGEN CINE"]],"applications":[["AIGC-A","电影/VFX 替代结局","删除触发事件并重写连锁后果，保留演员/场景身份。","OMNI","PHYS","OMNI 只给相关后效删除的已接收证据；PHYS 刚体限制与真实表演仍有距离。"],["AIGC-A","游戏机制与回放","同一 seed 的机关/碰撞干预及影响范围 GT，适合可证伪基准。","OMNI","PHYS","先验证真实机制后果，再研究视觉生成；并非完整游戏世界模型已解决。"],["AIGC-A","操作训练/安全演示","漏步、换序的后果对照。","VANS","GIVE","VANS 是视频答案；GIVE 是几何编辑。二者都不能直接替代操作风险真值。"],["AIGC-B","剧本杀/交互电影","新线索推翻角色身份猜测，局部修改未来互动。","BLACK","VIS","BlackSwan 验证推理缺口；VIS 处理上下文依赖，但含噪反证的裁决需新实验。"],["AIGC-B","动漫短剧连续制作","设定纠错之后保留不受影响人物与画面资产。","ONE","WEAVE","OneStory/WEAVE 保持或组合历史；需加 PermaVid 作为编辑后失效的直接 baseline。"],["AIGC-B","游戏 persistent world","回到旧区域后也遵守被证据修正的状态。","WORLD","PERMA","PermaVid 已支持编辑后重访；新价值要落在未知影响区和不可靠证据。"],["AIGC-C","剧本杀/悬疑电影","同一事件图制作不同线索揭示顺序，并测提前泄密。","SUSP","VIS","SUSP 已操纵揭示时机且该项不显著；VIS 提供约束续写，不证明观众体验。"],["AIGC-C","动漫短剧/电影剪辑","角色知道、观众知道与世界事实分开控制。","ONE","GRID","多镜头一致性/覆盖率是 renderer 基础，不是信息揭示指标。"],["AIGC-C","XR 叙事游戏","玩家自主换视角时保持线索边界与后期一致揭示。","DRAM","VIS","Dramatis 是符号/文本 reader model；XR 感知泄漏尚需实测。"],["AIGC-D","动漫 storyboard production","决定索取角色参考、草图、轨迹或关键帧中的哪一个。","SKETCH","COG","专业意图条件化已存在；剩余问题是查询选择与充分性。"],["AIGC-D","电影/VFX 导演工作流","高成本渲染前获取最少的拍摄/动作约束。","CONTROL","VGEN","ControlNet 奠基于图像；VideoGen-Agent 已学工具策略。不能把二者描述成电影制作验证。"],["AIGC-D","游戏/XR 教程","按任务影响选择问人、转视角、查规则、看示范。","GPS","S3","GPS 是规则 QA；S3 是 streaming spatial reasoning；最终视频/执行结果须另测。"]],"search_log":[{"area":"页面与库内基线","query":"soniasuns.github.io/topics/data/{xr,aigc}-directions.json；papers.json 与既有补充 JSON","result":"线上 4 XR＋4 AIGC 定义已重新获取；旧库 451 篇作为“库内”基线。","kind":"范围记录"},{"area":"procedure/state","query":"Procedure Planning Instructional Videos；SCHEMA；What Changed and What Could Have Changed；Video-as-Answer","result":"PROC 为库外早期代表；SCHEMA/STATE/VANS 任务类型分开；VANS 补核 CVPR 2026。","kind":"纳入"},{"area":"物理/反事实","query":"counterfactual video generation 2026 September October；VOID；GIVE；VideoPhysEdit；Generative Omnimatte","result":"纳入 object effects→dynamic consequences 及 geometry/physics 两条分支；不把所有 counterfactual 都视为 SCM 个体反事实。","kind":"纳入"},{"area":"反事实术语的相邻用法","query":"Counterfactual Video Generation Enables Scalable Humanoid Loco-Manipulation","result":"已打开 arXiv:2609.38172 摘要；PRISM 用 V2V 扩增 human-object interaction 数据并训练机器人。保留为应用邻居，未进入 38 篇精读卡：不是本表“固定同一历史的后果选择性”主问题。","kind":"边界候选","url":"https://arxiv.org/abs/2609.38172"},{"area":"memory revision","query":"video generation belief revision counter evidence；PermaVid；Weave Forcing；video memory contradictory generation","result":"PermaVid、Weave Forcing 为主表/既有补充之外的新近邻；PermaVid 改变 B/S 新颖性边界。","kind":"纳入"},{"area":"持续改进","query":"self improving video generation CineForge；VideoAgent Self-Improving；SPIRAL","result":"VideoAgent/SPIRAL 库外纳入；与已有 CineForge/MEMFAULT 联合，区分 inference repair、parameter update、persistent policy。","kind":"纳入"},{"area":"主动获取","query":"Robots That Ask For Help；Learning to Clarify Action Based；GPS Graph-guided Proactive Information Seeking","result":"纳入 KnowNo、ACT、GPS。