AI 中文总结
本研究提出 HERA 框架，通过 RRPM 适配器为冻结潜在预测器路由历史证据，在 IntPhys2 数据集上显著提升 V-JEPA 2-G 的物理预测准确率，验证了该策略的有效性。
AI 中文摘要
预测视频模型通过从大规模视频中学习潜在视觉动态，已成为颇具前景的世界模型。然而，这些模型在处理遮挡情况下的物理事件时仍面临挑战，后续预测可能依赖当前视图中不再存在的对象证据。要解决这一挑战，不仅需要保留历史证据，还需在其与后续预测相关时保持可访问性。现有方法主要通过扩大时间上下文、缓存通用视频特征或施加显式以对象为中心的状态，来提升保留历史的容量或结构，但未直接解决如何在不干扰预训练预测器固有潜在工作空间的情况下，选择性检索相关历史证据并将其整合。因此，我们提出 HERA（Historical Evidence Routing Adapter，历史证据路由适配器），这一用于将保留的历史证据路由至冻结潜在预测器的框架，并通过 Register-Routed Patch Memory（RRPM，寄存器路由补丁记忆）实现，该轻量适配器包含结构化记忆库、记忆寄存器和工作空间寄存器。在 IntPhys2 主数据集上，搭载 RRPM 的 HERA 将 V-JEPA 2-G 的成对平均惊喜度准确率从 52.57% 提升至 54.35%；子组分析显示，其在固定相机连续性任务上的提升尤为显著，从 46.15% 升至 57.69%，在固定相机不变性任务上则从 46.15% 升至 63.46%。这些结果表明，历史证据路由是潜在世界模型中物理预测的实用适配策略。
英文摘要
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.
