世界模型与评测 · World Models and Evaluation

LeWorldModel 技术谱系:JEPA 潜空间世界模型到规划控制

LeWorldModel: From JEPA to Planning and Control

更新于

30 秒结论

先读这个判断

本文聚焦 LeWorldModel(LeWM)及其最直接的技术前序、后续修补和反例,而不把所有带有 world model 名称的工作混在一起。检索窗口为 2026 年 1 月 24 日至 7 月 24 日,并为理解技术来源纳入窗口外的 V-JEPA 2 与 LeJEPA。本轮共获得 193 篇去重候选、38 篇可读完整正文,最终以 24 篇逐篇精读论文支撑判断。

精读论文
24
正式证据事件
76
知识单元
EA-MODEL, EA-EVAL, EA-4D
审计状态
当前结算 run 已通过仓库校验

v0 · 完整正文

研究综述

研究边界

本文聚焦 LeWorldModel(LeWM)及其最直接的技术前序、后续修补和反例,而不把所有带有 world model 名称的工作混在一起。检索窗口为 2026 年 1 月 24 日至 7 月 24 日,并为理解技术来源纳入窗口外的 V-JEPA 2 与 LeJEPA。本轮共获得 193 篇去重候选、38 篇可读完整正文,最终以 24 篇逐篇精读论文支撑判断。

分析对象不是单一预测指标,而是从像素输入到真实动作之间的整条链:表征能否稳定联合训练,潜动态是否对动作敏感,多步滚动是否覆盖规划时域,潜空间代价是否对应真实可达性,规划与控制又能否在计算预算内改善任务结果。DINOv3 一类静态视觉基础表征只作为“冻结通用编码器”的对照边界;MoE、统一模型和一般 VLA 路线不承担本文核心论证。

中心判断

LeWM 的贡献不是简单地把世界模型做小,而是证明了一条极简链路可以工作:从像素联合训练编码器和动作条件预测器,用一个预测损失加一个防坍塌正则学习潜状态,再在该潜空间中以 CEM/MPC 搜索短动作序列。这使“自监督视觉表征—潜动态—规划控制”第一次以很少的训练部件连成一个可复现基线。

但 2026 年紧随其后的研究也给出更重要的结论:稳定、可预测、动作敏感、可达、可搜索和闭环有效是六个不同命题。 LeWM 主要把前两三项接了起来,后续工作则逐层暴露并修补剩余接口。因此,评价这条路线不能只问 latent prediction loss 是否下降,也不能只看生成结果是否自然;必须检查每一层是否满足下一层所需的契约。

最短技术链:LeJEPA 的正则如何进入 LeWM

LeJEPA 的出发点是简化联合嵌入预测学习。它把多视图预测与 SIGReg 结合,移除 prototypes、teacher–student、EMA 和 stop-gradient 等常用防坍塌装置。SIGReg 将批量表征投影到多个随机一维方向,再用 Epps–Pulley 统计量把投影分布推向标准高斯;当投影方向足够多时,这相当于约束联合分布接近各向同性高斯。其价值首先是训练稳定和分布塑形,而不是控制本身。LeJEPA 的主要实验没有动作条件动力学、模型滚动或闭环规划,因此它只能解释 LeWM 的正则来源,不能替 LeWM 提供控制证据。

LeWorldModel 将这套思想放进离线动作轨迹。设像素观测经编码器得到 \(z_t=f_\theta(o_t)\),动作条件预测器根据历史潜状态和动作给出下一状态 \(\hat z_{t+1}=g_\phi(z_{\le t},a_t)\)。训练目标可概括为

\[ \mathcal{L}=\mathcal{L}_{pred}+\lambda\mathcal{L}_{SIGReg}. \]

动作通过 AdaLN 注入预测器每一层,预测器自回归推出候选动作序列下的未来潜状态。控制时,CEM 采样并迭代更新动作序列,最小化终端预测潜状态与目标图像潜状态的距离;系统只执行其中一段动作,获得新观测后重新规划。这种 receding-horizon 设计很关键:论文没有声称可以无限滚动,而是明确承认更长 rollout 同时增加计算和模型偏差。

在论文的模拟协议中,LeWM 在 Push-T 上比 PLDM 高 18 个百分点,固定规划配置完成一次完整规划少于一秒;三训练种子的对照也显示 PLDM 的成功率方差更大。不过,LeWM 并非在所有环境领先。简单的 Two-Room 反而表现较差,作者推测低数据多样性、低内在状态维度与高维各向同性先验存在冲突。这说明 SIGReg 是一种带数据条件的正则,不是“潜空间越高斯越适合所有控制任务”的普遍定律。

速度分支:Fast-LeWM 并行的是动力学查询,不是整个规划器

Fast-LeWorldModel 针对 LeWM 规划时必须顺序调用一步 predictor 的瓶颈,引入因果 action-prefix encoder:把同一个已观测状态作为锚点,一次处理多个动作前缀,并行预测不同未来时域的潜状态;训练时对每个前缀提供稠密目标。视觉编码器与 SIGReg 仍被保留,因此它主要改写的是动力学查询方式,而不是换掉表征目标、终端潜距离或 CEM。

在该论文的同协议四任务汇总中,平均成功率由 LeWM 的 85.8% 提升到 Fast-LeWM 的 90.5%,再加 self-conditioning 后为 92.0%,但主表没有跨种子误差条。Two-Room、同一 CEM 预算和单张 RTX 4090 的计时显示,动力学部分由 31.4 秒降到 8.0 秒,完整 CEM 则由 54.4 秒降到 28.3 秒。这个分账说明加速确实来自把五次顺序动力学调用合为一次 prefix pass,同时也说明模型前向之外的采样、精英筛选和分布更新仍占大量时间;28.3 秒的绝对延迟也不能称为实时控制。这里的协议与 LeWM 原论文“少于一秒”的配置不同,数字不能横向直接相减。超过单个最大 prefix window 的更长时间点仍要组合两次预测,而 naive 的长动作输入若缺少稠密前缀监督和 state token,并不能保持同样质量。

第一份独立契约:预测必须对动作敏感

动作作为网络输入,并不保证网络真正使用动作。若相邻帧变化平滑,模型可能主要靠当前状态预测下一状态,同时把不同动作造成的差异压在很小的潜位移里。这样的模型可以有较低预测误差,却会让 CEM 难以区分候选动作。

Delta-JEPA 对此给出直接修补:除预测未来潜状态外,再要求仅从潜状态差分 \(z_{t+1}-z_t\) 恢复已执行动作。这个 latent-difference action decoding 迫使局部转移几何携带动作信息。固定训练和规划协议后,它在四个模拟环境都优于从两个端点拼接后解码动作的版本;可视化也显示,当起始历史不变而只改变动作时,Delta-JEPA 的预测响应按动作分开,LeWM 的响应则更集中、更重叠。

然而动作敏感仍不是任务充分性。Predictive Objectives Discard Exogenous Control-Relevant Features 构造了一个受控反例:若某个与奖励相关的状态特征在时间上不可预测,纯粹的无奖励潜预测目标没有理由保留它。增加潜容量不能自动补回这个信号;逆动力学只有在行为策略的动作与该特征相关时才可能间接保留它。这个结果来自小型合成环境,不能直接宣布大型机器人模型也会失败,但它明确否定了一个常见推理——“把所有可预测因素建模好,就等于保留了控制所需的一切”。

第二份独立契约:潜距离必须对应有限预算内的可达性

LeWM 的默认规划接口把终端潜状态到目标潜状态的欧氏距离当作代价。这个选择计算简单,却暗含很强的几何假设:潜空间中的近邻,必须在动力学约束和剩余步数下也容易到达。

Beyond Euclidean Proximity 用 Two-Room 给出醒目的反例。只替换终端 selector、保持世界模型不变,full-horizon trajectory reachability metric 就把 LeWM 的平均成功率从 7.0% 提升到 97.0%。更有解释力的是,和二维位置相关的一个低维 rowspace 只贡献了终端—目标潜均方误差的 0.5%–0.7%,却能让规划成功率达到 90.8%;剩余的大部分潜误差几乎不帮助穿过门洞。这表明“潜空间中占能量最多的方向”与“规划中最重要的方向”可能完全不同。

RC-aux 进一步把时间和方向纳入接口。它一方面训练多时域开放环预测,缩小单步训练与多步规划的错配;另一方面学习预算条件、方向敏感的轨迹可达性,并在测试时把它作为搜索信号。在 Wall 任务中,匹配比较的成功率由 50.4% 提升到 83.6%。但它的可达性标签仍来自数据轨迹,只是现实可行性的代理;未观察到的捷径、不确定性和完整约束还没有解决。