GPS 的正式 proceedings/PDF 可访问；OpenReview 的一个 PDF 入口遇到浏览器验证，改用 proceedings。arXiv/DOI 未找到。","kind":"纳入／标识缺口"},{"area":"视觉主动探索","query":"Interactive Visual Reasoning；Generative World Explorer；See Remember Explore","result":"三者分别为实验、想象和 streaming observation；未混算为真实主动感知证据。","kind":"纳入"},{"area":"生成上下文/控制","query":"video generation active clarification context；Qwen-Image-Agent；VideoGen-Agent；SketchDynamics；CogOmniControl","result":"纳入 image/video/HCI 三类近邻并注明输出域；SketchDynamics 从 arXiv conditional 标签补核正式 CHI DOI。","kind":"纳入"},{"area":"AIGC-C 库外反证检索","query":"computational narrative suspense generation reader knowledge planning information disclosure；Creating Suspenseful Stories","result":"Dramatis 与 SUSP 实际打开正文；找到 revelation timing 的直接人类实验，推翻“时间分离本身是新问题”的宽主张。","kind":"纳入"},{"area":"AIGC-C 同题近期补查","query":"2025 2026 story generation \"information revelation\"；2025 2026 suspense generation reader knowledge narrative planning；2025 video generation reveal narrative","result":"在本次检索范围内未找到第三篇 2024–2026、直接测试 revelation timing 且已打开原文的代表作。普通 story planning、SeqBench、NarrLV 搜索命中未用来填 L9 缺额。","kind":"未找到／不凑数"},{"area":"venue 核实","query":"WorldMem NeurIPS/ICCV；OneStory CVPR；StreamingT2V CVPR；VANS CVPR；VideoDirectorGPT COLM；Voyager TMLR","result":"优先正式 proceedings。WorldMem 为 NeurIPS 2025；VIDEODIRECTORGPT 为 COLM 2024；VideoAgent 仅保留预印本，不将 ICLR under review 当录用。","kind":"书目更正"}],"scope_history":{"XR-L-before-2026-10-15":{"code":"XR-L","name":"主动获取充分上下文的 XR 教程生成","definition":"生成下一段 XR 教程前，选择询问、换视角、检索或示范，以有限交互成本取得会改变操作决策的信息。","lineages":["L3","L4"],"history":"KnowNo 从校准不确定性决定求助；ACT 学 dialogue actions；GPS 在条件图中找缺失变量。IVRE 把假设检验做成可交互实验，S3 把空间推理变成 streaming active perception；Qwen/VideoGen-Agent 将信息与工具链连接到生成。","boundary":"澄清和主动探索本身已有理论/benchmark/真实系统证据。未被这些证据回答的是：问人、看一眼、读手册、索取动作示范之间，何种预算分配能保证最终教程的决策相关信息足够。GenEx 的 imagined view 不能替代真实确认。","core":"如何依据下游操作分支的可区分性选择下一条证据，并在已足够时停止，而不是最大化通用信息量或把所有 context 收齐？","why":"(a) 关键在目标相关的充分性和停止准则，非更大 context window；(b) QA 正确率、生成质量或独立 help coverage 未建立跨模态 acquisition→video→execution 的条件保证；(c) 可构造只有一种证据渠道能辨别正确操作的任务。","experiment":{"case":"装配教程中随机隐藏零件型号、背面锁扣、用户目标或动作手法；分别只有手册、换视角、问人或示范能消歧。","operation":"控制信息缺失类型、噪声、证据价格与一次交互预算；同一生成器接受不同 acquisition policies 的输出。","controls":"不查询、固定全查询、随机同成本、entropy/information-gain、KnowNo-style confidence、GPS-style DAG、S3-style spatial exploration；oracle sufficient context。","metrics":"最终步骤/视频约束满足与执行成功、成本—成功率 Pareto curve、过早停止率、冗余查询率；删除被选证据的 counterfactual ablation 检查必要性。","falsifier":"若固定便宜渠道或普通 information gain 在混合任务中同样好，跨模态决策策略的额外复杂度不成立；若 oracle context 也失败，瓶颈在 renderer/控制执行。"},"neighbours":"GPS/ACT 已做有效澄清，S3 已做 streaming active perception，VideoGen-Agent 已做工具选择。候选差别是以最终视频/操作风险为终点、显式成本、可诊断的最小充分性；不能只声称“agent 会主动问问题”。","verdict":"与 AIGC-D 共用 acquisition policy 和 benchmark 协议；XR 引入感知动作与物理执行代价，可作为严格测试环境。","sources":["KNOW","ACT","GPS","IVR","S3","GENEX","QWEN","VGEN"]}},"recent_evidence":{"date":"2026-10-15","selection_rule":"按问题变量与方法接口的接近度排序；不按 venue 排名。