Organizations: Chongqing University · Everwise-Tech Co., Ltd. · Research Institute of Tsinghua University in Shenzhen · Fudan University · Tsinghua Shenzhen International Graduate School
Predictive video models have emerged as promising world models by learning latent visual dynamics from large-scale video. Yet these models remain challenged by physical events under occlusion, where later predictions may depend on object evidence that is no longer available in the current view. Addressing this challenge requires historical evidence not only to be preserved but also to remain accessible when it becomes relevant to a subsequent prediction. Existing approaches mainly enlarge the temporal context, cache generic video features, or impose explicit object-centric states, thereby improving the capacity or structure of retained history. However, they do not directly address how relevant historical evidence can be selectively retrieved and integrated into a pretrained predictor without interfering with its native latent workspace. Accordingly, we introduce HERA (Historical Evidence Routing Adapter), a framework for routing retained historical evidence into a frozen latent predictor, and instantiate it with Register-Routed Patch Memory (RRPM), a lightweight adapter comprising a Structured Memory Bank, Memory Registers, and Workspace Registers. On the IntPhys2 Main split, HERA with RRPM improves the pairwise AvgSurprise accuracy of V-JEPA 2-G from 52.57% to 54.35%. Subgroup analysis shows particularly strong improvements on fixed-camera continuity, from 46.15% to 57.69%, and fixed-camera immutability, from 46.15% to 63.46%. These results support historical evidence routing as a practical adaptation strategy for physical prediction in latent world models.
Self-supervised video models are increasingly framed as world models, yet their evaluation remains largely confined to a single top-1 accuracy score on clean benchmarks. This leaves a major gap in comprehending their potential as world models. We present the first systematic study addressing this gap, analyzing four matched-capacity frontier video foundation models, V-JEPA 2.1, V-JEPA 2, VideoPrism, and VideoMAEv2, across five robustness axes relevant to their deployment as video world models: feature discriminability, corruption robustness, fine-grained discrimination, occlusion robustness, and sensitivity to temporal direction. Our evaluations establish that across all five axes, latent-prediction models form a distinct and consistent profile. They degrade more gracefully under pixel corruption, preserve usable class structure rather than mere geometric stability under occlusion, capture fine-grained physical contact cues without reconstructing pixels, and uniquely encode the arrow of time. These advantages can even survive task adaptation: a frozen V-JEPA 2 backbone with a lightweight attentive probe outperforms a fully fine-tuned VideoMAE and a supervised TimeSformer on corruption and occlusion robustness. Our extensive results offer concrete new evidence in favor of latent prediction for robust world modeling.
Ali J Alrasheed, Aryan Yazdan Parast, Basim Azam +2
Learning predictive world models from visual observations is a core problem in embodied AI, with applications to model-based reinforcement learning and robotic planning. Existing latent world models typically generate future states with unconstrained neural transition functions, while modern video generation systems often prioritize perceptual plausibility or introduce physical structure through auxiliary losses, external guidance, or separate dynamics modules. As a result, long-horizon rollouts can remain weakly grounded in the physical principles that govern real dynamics, leading to compounding error, energy drift, and physically inconsistent futures. We propose Least Action World Models (LaWM), a latent world-modeling framework that operationalizes the Principle of Least Action in learned visual latent space: future rollouts are governed by a learned Lagrangian action functional rather than produced only by an unconstrained transition predictor. Our main technical realization is a latent variational integrator: LaWM encodes observations into learned generalized coordinates, learns a latent discrete Lagrangian over consecutive latent states, constructs a discrete action functional, and advances prediction by solving the corresponding discrete integration condition. Thus, physical structure is not merely used to score, regularize, or constrain a completed trajectory; it defines the latent transition rule itself. Because the transition is induced by a discrete variational principle, LaWM provides a structure-preserving bias for long-horizon visual prediction. Across physics-clean synthetic dynamics and embodied robot interaction benchmarks, LaWM improves physical invariance, background consistency, motion smoothness, and appearance and geometric prediction metrics over video-generation and world-model baselines.
Video world models are increasingly used to provide predictive visual representations, yet it remains unclear which pretraining signals induce action-relevant structure in their latent spaces. We study this question through a unified probe-based evaluation across diverse encoder families, including image-only self-supervision, video pretraining with and without latent prediction, reconstruction-based autoencoders, diffusion models, and shortcut-forcing dynamics models. Using a common inverse-dynamics probing objective, we find that action-relevant structure is driven primarily by temporal video pretraining rather than pixel reconstruction fidelity: models with strong pixel decoding quality can exhibit near-zero action recoverability, while video-pretrained self-supervised encoders consistently achieve the best Pareto trade-off between visual fidelity and action prediction. Comparing V-JEPA and VideoMAE further shows that most gains arise from natural-video temporal context, with feature-level latent prediction providing a smaller additional benefit. These trends transfer across robotic benchmarks, though CALVIN reveals that static-environment tasks can partially mask the importance of temporal structure by allowing strong image priors to suffice. Finally, inverse-dynamics supervision substantially improves robustness to visual corruption, suggesting that action-aware objectives regularize latent geometry beyond clean-setting performance. Our results identify temporal predictive structure -- not reconstruction fidelity -- as the primary ingredient underlying action-relevant video representations.