这两项工作把问题从“表征是否好”改写为“表征向规划器暴露了什么代价”。它们也解释了为什么仅比较预测损失常常无法预测控制结果:损失平均了所有潜维度,而规划依赖的可能只是少数方向、特定时域和有向可达关系。

第三份独立契约:更快的规划仍要保留闭环含义

Latent Geometry Beyond Search 尝试把 LeWM 上反复运行的 CEM 摊销掉。它冻结 LeWM,用离线轨迹训练目标条件逆动力学模型,根据当前潜状态、目标潜状态和剩余时域直接输出下一动作;执行后重新编码真实观测,因此仍是闭环控制。在论文八个“环境—协议”单元中,它有七个达到或超过 CEM,Push-T 是例外。

这不是无成本地消除规划,而是把重复在线搜索换成环境内离线监督。论文关于“各向同性 LeWM 几何让逆映射更好学”的理论解释也明确带条件:作者没有验证 LeWM 完全满足所列几何假设,也没有设置 non-isotropic 对照。Push-T 的成功率还会在测试预算从 25 步增至 100 步时由 93.0% 降至 75.0%,与超出训练范围后的 horizon clamp 一致。对于不可逆的多步决策,局部逆映射不能显式比较长期后果,仍可能需要偶发搜索或其他纠错机制。

第四份独立契约:层级结构不能自动制造长程能力

短时域模型自然会引出“增加高层子目标”的方案。但 Mind the Gap 表明,oracle 轨迹中的中间目标可执行,不代表模型生成的子目标同样可靠。生成 subgoal 可能时间错位,也可能落到低层控制器没有见过的支持分布之外。论文中的经验宏动作 CEM 通过从训练轨迹的 macro-action bank 采样锚点、只在其附近优化 residual,能改善部分设置;收益却依赖时间尺度和执行方式,最长 Push-T 时域中在线受约束重规划优于 staged execution。

因此,长程能力至少要分别评估子目标质量、低层可执行性、高层搜索支持和执行期纠偏。只报告 oracle subgoal 或只增加 hierarchy 层数,会把不同问题折叠成一个看似漂亮的总体成功率。

条件、分歧与评测边界

潜表征的另一个分歧是“重建细节”与“任务语义”的权衡。Reconstruction or Semantics? 在固定数据、动作条件转移和训练协议下比较两类冻结编码器。在 Bridge V2 的家族级统计中,语义潜空间相对重建潜空间的 VLA 成功率高 9.8 个百分点、OOD 成功率高 13.6 个百分点,一步 CEM 动作误差也更低。与此同时,语义潜变量可能损失精细几何和接触信息,压缩适配器还会在扩散易训练与细粒度控制之间产生取舍。结论应当是“视觉锐利不是控制效用的充分代理”,而不是某个通用视觉模型在所有机器人任务中必然更优。

评测也必须从模型内部指标走到动作后果。MiraBench 强调动作跟随、物理遵循与乐观偏差;长 rollout 还需要时间一致性和足够低的决策延迟。真实机器人、接触丰富任务、视角变化和分布外恢复目前仍是 LeWM 主线的薄弱处。V-JEPA 2-AC 虽然提供跨两座实验室的真实部署证据,但其无动作视频预训练不直接包含动作因果,控制来自随后少量机器人数据上的动作条件训练;约 16 秒的已验证时域、人工子目标和相机位置选择也限制了外推。

结论与下一步

LeWM 最值得学习的,不是某个单独的速度或成功率数字,而是一条极简、可拆解的研究基线。它把像素、JEPA、动作条件动态和 latent MPC 放进同一个系统,使后续研究可以逐层问出更精确的问题:

  1. SIGReg 是否在当前数据分布中既防坍塌又不扭曲任务几何?
  2. 固定历史、改变动作时,潜状态位移是否可分辨?
  3. 训练时域是否覆盖规划时域,多步误差是否按关键事件而非平均帧统计?
  4. 终端代价是否预测有向、预算条件的真实可达性?
  5. 加速来自减少模型计算、压缩 rollout,还是把在线搜索摊销成额外离线监督?
  6. 改进是否在真实闭环、接触任务、OOD 与失败恢复中仍成立?

下一步最有价值的复现不是简单跑通 LeWM,而是在同一数据和 planner 预算下做分层消融:记录坍塌指标、动作可分性、多时域误差、可达性校准、CEM 搜索效率和闭环成功率,再逐个加入 Delta-JEPA、TRM/RC-aux 或 GC-IDM。这样才能知道收益来自表征、动力学、代价还是控制器,而不是把所有改善归到一个含义过宽的“world model”上。

References

Accepted evidence

去重论文引用

  1. $\texttt{WEAVER}$, Better, Faster, Longer: An Effective World Model for Robotic ManipulationArnav Kumar Jain, Yilin Wu, Jesse Farebrother, Gokul Swamy, Andrea Bajcsy · 2026
  2. $τ_0$-WM: A Unified Video-Action World Model for Robotic ManipulationPengfei Zhou, Shengcong Chen, Di Chen, Jiaxu Wang, Rongjun Jin, Bingwen Zhu, Yike Pan, Songen Gu et al. · 2026
  3. Beyond Euclidean Proximity: Repairing Latent World Models with Horizon-Matched Trajectory Reachability MetricsLiangyu Li, Shengzhi Wang, Qingwen Liu · 2026
  4. ContactWorld: What Matters in Vision-Tactile World Models for Contact-Rich ManipulationZhiyuan Zhang, Pokuang Zhou, Kaidi Zhang, Adeesh Desai, Temitope Amosa, Davood Soleymanzadeh, Jiuzhou Lei, Minghui Zheng et al. · 2026
  5. Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference DecodingZhenghao Zhang, Yuanxiang Wang, Zhenyu Guan, Yujia Yang, Bingkang Shi, Tianyu Zong, Hongzhu Yi, Guoqing Chao et al. · 2026
  6. DreamDojo: A Generalist Robot World Model from Large-Scale Human VideosShenyuan Gao, William Liang, Kaiyuan Zheng, Ayaan Malik, Seonghyeon Ye, Sihyun Yu, Wei-Cheng Tseng, Yuzhu Dong et al. · 2026
  7. Fast LeWorldModelYuntian Gao, Xiangyu Xu · 2026
  8. Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent PlanningChi-Pin Huang, Yunze Man, Zhiding Yu, Min-Hung Chen, Jan Kautz, Yu-Chiang Frank Wang, Fu-En Yang · 2026
  9. GaussianDream: A Feed-Forward 3D Gaussian World Model for Robotic ManipulationZijian Zhang, Yuqing Jiang, Qian Cheng, Xiaofan Li, Si Liu, Ding Zhao, Ping Luo, Weitao Zhou et al. · 2026
  10. GEM-4D: Geometry-Enhanced Video World Models for Robot ManipulationKaichen Zhou, Yuzhen Chen, Fangneng Zhan, Hang Hua, Grace Chen, Xinhai Chang, Ao Qu, Yilun Du et al. · 2026
  11. Inference-time Policy Steering via Vision and TouchYilin Wu, Zilin Si, Zeynep Temel, Oliver Kroemer, Andrea Bajcsy · 2026
  12. Interactive World Simulator for Robot Policy Training and EvaluationYixuan Wang, Rhythm Syed, Fangyu Wu, Mengchao Zhang, Aykut Onol, Jose Barreiros, Hooshang Nayyeri, Tony Dear et al. · 2026
  13. Latent Geometry Beyond Search: Amortizing Planning in World ModelsHoang Nguyen, Xiaohao Xu, Xiaonan Huang · 2026
  14. LeJEPA: Provable and Scalable Self-Supervised Learning Without the HeuristicsRandall Balestriero, Yann LeCun · 2025
  15. LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from PixelsLucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, Randall Balestriero · 2026
  16. Mind the Gap: Promises and Pitfalls of Hierarchical Planning in LeWorldModelNiccolò Caselli, Francesco Massafra, Samuele Punzo, Salvatore Lo Sardo, Ippokratis Pantelidis, Sathya Kamesh Bhethanabhotla · 2026
  17. MiraBench: Evaluating Action-Conditioned Reliability in Robotic World ModelsTianzhuo Yang, Zihan Shen, Zirui Mi, Zhaoyi Zhang, Jiayi Zhou, Jiaming Ji, Juntao Dai, Jiawei Chen et al. · 2026
  18. Predictive but Not Plannable: RC-aux for Latent World ModelsWenyuan Li, Guang Li, Keisuke Maeda, Takahiro Ogawa, Miki Haseyama · 2026
  19. Predictive Objectives Discard Exogenous Control-Relevant Features: A Controlled Mechanistic StudyAyan Pendharkar · 2026
  20. Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World ModelsNilaksh, Saurav Jha, Artem Zholus, Sarath Chandar · 2026
  21. SKIP: Sparse Keyframe Interpolation Paradigm for Efficient Embodied World ModelsZiheng He, Yixiang Chen, Ning Yang, Zhanqian Wu, Qisen Ma, Yuan Xu, Jiabing Yang, Peiyan Li et al. · 2026
  22. V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and PlanningMahmoud Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Mojtaba Komeili, Matthew Muckley, Ammar Rizvi et al. · 2025
  23. World Model for Robot Learning: A Comprehensive SurveyBohan Hou, Gen Li, Jindou Jia, Tuo An, Xinying Guo, Sicong Leng, Haoran Geng, Yanjie Ze et al. · 2026
  24. World Pilot: Steering Vision-Language-Action Models with World-Action PriorsZefu Lin, Rongxu Cui, Junjia Xu, Xiaojuan Jin, Wenling Li, Lue Fan, Zhaoxiang Zhang · 2026
完整证据附录(76 条正式证据事件)
  • Time range: 2026-01-24..2026-07-24
  • Events: 76
  • 每个事件一节,标题即锚点;trace-map 中的 event 链接跳转到这里。