正式发表与预印本分别标注；新增卡不自动计入原 451 条主库。","papers":{"CLARIFIER":{"key":"CLARIFIER","title":"Plug-and-Play Clarifier: A Zero-Shot Multimodal Framework for Egocentric Intent Disambiguation","year":2026,"venue":"AAAI 2026","publication_status":"已发表／已确认会议归属","source":"https://ojs.aaai.org/index.php/AAAI/article/view/38851","venue_source":"https://ojs.aaai.org/index.php/AAAI/article/view/38851","arxiv":"2511.08971","finding":"通过文本对话、视觉采集引导与 3D 指向 grounding 分解第一人称多模态歧义。","boundary":"[推断] 最接近 XR 澄清接口；视觉模块仍让用户调整位置，不能说全部取证动作由 agent 自己完成。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"GPS":{"key":"GPS","title":"GPS: Graph-guided Proactive Information Seeking in Large Language Models","year":2026,"venue":"ICLR 2026","publication_status":"已发表／已确认会议归属","source":"https://proceedings.iclr.cc/paper_files/paper/2026/hash/1f49b7fbd7c7afc52c1db0d4ed1a338d-Abstract-Conference.html","venue_source":"https://proceedings.iclr.cc/paper_files/paper/2026/hash/1f49b7fbd7c7afc52c1db0d4ed1a338d-Abstract-Conference.html","arxiv":"","finding":"以条件规则 DAG 引导询问、遍历和裁剪，减少无关澄清。","boundary":"[推断] 适合用户持有缺失条件的任务；规则预先给定，未验证 XR 交互呈现和视频执行。","read_date":"2026-10-15","read_scope":"本轮摘要与发表信息复核；既有全文拆解另列，阅读日期与版本保留","formal_ai_venue":true,"deep_analysis_key":"GPS"},"CLARITI":{"key":"CLARITI","title":"Asking What Matters: Reward-Driven Clarification for Software Engineering Tasks","year":2026,"venue":"ICML 2026","publication_status":"已发表／已确认会议归属","source":"https://proceedings.mlr.press/v306/vijayvargiya26a.html","venue_source":"https://proceedings.mlr.press/v306/vijayvargiya26a.html","arxiv":"","finding":"用 task relevance 与 user answerability 奖励训练 8B CLARITI，通过模拟用户评价澄清。","boundary":"[推断] 如何问最有价值且用户能答的问题很接近；原任务为软件工程，原文也研究 when，本方向只借用 how。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"POINTER":{"key":"POINTER","title":"Uncertain Pointer: Situated Feedforward Visualizations for Ambiguity-Aware AR Target Selection","year":2026,"venue":"CHI 2026","publication_status":"已发表／已确认会议归属","source":"https://arxiv.org/abs/2602.13433","venue_source":"https://www.chingyitsai.com/_posts/projects/UncertainPointer.html","arxiv":"2602.13433","finding":"探索 AR 指向歧义的前馈可视化；两项在线实验分别 N=60、N=40。","boundary":"[推断] 提供候选对象与置信度呈现依据，不是自主澄清策略，也不直接生成教程。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":false},"IGCLAR":{"key":"IGCLAR","title":"Uncertainty-Aware Clarification in LLM Agents with Information Gain","year":2026,"venue":"ICML 2026","publication_status":"已发表／已确认会议归属","source":"https://proceedings.mlr.press/v306/deng26l.html","venue_source":"https://proceedings.mlr.press/v306/deng26l.html","arxiv":"","finding":"将澄清引起的、朝 ground-truth goal 的 Bayesian belief update 作为 Information Gain Reward；在扩展 τ-Bench 测试。","boundary":"[推断] 适合作为如何选择问题的训练基线；没有 XR 行动或视频执行验证。