EA-JIMFAN-READ-0015

  • Claim: DreamDojo uses large-scale egocentric human video and continuous latent actions to pretrain a robot world model, then reports bounded policy-evaluation and planning benefits after robot post-training.
  • Stance: conditional | Confidence: direct
  • Paper: 2602.06949 DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos
  • Locator: 4.7 Downstream Applications
  • Evidence: Its policy-evaluation correlation is measured on 20 fruit-packing scenes; the paper also acknowledges optimistic simulation and coverage limitations.
  • Authors: shenyuan-gao; william-liang; kaiyuan-zheng; et al.

EA-WMDATA-READ-0007

  • Claim: A robot world model can be trained from a moderate-sized robot interaction dataset and then used to generate policy-training demonstrations, but its value depends on interaction-consistent long-horizon rollouts and sim-real correlation.
  • Stance: support | Confidence: direct
  • Paper: 2603.08546 Interactive World Simulator for Robot Policy Training and Evaluation
  • Locator: IV-C Data Generation for Policy Training
  • Evidence: The paper builds an Interactive World Simulator from a moderate-sized robot interaction dataset, reports world-model-generated policy data comparable to the same amount of real-world data, and evaluates sim-real performance correlation.
  • Authors: yixuan-wang; rhythm-syed; fangyu-wu; et al.

EA-WMDATA-READ-0008

  • Claim: Robot videos can be converted into richer world-model supervision by pairing RGB with depth and pseudo 3D scene-flow targets, training models to represent current 3D structure and short-horizon future evolution rather than only behavior-cloning actions.
  • Stance: support | Confidence: direct
  • Paper: 2605.20752 GaussianDream: A Feed-Forward 3D Gaussian World Model for Robotic Manipulation
  • Locator: 3.4 GaussianDream Training and Efficient Inference
  • Evidence: GaussianDream trains current Gaussian reconstruction and future Gaussian prediction heads with RGB rendering, depth, and pseudo 3D scene-flow supervision, then retains only a compact prefix for control at inference.
  • Authors: zijian-zhang; yuqing-jiang; qian-cheng; et al.

EA-TWM-READ-0003

  • Claim: 在接触丰富操作中,触觉世界模型的关键不是简单增加模态,而是让表征同时具备空间结构、时间连续性和跨模态兼容性。
  • Stance: support | Confidence: direct
  • Paper: 2606.13877 ContactWorld: What Matters in Vision-Tactile World Models for Contact-Rich Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: ContactWorld 在 12 个接触丰富任务上比较视觉与触觉表征;点云把平均规划成功率从腕部视角 20.7% 和前视 22.0% 提升到 32.1%,点云加触觉力场进一步到 36.1%。作者强调触觉效果取决于跨模态表征兼容,而非模态数量本身。
  • Authors: zhiyuan-zhang; pokuang-zhou; kaidi-zhang; et al.

EA-TWM-READ-0004

  • Claim: 触觉世界模型也可以在推理期作为候选动作验证器,而不只是训练期的动态模型。
  • Stance: support | Confidence: direct
  • Paper: 2606.14981 Inference-time Policy Steering via Vision and Touch
  • Locator: 5 Experiments
  • Evidence: ViTaL 学习 visuo-tactile latent world model,结合视觉和文本条件触觉 verifier,对候选动作进行长时域视觉模式选择和短时域触觉 refinement;真实机器人任务包括 wiping、insertion 和 pipette transfer。
  • Authors: yilin-wu; zilin-si; zeynep-temel; et al.

EA-WMEVAL-READ-0005

  • Claim: GEM-4D supports geometry-feature distillation as a way to make video world models more actionable for robot manipulation without adding inference-time cost.
  • Stance: support | Confidence: direct
  • Paper: 2605.22882 GEM-4D: Geometry-Enhanced Video World Models for Robot Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: The model distills 4D geometry foundation-model representations into a video backbone during training, discards the geometry branch at inference, and uses an inverse dynamics module to convert generated rollouts into executable trajectories; the paper reports real-world manipulation success improving from 61% to 81%.
  • Authors: kaichen-zhou; yuzhen-chen; fangneng-zhan; et al.

EA-WMEVAL-READ-0003

  • Claim: Training data for efficient embodied world-model rollouts must preserve sparse task-relevant events such as approach, contact, grasp, and release; generic frame dropping can remove the information downstream policies need.
  • Stance: support | Confidence: direct
  • Paper: 2606.00664 SKIP: Sparse Keyframe Interpolation Paradigm for Efficient Embodied World Models
  • Locator: Abstract (full-text section)
  • Evidence: SKIP argues that manipulation rollouts concentrate task-relevant information in sparse events, selects event-preserving keyframes through robot-aware multimodal fusion, and reports that generated videos can serve as policy-training data.
  • Authors: ziheng-he; yixiang-chen; ning-yang; et al.

EA-WMEVAL-READ-0001

  • Claim: τ0-WM 使用真实机器人遥操作、UMI 式交互、人类第一视角视频与 rollout/失败轨迹的异构语料,并用按模态的监督掩码训练统一视频—动作世界模型。
  • Stance: support | Confidence: direct
  • Paper: 2606.01027 $τ_0$-WM: A Unified Video-Action World Model for Robotic Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: 摘要直接报告了异构数据组成与 modality-specific supervision masks。
  • Authors: pengfei-zhou; shengcong-chen; di-chen; et al.

EA-WMEVAL-READ-0010

  • Claim: WEAVER defines useful robot world models by three joint requirements: fidelity to real outcomes, temporal consistency over long horizons, and enough efficiency for evaluation, improvement, and planning.
  • Stance: support | Confidence: direct
  • Paper: 2606.13672 $\texttt{WEAVER}$, Better, Faster, Longer: An Effective World Model for Robotic Manipulation
  • Locator: 3 WEAVER : World Estimation Across Views for Embodied Reasoning
  • Evidence: The paper argues that manipulation world models must satisfy fidelity, consistency, and efficiency together, then designs a multi-view latent world model with reward/value prediction to support policy evaluation, synthetic policy improvement, and test-time planning.
  • Authors: arnav-kumar-jain; yilin-wu; jesse-farebrother; et al.

EA-WMEVAL-READ-0013

  • Claim: Static image-text pretraining is insufficient training signal for contact-rich manipulation; world-model data should expose action-conditioned scene evolution and contact dynamics.
  • Stance: limit | Confidence: direct
  • Paper: 2606.12403 World Pilot: Steering Vision-Language-Action Models with World-Action Priors
  • Locator: Abstract (full-text section)
  • Evidence: World Pilot argues that VLA semantic grounding from static image-text pairs cannot capture continuous contact-rich dynamics, and uses WAM-derived scene-evolution and trajectory priors to complement the policy.
  • Authors: zefu-lin; rongxu-cui; junjia-xu; et al.

EA-WMEVAL-READ-0004

  • Claim: Robotic world-model evaluation should move beyond visual fidelity toward action-conditioned reliability, including physical adherence, action-following fidelity, and optimism-bias detection.
  • Stance: gap | Confidence: direct
  • Paper: 2605.29360 MiraBench: Evaluating Action-Conditioned Reliability in Robotic World Models
  • Locator: Abstract (full-text section)
  • Evidence: The paper frames existing evaluations as weak evidence for whether action-conditioned predictions are reliable, then defines MiraBench around physics adherence, action fidelity, and failure-case optimism bias.
  • Authors: tianzhuo-yang; zihan-shen; zirui-mi; et al.

EA-LEWM-READ-0081

  • Claim: V-JEPA 2 的无动作视频预训练预测不直接包含机器人动作的因果效应;论文是在冻结视频编码器上另行训练帧因果、动作条件预测器,才把表征接到规划。
  • Stance: support | Confidence: direct
  • Paper: 2506.09985 V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
  • Locator: 3 V-JEPA 2-AC: Learning an Action-Conditioned World Model
  • Evidence: 方法章节明确说明预训练后的缺失视频预测不考虑智能体动作因果,并把小规模交互数据后训练定义为使模型可用于规划的下一阶段。
  • Authors: mahmoud-assran; adrien-bardes; david-fan; et al.

EA-LEWM-READ-0082

  • Claim: V-JEPA 2 的表征目标是预测场景中可预测的方面,而不是复原生成目标强调的所有像素级不可预测细节。
  • Stance: support | Confidence: direct
  • Paper: 2506.09985 V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
  • Locator: 1 Introduction
  • Evidence: 引言把 JEPA 的可预测表征目标与像素生成目标并列,并用运动轨迹和草叶细节说明选择性。
  • Authors: mahmoud-assran; adrien-bardes; david-fan; et al.

EA-LEWM-READ-0084

  • Claim: 在论文展示的三个单目标 reaching 实例中,V-JEPA 2-AC 把末端移动到距目标不足 4 cm,并选择使误差单调下降的动作。
  • Stance: support | Confidence: direct
  • Paper: 2506.09985 V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
  • Locator: 4.2 Results
  • Evidence: 结果章节用末端目标距离报告三个单目标 reaching 执行,并同时观察到误差单调下降。
  • Authors: mahmoud-assran; adrien-bardes; david-fan; et al.

EA-LEWM-READ-0026

  • Claim: LeJEPA selects an Epps–Pulley empirical-characteristic-function regularizer because the implementation is DDP-friendly, has uniformly bounded gradients and curvature, and has linear memory and computational complexity.
  • Stance: support | Confidence: direct
  • Paper: 2511.08544 LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics
  • Locator: 4.2.3 Characteristic Functions are Stable, Scalable and Identifiable
  • Evidence: Section 4.2.3 明确比较统计检验家族后列出 Epps–Pulley 的 distributed、bounded 与 linear-complexity 性质。
  • Authors: randall-balestriero; yann-lecun

EA-LEWM-READ-0027

  • Claim: The LeJEPA objective combines the multi-view prediction loss with SIGReg and removes prototypes, stop-gradients, and teacher–student networks, leaving one coefficient to trade prediction against isotropic-Gaussian regularization.
  • Stance: support | Confidence: direct
  • Paper: 2511.08544 LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics
  • Locator: 5.1 LeJEPA: SIGReg + Prediction Loss
  • Evidence: Section 5.1 给出最终组合损失并明确列出不再需要的 collapse-prevention 组件和 mixing coefficient 的含义。
  • Authors: randall-balestriero; yann-lecun

EA-LEWM-READ-0029

  • Claim: On ImageNet-10, LeJEPA pretrained approximately 50 architectures from eight families with fewer than 20 million parameters, and all reached 91.5%–95% top-1 accuracy under frozen-backbone linear probing.
  • Stance: support | Confidence: direct
  • Paper: 2511.08544 LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics
  • Locator: 6.1 LeJEPA’s Stability Across Hyper-Parameters and Architectures
  • Evidence: Section 6.1 直接给出模型数量、架构族、参数上限和 frozen linear-probe top-1 区间。
  • Authors: randall-balestriero; yann-lecun

EA-ALIGN-READ-0013

  • Claim: 长程规划、失败后自我纠正、少样本适应属于认知层能力;显式 CoT 能提升它们但推理延迟高,认知处理可以压缩为 latent 推理而保持规划与恢复能力。
  • Stance: support | Confidence: direct
  • Paper: 2601.09708 Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent Planning
  • Locator: 5 Conclusion
  • Evidence: 论文指出 VLA 靠动作监督擅长基本技能,但在长程规划、失败自我纠正、新场景适应上泛化差;Fast-ThinkAct 用 preference-guided 蒸馏把冗长文本推理压缩为紧凑 latent CoT,在保持 long-horizon planning、few-shot adaptation 和 failure recovery 的同时推理延迟最多降 89.3%。
  • Authors: chi-pin-huang; yunze-man; zhiding-yu; et al.

EA-LEWM-READ-0001

  • Claim: LeWorldModel jointly trains a pixel encoder and an action-conditioned latent predictor; actions enter the predictor through AdaLN at every layer, and the predictor autoregressively forecasts the next latent from a history of frame representations.
  • Stance: support | Confidence: direct
  • Paper: 2603.19312 LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels
  • Locator: 3 Method: LeWorldModel
  • Evidence: 方法段明确描述 encoder、predictor、AdaLN 动作注入、causal autoregression,并说明所有组件联合学习。
  • Authors: lucas-maes; quentin-le-lidec; damien-scieur; et al.

EA-LEWM-READ-0002

  • Claim: SIGReg regularizes LeWM latents toward an isotropic Gaussian by projecting them onto random unit directions and matching each one-dimensional projection with an Epps–Pulley normality statistic; the joint-distribution argument is asymptotic in the number of projections.
  • Stance: support | Confidence: direct
  • Paper: 2603.19312 LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels
  • Locator: Appendix A SIGReg
  • Evidence: SIGReg 附录明确给出随机方向、一维 Epps–Pulley 匹配、Cramér–Wold 连接及其渐近限定。
  • Authors: lucas-maes; quentin-le-lidec; damien-scieur; et al.

EA-LEWM-READ-0004

  • Claim: In the paper's simulated benchmark setup, LeWM achieved an 18% higher Push-T success rate than PLDM, while the fixed planning configuration completed a full plan in under one second; this did not translate into dominance on every environment.
  • Stance: support | Confidence: direct
  • Paper: 2603.19312 LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels
  • Locator: 4.2 Towards Efficient Planning with WMs, Figure 3 and Figure 6
  • Evidence: 4.2 同时报告 Push-T 相对 PLDM 的成功率差、低于一秒的完整规划,以及 Two-Room 反例。
  • Authors: lucas-maes; quentin-le-lidec; damien-scieur; et al.

EA-LEWM-READ-0005

  • Claim: The paper's direct cross-seed stability check is narrow: on Push-T, across three training seeds evaluated on the same 50 trajectories, PLDM exhibited higher success-rate variance than LeWM.
  • Stance: support | Confidence: direct
  • Paper: 2603.19312 LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels
  • Locator: Appendix G Ablations., Table 5
  • Evidence: Appendix G Table 5 报告三 seed、同一 50 条 Push-T 轨迹的 success rate 与方差,并明确 PLDM 方差更高。
  • Authors: lucas-maes; quentin-le-lidec; damien-scieur; et al.

EA-ALIGN-READ-0014

  • Claim: 纯反应式 VLA 在复杂物理环境中仍受长时程推理、时序归因和误差累积限制,这构成引入显式预测结构的主要动机。
  • Stance: support | Confidence: direct
  • Paper: 2605.00080 World Model for Robot Learning: A Comprehensive Survey
  • Locator: 1 Introduction
  • Evidence: 引言直接将纯反应 VLA 的三类困难列为长时程推理、temporal credit assignment 与 compounding errors。
  • Authors: bohan-hou; gen-li; jindou-jia; et al.

EA-LEWM-READ-0087

  • Claim: 该研究固定数据、历史长度、动作条件、转移架构、优化器和训练日程,仅改变冻结编码器、可选适配器与解码路径,从而把编码器定义的潜接口作为主要实验变量。
  • Stance: support | Confidence: direct
  • Paper: 2605.06388 Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models
  • Locator: 3.1 Dataset and Training
  • Evidence: 训练协议章节逐项列出固定量和变化量,并为每个编码器—适配器组合从头训练 LDM。
  • Authors: nilaksh; saurav-jha; artem-zholus; et al.