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"SC2":{"key":"SC2","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","publication_status":"已发表／已确认会议归属","source":"https://proceedings.mlr.press/v306/yao26u.html","venue_source":"https://proceedings.mlr.press/v306/yao26u.html","arxiv":"","finding":"利用 world-model foresight 产生内部反馈，修订执行前计划；必要时选择性进行 test-time 模型适应。","boundary":"[推断] 直接涉及 state-level 与 model-level correction，但任务是导航；内部反馈不等于外来反证，输出也不是教程视频。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"SPATIA":{"key":"SPATIA","title":"Spatia: Video Generation with Updatable Spatial Memory","year":2026,"venue":"CVPR 2026","publication_status":"已发表／已确认会议归属","source":"https://arxiv.org/abs/2512.15716","venue_source":"https://openaccess.thecvf.com/content/CVPR2026/papers/Zhao_Spatia_Video_Generation_with_Updatable_Spatial_Memory_CVPR_2026_paper.pdf","arxiv":"2512.15716","finding":"以可更新空间记忆支持视频生成，利用生成片段和视觉 SLAM 更新空间表示。","boundary":"[推断] 很适合研究记忆到视频的执行接口；空间更新本身不建立外部证据推翻旧信念的选择性撤销。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":false},"COME":{"key":"COME","title":"Composition of Memory Experts for Diffusion World Models","year":2026,"venue":"ICLR 2026","publication_status":"已发表／已确认会议归属","source":"https://proceedings.iclr.cc/paper_files/paper/2026/hash/19eae75beed66321d62272e794a9c2ac-Abstract-Conference.html","venue_source":"https://proceedings.iclr.cc/paper_files/paper/2026/hash/19eae75beed66321d62272e794a9c2ac-Abstract-Conference.html","arxiv":"","finding":"组合短期、长期与空间记忆专家；长期专家通过轻量 test-time finetuning 将历史保留在外部 diffusion 权重中。","boundary":"[推断] 解决历史保持与一致性；未据此建立冲突证据的来源裁决与依赖撤销。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"SPACE":{"key":"SPACE","title":"Theory of Space: Can Foundation Models Construct Spatial Beliefs through Active Exploration?","year":2026,"venue":"ICLR 2026","publication_status":"已发表／已确认会议归属","source":"https://arxiv.org/abs/2602.07055","venue_source":"https://arxiv.org/abs/2602.07055","arxiv":"2602.07055","finding":"用 spatial belief probing 和 false-belief 设置发现 Belief Inertia：新证据到来后模型仍保留过时先验。","boundary":"[推断] 与外来反证的失败定义非常接近；属于评测与分析，不提供视频生成修复方法。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"VOID":{"key":"VOID","title":"VOID: Video Object and Interaction Deletion","year":2026,"venue":"ECCV 2026","publication_status":"已发表／已确认会议归属","source":"https://void-model.github.io/","venue_source":"https://void-model.github.io/","arxiv":"2604.02296","finding":"以 Kubric／HUMOTO 配对反事实数据训练；识别删除对象后受影响区域，并重生成下游交互。","boundary":"[推断] 最直接的选择性视频后果编辑先例；主要是删除，不能外推为任意动作或未知机制修改。","read_date":"2026-10-15","read_scope":"本轮摘要与发表信息复核；既有全文拆解另列，阅读日期与版本保留","formal_ai_venue":false,"deep_analysis_key":"VOID"},"CAUSAL":{"key":"CAUSAL","title":"Causal-Adapter: Taming Text-to-Image Diffusion for Faithful Counterfactual Generation","year":2026,"venue":"ICML 2026","publication_status":"已发表／已确认会议归属","source":"https://leitong02.github.io/causaladapter/","venue_source":"https://leitong02.github.io/causaladapter/","arxiv":"2509.24798","finding":"以 SCM 传播属性干预，并通过条件注入和 token regularization 兼顾依赖属性改变与图像身份保持。","boundary":"[推断] 最适合读 Δ／Keep 的因果条件化思路；原任务为图像，未验证跨时间的视频因果链。