EA-LEWM-READ-0041

  • Claim: RC-aux 通过多时域开放环预测修正时间错配,并以预算条件、方向敏感的轨迹可达性监督修正空间几何;测试时还可按剩余预算把可达性作为显式搜索信号。
  • Stance: support | Confidence: direct
  • Paper: 2605.07278 Predictive but Not Plannable: RC-aux for Latent World Models
  • Locator: 1 Introduction
  • Evidence: 引言明确把 RC-aux 分成时间与空间两条监督轴,并说明测试时规划器如何使用可达性信号。
  • Authors: wenyuan-li; guang-li; keisuke-maeda; et al.

EA-LEWM-READ-0042

  • Claim: 在论文的五个匹配 LeWM-family 比较中,RC-aux 改善四项;Wall 成功率从 50.4±6.5 提升到 83.6±3.6,即提高 33.2 个百分点。
  • Stance: support | Confidence: direct
  • Paper: 2605.07278 Predictive but Not Plannable: RC-aux for Latent World Models
  • Locator: 4.2 Main Results
  • Evidence: 表 1 同时给出 LeWM、RC-aux 和 matched delta,正文说明 Wall 是最大增益且该任务没有 continuation checkpoint。
  • Authors: wenyuan-li; guang-li; keisuke-maeda; et al.

EA-LEWM-READ-0021

  • Claim: GC-IDM freezes the LeWorldModel representation and amortizes planning into a horizon-conditioned inverse map from current and goal latents to the next action; at test time it re-encodes the actual observation each step and performs no latent rollout or online search.
  • Stance: support | Confidence: direct
  • Paper: 2605.08732 Latent Geometry Beyond Search: Amortizing Planning in World Models
  • Locator: 4 Method: Goal-Conditioned Inverse Dynamics Model (GC-IDM)
  • Evidence: 方法图和正文明确给出冻结 LeWM embeddings、随机 horizon 训练元组、只更新 inverse module,以及逐步重编码/单次前向的推理路径。
  • Authors: hoang-nguyen; xiaohao-xu; xiaonan-huang

EA-LEWM-READ-0022

  • Claim: On the eight environment–protocol cells in the main LeWM benchmark comparison, GC-IDM matched or outperformed CEM in seven, with Push-T as the lone exception.
  • Stance: support | Confidence: direct
  • Paper: 2605.08732 Latent Geometry Beyond Search: Amortizing Planning in World Models
  • Locator: 5.3 Main Results
  • Evidence: 主结果按 environment–protocol cell 汇总,明确报告 seven of eight,并指出 Push-T 是唯一例外。
  • Authors: hoang-nguyen; xiaohao-xu; xiaonan-huang

EA-LEWM-READ-0046

  • Claim: 在固定 hard n100 TwoRoom 协议中,只替换终端 selector 的 full-horizon TRM 将 LeWM 平均成功率从 7.0% 提升到 97.0%,将本地 PLDM 从 32.7% 提升到 84.0%;同架构 shuffled-label heads 在两类模型上均为 0.0%。
  • Stance: support | Confidence: direct
  • Paper: 2605.22164 Beyond Euclidean Proximity: Repairing Latent World Models with Horizon-Matched Trajectory Reachability Metrics
  • Locator: 6.1 TRM Repairs Hard TwoRoom in Fixed Terminal Selection
  • Evidence: 6.1 节声明所有 checkpoint、cache、manifest、CEM 与候选采样保持不变,并在表 1 中报告真实时间标签与 shuffled control。
  • Authors: liangyu-li; shengzhi-wang; qingwen-liu

EA-LEWM-READ-0047

  • Claim: 在 LeWM seed 3072 的 matched-b50 消融中,平衡满 episode TRM 使用 100,000 pairs 达到 97.5% 成功率,而相同 pair 预算的平衡短时域 max 版本仅 35.0%。
  • Stance: support | Confidence: direct
  • Paper: 2605.22164 Beyond Euclidean Proximity: Repairing Latent World Models with Horizon-Matched Trajectory Reachability Metrics
  • Locator: 6.2 Horizon Matching and Temporal Structure Matter
  • Evidence: 6.2 节与附录 C.5 在相同 pair 预算下比较 full-episode 和 max-horizon,并把 horizon matching 识别为最强因素。
  • Authors: liangyu-li; shengzhi-wang; qingwen-liu

EA-LEWM-READ-0048

  • Claim: 在 matched-b50 TwoRoom 中,XY probe rowspace 只贡献 0.5–0.7% 的终端—目标潜 MSE,却使 rowspace-only planning 达到 90.8% 平均成功率;raw latent MSE 和 residual-only 均为 1.7%。
  • Stance: support | Confidence: direct
  • Paper: 2605.22164 Beyond Euclidean Proximity: Repairing Latent World Models with Horizon-Matched Trajectory Reachability Metrics
  • Locator: 6.3 Mechanism: Raw Latent MSE Hides the Useful State
  • Evidence: 6.3 节的 rowspace surgery 直接保留或移除 XY probe 子空间,且规划时不使用真实 simulator state。
  • Authors: liangyu-li; shengzhi-wang; qingwen-liu

EA-LEWM-READ-0008

  • Claim: Fast-LeWM replaces LeWM's within-window autoregressive latent chain with action-prefix prediction: every future latent is predicted directly from the same observed anchor latent and its causal action prefix, so all queried horizons can be produced in parallel.
  • Stance: support | Confidence: direct
  • Paper: 2606.26217 Fast LeWorldModel
  • Locator: 3.3 Fast LeWorldModel
  • Evidence: 3.3 明确给出 anchor latent、prefix token、dense horizon targets 与并行预测,并说明 future predictions 不再顺序依赖彼此。
  • Authors: yuntian-gao; xiangyu-xu

EA-LEWM-READ-0009

  • Claim: Under the same four-task LeWM planning protocol, base Fast-LeWM increased average success from 85.8% to 90.5%, and optional self-consistency increased it to 92.0%; no environment had lower point-estimate success than LeWM.
  • Stance: support | Confidence: direct
  • Paper: 2606.26217 Fast LeWorldModel
  • Locator: 3.7 Planning with Action Prefixes and Self-Consistency
  • Evidence: Table 1 完整列出四环境和平均成功率,base 与 self-consistency variant 在每列都不低于 LeWM。
  • Authors: yuntian-gao; xiangyu-xu

EA-LEWM-READ-0011

  • Claim: Across all four tasks, Fast-LeWM had lower initial open-loop latent error and a smaller least-squares error-growth slope than LeWM, while qualitative decoded rollouts showed less long-horizon drift.
  • Stance: support | Confidence: direct
  • Paper: 2606.26217 Fast LeWorldModel
  • Locator: 4.3 Open-Loop Latent Prediction
  • Evidence: 4.3 以 ground-truth future-frame latents 对齐,比较 loss curve 起点、增长斜率和 decoder rollout。
  • Authors: yuntian-gao; xiangyu-xu

EA-LEWM-READ-0061

  • Claim: LDAD decodes the executed action from the latent displacement alone, so successful recovery requires local transition geometry to encode the action and encourages different actions to induce distinguishable latent changes for planning.
  • Stance: support | Confidence: direct
  • Paper: 2606.31232 Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference Decoding
  • Locator: Action-Supervised Displacement Mechanism.
  • Evidence: 方法段明确将 action decoder 的唯一输入设为 displacement,并解释 action recovery、transition distinguishability 与 planner candidate comparison 的关系。
  • Authors: zhenghao-zhang; yuanxiang-wang; zhenyu-guan; et al.

EA-LEWM-READ-0062

  • Claim: Across the four evaluated environments, Delta-JEPA had the highest mean planning success; it exceeded the strongest OGB-Cube baseline by 15.14 percentage points and LeWM on Push-T by 4.54 points.
  • Stance: support | Confidence: direct
  • Paper: 2606.31232 Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference Decoding
  • Locator: Planning Performance
  • Evidence: 主结果段给出四环境全部最高及 OGB-Cube、Push-T 的明确百分点差。
  • Authors: zhenghao-zhang; yuanxiang-wang; zhenyu-guan; et al.

EA-LEWM-READ-0063

  • Claim: With the training and planning protocol held fixed, decoding actions from latent displacements improved planning relative to decoding from concatenated endpoint embeddings in every evaluated environment.
  • Stance: support | Confidence: direct
  • Paper: 2606.31232 Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference Decoding
  • Locator: Ablation Study > Displacement-Based Action Decoding.
  • Evidence: 该消融明确说明两变体只在 action-decoder input 上不同,并报告 displacement variant 跨四环境一致改善。
  • Authors: zhenghao-zhang; yuanxiang-wang; zhenyu-guan; et al.

EA-LEWM-READ-0064

  • Claim: When the starting history is fixed and only the action input is varied, Delta-JEPA produces separated action-wise predictor responses, whereas LeWM's responses remain concentrated near the origin and substantially overlap.
  • Stance: support | Confidence: direct
  • Paper: 2606.31232 Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference Decoding
  • Locator: Action-Sensitive Latent Dynamics
  • Evidence: 作者固定 512 个 Two-Room histories 并逐 action 干预 predictor,比较 action-wise mean displacement。
  • Authors: zhenghao-zhang; yuanxiang-wang; zhenyu-guan; et al.

EA-LEWM-READ-0052

  • Claim: Empirical-macro CEM 不在连续潜 macro-action 空间中无约束搜索,而从训练轨迹编码的 macro-action 序列 bank 采样 anchor,并只在其附近拟合局部 residual。
  • Stance: support | Confidence: direct
  • Paper: 2607.12547 Mind the Gap: Promises and Pitfalls of Hierarchical Planning in LeWorldModel
  • Locator: 3.3 Planning-Side Mitigations
  • Evidence: 3.3 节给出方法动机,附录 C.6 进一步说明每轮重采样 anchor、只 refit residual 并保留零残差候选。
  • Authors: niccol-caselli; francesco-massafra; samuele-punzo; et al.

EA-LEWM-READ-0083

  • Claim: V-JEPA 2-AC 与论文基线使用同一套目标实验室协议部署在两座 Droid 未覆盖的实验室,输入为未标定、低分辨率单目 RGB,相同 V-JEPA 2-AC 权重和推理代码跨实验室复用。
  • Stance: conditional | Confidence: direct
  • Paper: 2506.09985 V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
  • Locator: 4 Planning: Zero-shot Robot Control
  • Evidence: 规划章节明确记录两座新实验室、Franka 平台、未标定单目相机及跨机器人复用权重和推理代码。
  • Authors: mahmoud-assran; adrien-bardes; david-fan; et al.

EA-LEWM-READ-0028

  • Claim: The nonlinear-probe optimality result is conditional on a smooth embedding density with finite Fisher information and covariance, test queries drawn from the training distribution, an isotropic mean-zero target-gradient prior, and a fixed scalar covariance constraint.
  • Stance: conditional | Confidence: direct
  • Paper: 2511.08544 LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics
  • Locator: A.1 kNN Probing
  • Evidence: Appendix A.1 在给出 kNN 唯一最优解前逐项列出 density、query、target function 与 covariance constraints。
  • Authors: randall-balestriero; yann-lecun

EA-LEWM-READ-0003

  • Claim: LeWM planning is explicitly short-horizon and conditional: longer autoregressive rollouts trade additional lookahead for more compute and model bias, so the controller executes only a planned action block before replanning from a new observation.
  • Stance: conditional | Confidence: direct
  • Paper: 2603.19312 LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels
  • Locator: 3.2 Latent Planning
  • Evidence: 规划段直接陈述 horizon 的计算/偏差权衡、误差累积和 MPC 重规划缓解机制。
  • Authors: lucas-maes; quentin-le-lidec; damien-scieur; et al.

EA-LEWM-READ-0088

  • Claim: 在论文的共享 Bridge V2 任务统计中,语义潜空间家族相对重建家族的 VLA 成功率高 9.8 个百分点、OOD 成功率高 13.6 个百分点,并把一步 CEM 动作误差降低 0.0266;三项 95% 配对 bootstrap 区间均不跨零。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.06388 Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models
  • Locator: D.3 Statistical Analyses
  • Evidence: 附录统计以共享任务做家族级配对 bootstrap,同时报告策略成功、OOD 成功和 CEM 动作恢复三个面向控制的指标。
  • Authors: nilaksh; saurav-jha; artem-zholus; et al.

EA-LEWM-READ-0089

  • Claim: 在 DiT-S 的真实编码潜变量上,语义编码器的 IDM Pearson 为 0.772–0.829、轨迹成功分类准确率为 0.903–0.906;重建编码器对应范围为 0.507–0.626 和 0.835–0.868,且动作信息优势大体延续到世界模型生成潜变量。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.06388 Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models
  • Locator: 4.2 Does the latent space affect action recoverability and preservation of task semantics?
  • Evidence: 表 2 同时报告真实编码潜变量与生成潜变量的 IDM 和成功分类结果,正文总结语义家族保留更多动作信息。
  • Authors: nilaksh; saurav-jha; artem-zholus; et al.

EA-LEWM-READ-0090

  • Claim: 在该受控实验中,重建对齐潜空间仍能保持较锐利的局部外观,但语义潜空间通常在全局结构和时序生成指标上更强;因此低层像素质量不足以解释策略效用。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.06388 Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models
  • Locator: 4.3 How does the latent space affect visual fidelity?
  • Evidence: 视觉结果章节同时报告语义模型的结构与视频指标优势和 VAE 类模型的局部外观竞争力。
  • Authors: nilaksh; saurav-jha; artem-zholus; et al.

EA-LEWM-READ-0091

  • Claim: 论文的质性 rollout 显示两类潜空间会以不同方式失败:重建潜变量可能生成外观连贯但任务语义错误的状态,语义潜变量较能保持任务意图,却可能损失几何和接触精度。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.06388 Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models
  • Locator: 4.5 Do reconstruction-aligned and semantic encoders fail differently?
  • Evidence: 失败分析用篮子、毛巾、抽屉和指令变化案例区分任务语义幻觉与几何接触不足。
  • Authors: nilaksh; saurav-jha; artem-zholus; et al.

EA-LEWM-READ-0092

  • Claim: 把高维语义特征压缩到适配器潜空间通常更利于扩散去噪和高层任务完成,但会恶化潜空间 CEM 动作误差、OOD 鲁棒性或点跟踪等精细控制指标。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.06388 Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models
  • Locator: 4.6 Do compressed adapter latents aid semantic encoders further for world modeling?
  • Evidence: 适配器消融把多数指标的改善与 CEM、OOD 和 PCK 的例外并列,并给出压缩损失精细动作信息的解释。
  • Authors: nilaksh; saurav-jha; artem-zholus; et al.

EA-LEWM-READ-0043

  • Claim: Wall 消融中,RC-aux 训练后的模型即使仍用基础终端潜距离规划也达到 72.4±3.6%,高于 LeWM-family control 的 50.4±6.5%;加入可达性感知规划后进一步达到 83.6±3.6%。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.07278 Predictive but Not Plannable: RC-aux for Latent World Models
  • Locator: 4.2 Main Results
  • Evidence: 表 2 在相同任务上并列 control、RC-aux-trained/base planner 与完整 RC-aux,能区分两层干预。
  • Authors: wenyuan-li; guang-li; keisuke-maeda; et al.

EA-LEWM-READ-0044

  • Claim: 在 LIBERO-Goal 的可训练 no-repeat OFT-style 动作头协议中,RC-aux 表征的平均成功率为 0.812,高于 LeWM 表征的 0.712;repeat tuning 后 RC-aux 为 0.864。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.07278 Predictive but Not Plannable: RC-aux for Latent World Models
  • Locator: 4.4 Model Size and Computational Overhead, Table 5
  • Evidence: 表 5 明确区分 matched no-repeat 比较、repeat-tuned 结果和外部 OpenVLA-OFT 参考。
  • Authors: wenyuan-li; guang-li; keisuke-maeda; et al.