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"MULTI":{"key":"MULTI","title":"Multiverse Mechanica: A Testbed for Learning Game Mechanics via Counterfactual Worlds","year":2026,"venue":"ICLR 2026","publication_status":"已发表／已确认会议归属","source":"https://proceedings.iclr.cc/paper_files/paper/2026/hash/f9bc7653cf176075e88203d1e83e8c65-Abstract-Conference.html","venue_source":"https://proceedings.iclr.cc/paper_files/paper/2026/hash/f9bc7653cf176075e88203d1e83e8c65-Abstract-Conference.html","arxiv":"","finding":"构建带 ground-truth 因果 DAG 的游戏机制 testbed，产生反事实世界数据并展示机制学习的微调概念验证。","boundary":"[推断] 可借鉴配对真值与机制一致性评测；不是现实 XR 视频编辑器。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"GIVE":{"key":"GIVE","title":"Geometry-Instructed Video Editing","year":2026,"venue":"arXiv 2026","publication_status":"预印本；正式 venue 未核实","source":"https://arxiv.org/abs/2606.24225","venue_source":"https://arxiv.org/abs/2606.24225","arxiv":"2606.24225","finding":"用编辑前后几何条件和配对数据控制视频对象变换，包含阴影、反射等次生效果。","boundary":"[推断] 与选择性执行直接接近；给定几何的控制不等于自行推断未知机制的全部后果。","read_date":"2026-10-15","read_scope":"本轮摘要与发表信息复核；既有全文拆解另列，阅读日期与版本保留","formal_ai_venue":false,"deep_analysis_key":"GIVE"},"ARIFY":{"key":"ARIFY","title":"ARify: Leveraging Narrated Instructional Videos to Create Augmented Reality Tutorials for Procedural Tasks","year":2026,"venue":"CHI 2026","publication_status":"已发表／已确认会议归属","source":"https://engineering.purdue.edu/cdesign/wp/arify/","venue_source":"https://doi.org/10.1145/3772318.3790715","arxiv":"","finding":"从带解说的教学视频构建 AR 教程计划与内容，支持用户修改；报告三类机器任务与 N=18 用户研究。","boundary":"[推断] 最接近 XR 教程生产流程；没有据此建立跨任务持续改善。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":false},"COACH":{"key":"COACH","title":"AgentCoach: LLM-Based Adaptive Coaching Feedback for Motor Skill Learning","year":2026,"venue":"CHI 2026","publication_status":"已发表／已确认会议归属","source":"https://engineering.purdue.edu/cdesign/wp/agentcoach/","venue_source":"https://doi.org/10.1145/3772318.3791652","arxiv":"","finding":"将 coaching points 编译为运动学评价器，结合错误诊断与历史记录生成适应性反馈。","boundary":"[推断] 连接学员反馈与教学策略；当前核实内容不足以声称跨任务策略学习，样本量在本补充记录中未确认。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":false},"ACE":{"key":"ACE","title":"Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models","year":2026,"venue":"ICLR 2026","publication_status":"已发表／已确认会议归属","source":"https://arxiv.org/abs/2510.04618","venue_source":"https://ace-agent.github.io/blogs/2026-01-30-iclr-acceptance/","arxiv":"2510.04618","finding":"通过 generation、reflection、curation 持续更新可复用 context playbook。","boundary":"[推断] 适合先实现跨任务经验积累；context 演化不是模型权重训练，也不自动保证教学迁移。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"MCE":{"key":"MCE","title":"Meta Context Engineering via Agentic Skill Evolution","year":2026,"venue":"ICML 2026","publication_status":"已发表／已确认会议归属","source":"https://proceedings.mlr.press/v306/ye26g.html","venue_source":"https://proceedings.mlr.press/v306/ye26g.html","arxiv":"2601.21557","finding":"联合演化 context-engineering skills 和 context artifacts，在五类领域的离线、在线设置中评价。","boundary":"[推断] 比仅记录成功提示更接近策略演化；多领域实验不能直接解释成跨领域无遗忘迁移。