EA-LEWM-READ-0023

  • Claim: The paper's explanation that isotropic LeWM geometry makes inverse dynamics well conditioned is explicitly conditional: the authors did not verify that LeWM fully satisfies the stated geometric assumptions and did not ablate against a non-isotropic baseline.
  • Stance: conditional | Confidence: direct
  • Paper: 2605.08732 Latent Geometry Beyond Search: Amortizing Planning in World Models
  • Locator: Appendix B Theoretical Analysis: Isotropy and Inverse-Dynamics Conditioning
  • Evidence: 理论附录开头明确将分析标为 conditional,并把假设验证与 non-isotropic 对照列为未完成。
  • Authors: hoang-nguyen; xiaohao-xu; xiaonan-huang

EA-LEWM-READ-0049

  • Claim: PushT go50 的 true hybrid cost 在论文权重 sweep 中达到 52.7% 平均成功率,高于 raw latent 的 40.0% 和最佳 shuffled hybrid 的 42.7%,但该提升被作者定位为连续接触任务中的边界结果。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.22164 Beyond Euclidean Proximity: Repairing Latent World Models with Horizon-Matched Trajectory Reachability Metrics
  • Locator: C.6 PushT Details
  • Evidence: 附录 C.6 表 19 给出 go50 的 weight sweep;正文 6.4 明确指出 SCSA 排序和 selected distance 的证据比闭环成功更清楚。
  • Authors: liangyu-li; shengzhi-wang; qingwen-liu

EA-LEWM-READ-0010

  • Claim: On Two-Room with the same CEM budget and one NVIDIA 4090, one prefix-model call reduced dynamics time from 31.4s to 8.0s, but unchanged encoding/scoring/data overhead limited the full CEM solve reduction from 54.4s to 28.3s.
  • Stance: conditional | Confidence: direct
  • Paper: 2606.26217 Fast LeWorldModel
  • Locator: 4.2 Planning Performance and Efficiency, Table 2
  • Evidence: Table 2 报告相同 CEM budget、GPU、model calls、dynamics time 与 full solve time;正文说明 full CEM 还含图像编码、评分和数据操作。
  • Authors: yuntian-gao; xiangyu-xu

EA-LEWM-READ-0013

  • Claim: The quality-speed gain depends on structured prefixes and intermediate supervision: simply enlarging LeWM's action block performed poorly, terminal-only Fast-LeWM remained below the dense-prefix model, and conditioning the prefix encoder on the current-state token improved results.
  • Stance: conditional | Confidence: direct
  • Paper: 2606.26217 Fast LeWorldModel
  • Locator: 4.5 Ablation Studies
  • Evidence: 4.5 的三组消融分别否定 naive long-action、terminal-only 和无 state token 的简化方案。
  • Authors: yuntian-gao; xiangyu-xu

EA-LEWM-READ-0069

  • Claim: Inverse-dynamics retention of the exogenous feature depends on the behavior policy: random actions provide no anchor, whereas an informative policy can retain it only because actions correlate with the feature.
  • Stance: conditional | Confidence: direct
  • Paper: 2606.30068 Predictive Objectives Discard Exogenous Control-Relevant Features: A Controlled Mechanistic Study
  • Locator: 3.3 Objectives
  • Evidence: Objectives section 明确区分 random-policy failure 与 informative-policy rescue,并把后者归为 action supervision。
  • Authors: ayan-pendharkar

EA-LEWM-READ-0071

  • Claim: The trained JEPA latent realized substantially less class separation for the exogenous reward-relevant feature than reconstruction and supervised references, but the authors characterize this as a single-seed controlled measurement rather than a theorem or large-model result.
  • Stance: conditional | Confidence: direct
  • Paper: 2606.30068 Predictive Objectives Discard Exogenous Control-Relevant Features: A Controlled Mechanistic Study
  • Locator: 5 Empirical Comparison with Bisimulation Predictions
  • Evidence: Section 5 把 analytical reward-sensitive state distinction与实际 latent separation 对比,并明示 single seed、finite budget、large pretrained generality open。
  • Authors: ayan-pendharkar

EA-LEWM-READ-0065

  • Claim: LDAD is load-bearing but requires balance: removing or weakening its action-reconstruction term yields near-collapse or poor Push-T planning, while an excessively large weight also degrades performance.
  • Stance: conditional | Confidence: direct
  • Paper: 2606.31232 Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference Decoding
  • Locator: Action Reconstruction Weight.
  • Evidence: Push-T weight sweep 同时显示 zero/weak signal 失败、合理区间稳定和 excessive weight 退化。
  • Authors: zhenghao-zhang; yuanxiang-wang; zhenyu-guan; et al.

EA-LEWM-READ-0053

  • Claim: Support-constrained high-level search 的收益取决于执行模式和时间尺度:staged execution 在中等时域最有帮助,但在最长 PushT 时域低于 online constrained replanning。
  • Stance: conditional | Confidence: direct
  • Paper: 2607.12547 Mind the Gap: Promises and Pitfalls of Hierarchical Planning in LeWorldModel
  • Locator: 3.3 Planning-Side Mitigations
  • Evidence: 3.3 节在同一 PushT sweep 中比较 online 与 staged,并明确把最长时域的 staged 退化解释为方法自然边界。
  • Authors: niccol-caselli; francesco-massafra; samuele-punzo; et al.

EA-LEWM-READ-0085

  • Claim: V-JEPA 2-AC 的已验证规划时域约为 16 秒;简单 grasp 和 reach-with-object 可用单一目标图像,但更长程 pick-and-place 若不提供子目标仍需新的建模方法。
  • Stance: limit | Confidence: direct
  • Paper: 2506.09985 V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
  • Locator: 9 Conclusion
  • Evidence: 结论把约 16 秒预测、单目标简单操作和无子目标长时程任务明确分开。
  • Authors: mahmoud-assran; adrien-bardes; david-fan; et al.

EA-LEWM-READ-0086

  • Claim: 未做显式相机标定时,V-JEPA 2-AC 必须从单目图像隐式推断笛卡尔动作坐标轴;机器人基座不可见会导致欠定和预测错误,作者实际人工尝试多个相机位置后才固定配置。
  • Stance: limit | Confidence: direct
  • Paper: 2506.09985 V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
  • Locator: 4.3 Limitations
  • Evidence: 限制章节把相机坐标轴推断、基座不可见和人工选位连成明确的部署失败机制。
  • Authors: mahmoud-assran; adrien-bardes; david-fan; et al.

EA-LEWM-READ-0030

  • Claim: The practical minibatch SIGReg gradient is biased; the authors report the bias as small in their experiments but do not explore unbiased U-statistic or sample-splitting alternatives.
  • Stance: limit | Confidence: direct
  • Paper: 2511.08544 LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics
  • Locator: 5.1 LeJEPA: SIGReg + Prediction Loss
  • Evidence: Section 5.1 明确承认 minibatch-induced bias、经验上影响小,以及未探索两类无偏替代。
  • Authors: randall-balestriero; yann-lecun

EA-LEWM-READ-0031

  • Claim: LeJEPA's empirical validation does not directly test action-conditioned dynamics, world-model rollouts, planning, or closed-loop control; using SIGReg in LeWorldModel is therefore a downstream mechanism transfer rather than direct control evidence from LeJEPA.
  • Stance: limit | Confidence: inference
  • Paper: 2511.08544 LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics
  • Locator: 6 LeJEPA: Empirical Validation
  • Evidence: 经验章节的完整任务清单只包括跨架构/数据稳定性、loss correlation、视觉 in-domain transfer、scaling 和语义分割;六遍结构图未发现动作条件 dynamics 或控制评测。
  • Authors: randall-balestriero; yann-lecun

EA-LEWM-READ-0006

  • Claim: LeWM underperformed on the simple Two-Room environment; the authors hypothesize that low data diversity and low intrinsic dimensionality conflict with matching a high-dimensional isotropic Gaussian prior, making this a potential SIGReg failure condition rather than a universal causal conclusion.
  • Stance: limit | Confidence: direct
  • Paper: 2603.19312 LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels
  • Locator: 4.2 Towards Efficient Planning with WMs
  • Evidence: 正文直接报告 Two-Room 更差,并用 possible explanation/potential limitation 表明几何归因仍是假设。
  • Authors: lucas-maes; quentin-le-lidec; damien-scieur; et al.

EA-LEWM-READ-0007

  • Claim: The authors explicitly bound LeWM to short-horizon planning and sufficiently covering offline action-labeled datasets; low diversity weakens SIGReg, and removing the need for action labels remains future work.
  • Stance: limit | Confidence: direct
  • Paper: 2603.19312 LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels
  • Locator: 6 Conclusion
  • Evidence: 结论限制段同时列出 short horizons、offline coverage、low diversity 和 action-label dependence。
  • Authors: lucas-maes; quentin-le-lidec; damien-scieur; et al.

EA-LEWM-READ-0093

  • Claim: 该研究的结论局限于 Bridge V2 和共享机器人本体;策略内环只评估固定 VLA 在生成 rollout 中的表现,且部分成功判断依赖 VLM,尚未验证策略改进、sim-to-real 或更广本体。
  • Stance: limit | Confidence: direct
  • Paper: 2605.06388 Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models
  • Locator: 7 Future Work and Limitations
  • Evidence: 限制章节连续列出数据与本体范围、固定策略代理和 VLM 评估偏差。
  • Authors: nilaksh; saurav-jha; artem-zholus; et al.

EA-LEWM-READ-0045

  • Claim: RC-aux 的当前可达性监督仍以轨迹派生标签代理真实环境可达性,测试时也只采用简单可达性门控,因此尚未解决未观察捷径、不确定性和完整可行性决策。
  • Stance: limit | Confidence: direct
  • Paper: 2605.07278 Predictive but Not Plannable: RC-aux for Latent World Models
  • Locator: 5 Conclusion
  • Evidence: 结论明确列出轨迹代理和简单门控两项当前限制,并把不确定性感知可达性列为后续方向。
  • Authors: wenyuan-li; guang-li; keisuke-maeda; et al.

EA-LEWM-READ-0024

  • Claim: For Push-T, GC-IDM success peaked at 93.0% with a 25-step evaluation budget and declined to 75.0% at 100 steps, which the authors associate with clamped horizon inputs outside the training-time range.
  • Stance: limit | Confidence: direct
  • Paper: 2605.08732 Latent Geometry Beyond Search: Amortizing Planning in World Models
  • Locator: E.7 Evaluation Budget and Long-Horizon Control
  • Evidence: Table J 的预算扫描提供 Push-T 的完整序列,正文将长预算退化与未在训练中出现的 clamped horizon signal 联系起来。
  • Authors: hoang-nguyen; xiaohao-xu; xiaonan-huang

EA-LEWM-READ-0025

  • Claim: The evaluation is confined mainly to one LeWM backbone and the stable-worldmodel suite, and GC-IDM does not reason about multi-step consequences; cross-world-model transfer and severely irreversible tasks therefore remain open.
  • Stance: limit | Confidence: direct
  • Paper: 2605.08732 Latent Geometry Beyond Search: Amortizing Planning in World Models
  • Locator: Appendix G More Limitations and Future Works
  • Evidence: 限制附录直接列出单 backbone、跨模型架构修改和 reactive-policy 无多步后果推理,并把 hybrid GC-IDM+CEM 留作未来方向。
  • Authors: hoang-nguyen; xiaohao-xu; xiaonan-huang

EA-LEWM-READ-0050

  • Claim: TRM 的时间度量依赖轨迹覆盖,且同 episode 时间标签只是对称标量可达性代理,不是有向或预算条件的目标可达性值。
  • Stance: limit | Confidence: direct
  • Paper: 2605.22164 Beyond Euclidean Proximity: Repairing Latent World Models with Horizon-Matched Trajectory Reachability Metrics
  • Locator: 8 Limitations
  • Evidence: 限制节同时指出低覆盖/短时域显著退化,以及标签在方向性和预算条件上的结构缺失。
  • Authors: liangyu-li; shengzhi-wang; qingwen-liu

EA-LEWM-READ-0012

  • Claim: Fast-LeWM removes sequential dependence only within one maximum encoded prefix window; the paper's longer open-loop evaluation already requires composing two maximum-horizon predictions, so long-range recurrence is reduced rather than eliminated.
  • Stance: limit | Confidence: direct
  • Paper: 2606.26217 Fast LeWorldModel
  • Locator: 4.3 Open-Loop Latent Prediction
  • Evidence: 4.3 明确区分一个 maximum-horizon prediction 与需要两次该预测的更远时间点。
  • Authors: yuntian-gao; xiangyu-xu

EA-LEWM-READ-0067

  • Claim: A reward-relevant feature can be temporally unpredictable, in which case a latent self-prediction objective receives no signal to retain information that a controller still needs.
  • Stance: limit | Confidence: direct
  • Paper: 2606.30068 Predictive Objectives Discard Exogenous Control-Relevant Features: A Controlled Mechanistic Study
  • Locator: 1 Introduction
  • Evidence: 引言把 feature 的不可预测性与 payoff relevance 分离,并明确预测目标会把它视为 noise。
  • Authors: ayan-pendharkar

EA-LEWM-READ-0068

  • Claim: Across both controlled environments, every evaluated reward-free predictive variant left the exogenous control-relevant feature near chance, while the reward-grounded variant recovered it at the one-bit ceiling alongside reconstruction and supervised references.
  • Stance: limit | Confidence: direct
  • Paper: 2606.30068 Predictive Objectives Discard Exogenous Control-Relevant Features: A Controlled Mechanistic Study
  • Locator: 4.1 Reward-free predictive objectives do not retain the exogenous control-relevant feature
  • Evidence: 主结果段按两环境和三 seeds 汇总 near-chance predictive variants 与 ceiling reward-grounded/reference outcomes。
  • Authors: ayan-pendharkar

EA-LEWM-READ-0070

  • Claim: Increasing latent capacity did not rescue cell-4 retention for JEPA, while the reward-grounded variant remained high across the sweep, supporting an objective-structural rather than capacity explanation.
  • Stance: limit | Confidence: direct
  • Paper: 2606.30068 Predictive Objectives Discard Exogenous Control-Relevant Features: A Controlled Mechanistic Study
  • Locator: 4.3 The failure is not a capacity problem
  • Evidence: capacity sweep 显示 JEPA 始终不过 retain threshold,reward-grounded variant 跨容量保持近完美。
  • Authors: ayan-pendharkar

EA-LEWM-READ-0072

  • Claim: The study establishes an objective-level failure mode only in small synthetic environments; it does not test large pretrained models, robotics, standard RL benchmarks, real data, or how often the failure occurs naturally.
  • Stance: limit | Confidence: direct
  • Paper: 2606.30068 Predictive Objectives Discard Exogenous Control-Relevant Features: A Controlled Mechanistic Study
  • Locator: 6.7 Limitations
  • Evidence: Limitations section 明确列出 synthetic-only、无大模型/机器人/Atari/真实数据和非 benchmark/频率结论。
  • Authors: ayan-pendharkar

EA-LEWM-READ-0066

  • Claim: Delta-JEPA does not establish that action-sensitive latents are sufficient for task control: its offline training data contain no rewards, and its two losses supervise only latent prediction and executed-action reconstruction.
  • Stance: limit | Confidence: inference
  • Paper: 2606.31232 Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference Decoding
  • Locator: Problem Formulation
  • Evidence: Problem formulation 明示 reward-free/unknown-policy data;完整 method map 显示总目标仅 prediction + action reconstruction,未测试 exogenous reward-relevant feature retention。
  • Authors: zhenghao-zhang; yuanxiang-wang; zhenyu-guan; et al.

EA-LEWM-READ-0051

  • Claim: 在冻结低层 LeWM 的 Hi-LeWM 中,简单增加时间抽象层并不足以改善长时程控制;数据轨迹上的 oracle 中间 subgoal 通常可执行,而生成 subgoal 更不可靠、时间错位并对高层搜索空间敏感。
  • Stance: limit | Confidence: direct
  • Paper: 2607.12547 Mind the Gap: Promises and Pitfalls of Hierarchical Planning in LeWorldModel
  • Locator: 1 Introduction
  • Evidence: 引言综合了 acting decomposition 的结论,并明确 naive hierarchy 经常低于 flat LeWM。
  • Authors: niccol-caselli; francesco-massafra; samuele-punzo; et al.

EA-LEWM-READ-0054

  • Claim: Hi-LeWM 的正面结论不能外推为 hierarchy 的一般优势:主分析集中于 PushT,VQ macro-actions 尚未充分评估,而且更高容量世界模型可能呈现不同结果。
  • Stance: limit | Confidence: direct
  • Paper: 2607.12547 Mind the Gap: Promises and Pitfalls of Hierarchical Planning in LeWorldModel
  • Locator: 4 Discussion and Conclusion
  • Evidence: 讨论节逐项说明环境范围、未充分探索的 VQ 路线和模型规模条件。
  • Authors: niccol-caselli; francesco-massafra; samuele-punzo; et al.

References

Interactive Wiki

进入交互阅读

静态页采用默认版本;v0、v0、v0可在同一交互 Wiki 中阅读,不建立相互竞争的索引页。

Method and limits

研究方法、证据边界与引用说明

本页由当前结算的证据层自动派生,不改写已结算 run。论文只有在完整非 OCR 全文、结构化论文笔记和 claim-support audit 通过后,才进入接受证据。

页面结论是对所列论文与研究窗口的综合,不等同于无限外推。引用本专题时,请同时保留本页 canonical 地址、访问日期,并在需要时回溯原论文。