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"DARWIN":{"key":"DARWIN","title":"Darwinian Memory: A Training-Free Self-Regulating Memory System for GUI Agent Evolution","year":2026,"venue":"ICML 2026","publication_status":"已发表／已确认会议归属","source":"https://proceedings.mlr.press/v306/mi26b.html","venue_source":"https://proceedings.mlr.press/v306/mi26b.html","arxiv":"2601.22528","finding":"将 GUI 轨迹拆为可组合记忆单元，并依据使用价值管理与裁剪经验。","boundary":"[推断] 可借鉴经验失效与保留机制；不自动提供教学策略无回归的保证。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"SILVR":{"key":"SILVR","title":"SILVR: Self-Improving Loops for Visual Robotic Planning","year":2026,"venue":"ICLR 2026","publication_status":"已发表／已确认会议归属","source":"https://arxiv.org/abs/2506.06658","venue_source":"https://arxiv.org/abs/2506.06658","arxiv":"2506.06658","finding":"用自生成成功轨迹改进视觉机器人规划，在 MetaWorld 与两个真实机器人任务中测试。","boundary":"[推断] 连接生成与执行反馈，但目标任务改善不等于跨任务 lifelong retention。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"JARVIS":{"key":"JARVIS","title":"JarvisEvo: Towards a Self-Evolving Photo Editing Agent with Synergistic Editor-Evaluator Optimization","year":2026,"venue":"CVPR 2026","publication_status":"已发表／已确认会议归属","source":"https://arxiv.org/abs/2511.23002","venue_source":"https://openaccess.thecvf.com/content/CVPR2026/html/Lin_JarvisEvo_Towards_a_Self-Evolving_Photo_Editing_Agent_with_Synergistic_Editor-Evaluator_CVPR_2026_paper.html","arxiv":"2511.23002","finding":"通过编辑器与评价器协同优化改进图像编辑 agent，并关注 reward hacking。","boundary":"[推断] 可借鉴训练反馈与评价器演化；图像编辑收益不是学员学习收益。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":false},"VISTA":{"key":"VISTA","title":"VISTA: A Test-Time Self-Improving Video Generation Agent","year":2026,"venue":"CVPR 2026","publication_status":"已发表／已确认会议归属","source":"https://research.google/pubs/vista-towards-test-time-self-improving-video-generation-agent/","venue_source":"https://research.google/pubs/vista-towards-test-time-self-improving-video-generation-agent/","arxiv":"2510.15831","finding":"通过时间计划、候选比较与多维反馈迭代修订视频生成提示。","boundary":"[推断] 主要是当前任务 test-time improvement；不能因 self-improving 名称就计为跨任务学习。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":false},"DGM":{"key":"DGM","title":"Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents","year":2026,"venue":"ICLR 2026","publication_status":"已发表／已确认会议归属","source":"https://arxiv.org/abs/2505.22954","venue_source":"https://www.jennyzhangzt.com/","arxiv":"2505.22954","finding":"修改 agent 自身代码，维护候选归档并在编码任务中检验新变体。","boundary":"[推断] 是 RSI／自修改 agent 的方法近邻，离 XR 教学策略和视频生成较远。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":true},"CINE":{"key":"CINE","title":"CineForge: Self-Improving Agents for Long-Horizon Video Generation","year":2026,"venue":"arXiv 2026","publication_status":"预印本；正式 venue 未核实","source":"https://arxiv.org/abs/2608.29621","venue_source":"https://arxiv.org/abs/2608.29621","arxiv":"2608.29621","finding":"以阶段性策略修改与回放验证研究跨故事的视频 agent 改进。","boundary":"[推断] 跨任务视频流程比单次 prompt 修补更接近，但教学有效性仍需独立验证。","read_date":"2026-10-15","read_scope":"本轮摘要与发表信息复核；既有全文拆解另列，阅读日期与版本保留","formal_ai_venue":false,"deep_analysis_key":"CINE"},"TUTOR":{"key":"TUTOR","title":"Generative Tutorial: Towards Live Contextualized Visual Instructions for Physical Tasks","year":2026,"venue":"arXiv 2026","publication_status":"预印本；正式 venue 未核实","source":"https://arxiv.org/abs/2609.24955","venue_source":"https://arxiv.org/abs/2609.24955","arxiv":"2609.24955","finding":"根据工作区上下文生成目标图像与演示视频，在 AR 原型中报告 N=24 用户研究。","boundary":"[推断] 直接接近物理任务视觉教程；尚不能作为跨任务 self-evolution 的证据。","read_date":"2026-10-15","read_scope":"摘要／作者项目页与发表信息；不冒充本轮阅读全文","formal_ai_venue":false},"PERMA":{"key":"PERMA","title":"PermaVid: Consistent Video Generation Across Edits via Disentangled Context Memory","year":2026,"venue":"NeurIPS 2026","publication_status":"已录用；作者发表列表确认，正式 proceedings 未核实","source":"https://arxiv.org/abs/2606.16449","venue_source":"https://liuziwei7.github.io/publications.html","arxiv":"2606.16449","finding":"分离 RGB 外观与 depth 几何记忆，并以 edit-aware 更新和检索处理编辑后失效的上下文。","boundary":"[推断] 最接近记忆有效性与后续视频执行；编辑导致状态变化，不等同于新观测证明历史判断错误。","read_date":"2026-10-15","read_scope":"本轮摘要与发表信息复核；既有全文拆解另列，阅读日期与版本保留","formal_ai_venue":true,"deep_analysis_key":"PERMA"},"SUSP":{"key":"SUSP","title":"Creating Suspenseful Stories: Iterative Planning with Large Language Models","year":2024,"venue":"EACL 2024","publication_status":"已发表／已确认会议归属","source":"https://aclanthology.org/2024.eacl-long.147/","venue_source":"https://aclanthology.org/2024.eacl-long.147/","arxiv":"2402.17119","finding":"以迭代故事规划研究悬念，并直接测试 revelation timing；原拆解中该项人评差异不显著。","boundary":"[推断] 最接近事件与读者获知时间的分离；是文本故事，而且不能宣称其揭示时间实验已经证明提升。","read_date":"2026-10-15","read_scope":"本轮摘要与发表信息复核；既有全文拆解另列，阅读日期与版本保留","formal_ai_venue":false,"deep_analysis_key":"SUSP"},"VIS":{"key":"VIS","title":"VIS-Ground: Video Interactive Storytelling with Contextual Grounding","year":2026,"venue":"arXiv 2026","publication_status":"预印本；正式 venue 未核实","source":"https://arxiv.org/abs/2610.09326","venue_source":"https://arxiv.org/abs/2610.09326","arxiv":"2610.09326","finding":"连接源叙事、已渲染历史与新交互意图，以约束支持规划、校验、修订和视频渲染。","boundary":"[推断] 适合借鉴事实—呈现依赖与约束检查；不等于已有完整的观众知识时间控制方法。","read_date":"2026-10-15","read_scope":"本轮摘要与发表信息复核；既有全文拆解另列，阅读日期与版本保留","formal_ai_venue":false,"deep_analysis_key":"VIS"},"ONE":{"key":"ONE","title":"OneStory: Coherent Multi-Shot Video Generation with Adaptive Memory","year":2026,"venue":"CVPR 2026","publication_status":"已发表／已确认会议归属","source":"https://arxiv.org/abs/2512.07802","venue_source":"https://openaccess.thecvf.com/content/CVPR2026/html/An_OneStory_Coherent_Multi-Shot_Video_Generation_with_Adaptive_Memory_CVPR_2026_paper.html","arxiv":"2512.07802","finding":"以自适应记忆维持多镜头生成的历史内容与实体一致性。","boundary":"[推断] 提供晚揭示与跨镜头重现所需的生成底座；未直接评测信息泄露或观众知识状态。","read_date":"2026-10-15","read_scope":"本轮摘要与发表信息复核；既有全文拆解另列，阅读日期与版本保留","formal_ai_venue":false,"deep_analysis_key":"ONE"},"QWEN":{"key":"QWEN","title":"Qwen-Image-Agent: Bridging the Context Gap in Real-World Image Generation","year":2026,"venue":"arXiv 2026","publication_status":"预印本；正式 venue 未核实","source":"https://arxiv.org/abs/2606.26907","venue_source":"https://arxiv.org/abs/2606.26907","arxiv":"2606.26907","finding":"通过规划、推理、搜索、记忆和反馈弥补真实图像生成任务的 context gap。","boundary":"[推断] 最接近缺失上下文到生成的链路；是图像任务，未证明“最少且足够”的信息保证。","read_date":"2026-10-15","read_scope":"本轮摘要与发表信息复核；既有全文拆解另列，阅读日期与版本保留","formal_ai_venue":false,"deep_analysis_key":"QWEN"},"VGEN":{"key":"VGEN","title":"VideoGen-Agent: Reinforcing Video Generation Agents","year":2026,"venue":"arXiv 2026","publication_status":"预印本；正式 venue 未核实","source":"https://arxiv.org/abs/2609.24997","venue_source":"https://arxiv.org/abs/2609.24997","arxiv":"2609.24997","finding":"以 SFT 和 agentic RL 学习信息补充、视频生成与验证工具的使用。","boundary":"[推断] 最直接的视频 agent 基线；选择工具不等于证明最少充分控制，也不保证每条条件被生成器执行。","read_date":"2026-10-15","read_scope":"本轮摘要与发表信息复核；既有全文拆解另列，阅读日期与版本保留","formal_ai_venue":false,"deep_analysis_key":"VGEN"}},"directions":{"XR-L":{"primary":["CLARIFIER","GPS","CLARITI"],"more":["POINTER","IGCLAR"],"focus":"Proactive clarification for XR Tutorials：已知需要澄清，GT 在用户那里；只研究 how。","boundary":"Agent 自己选择并发起问题、对象高亮、候选对照或确认界面；用户提供 GT。主动感知未知物理事实、何时提问和跨任务数据选择均不是当前主问题。"},"XR-S":{"primary":["SC2","SPATIA","COME"],"more":["SPACE","PERMA"],"focus":"区分内部自检、外部反证、真实世界变化与观测噪声，再更新记忆和后续教程。","boundary":"优先研究外部反证下的选择性修订。导航纠错、空间记忆更新和历史保持分别是近邻，不能合起来声称问题已解决。"},"XR-C":{"primary":["VOID","CAUSAL","MULTI"],"more":["GIVE"],"focus":"对动作、步骤或机制实施干预，传播必要后果，并保留无关内容。","boundary":"VOID 是视频直接先例；Causal-Adapter 和 Multiverse Mechanica 提供两篇 AI 会议方法／评测支撑，分别属于图像与游戏机制。"},"XR-E":{"primary":["ARIFY","ACE","MCE"],"more":["COACH","DARWIN","SILVR","JARVIS","VISTA","DGM","CINE","TUTOR"],"focus":"跨任务保留并改善教学策略：从学员执行反馈中更新可复用经验，并检验新任务迁移与旧任务回归。","boundary":"ARify／AgentCoach 对应教学端；ACE／MCE 对应持续改进机制；VISTA 的单任务修补与 DGM 的代码自修改分别列作不同距离的近邻。"},"AIGC-A":{"primary":["VOID","GIVE","CAUSAL"],"more":["MULTI"],"focus":"优先阅读：视频后果传播 → 几何条件执行 → 因果属性干预与保持。","boundary":"[推断] 按任务接近度选三篇，允许明确标记的预印本；第三篇是图像方法，不冒充视频证据。"},"AIGC-B":{"primary":["PERMA","SPACE","SC2"],"more":["SPATIA","COME","VIS"],"focus":"优先阅读：失效记忆怎样更新 → 外来反证怎样揭示信念惯性 → 怎样闭环修订状态／模型。","boundary":"[推断] 三篇分别覆盖生成记忆、反证评测、纠错方法；本论文池内未见它们完整实现外来反证驱动的选择性教程视频修复。"},"AIGC-C":{"primary":["SUSP","VIS","ONE"],"more":[],"focus":"优先阅读：揭示时间的任务定义 → 事实与已呈现历史的约束 → 跨镜头记忆执行。","boundary":"[推断] 这是互补阅读组合，不能说三篇都直接研究 video revelation。保留 EACL 2024，因为它比普通 2026 长视频论文更接近核心变量。"},"AIGC-D":{"primary":["QWEN","VGEN","GPS"],"more":["CLARIFIER","CLARITI"],"focus":"优先阅读：context gap → 视频工具执行 → 面向缺失条件的高效询问。","boundary":"[推断] 前两篇是预印本，GPS 是 ICLR 2026。D 仍研究广义信息与控制获取；XR-L 是 GT 在用户处的窄子问题，不能直接合并。"}}}}