Does LeWorldModel Preserve Local Action Geometry?

Computer Science > Machine Learning arXiv:2610.31716 (cs) [Submitted on 22 Oct 2026] Title: Does LeWorldModel Preserve Local Action Geometry? Authors: Vasil Vasilev, Siddharth Nayak.

计算机科学 > 机器学习 arXiv:2610.31716 (cs) [提交于 2026 年 10 月 22 日] 标题：LeWorldModel 是否保留了局部动作几何结构？作者：Vasil Vasilev 及另一位作者。

Abstract: We examine how closely pretrained LeWorldModel (LeWM) reproduces the effects of small action changes. Using matched simulator and model responses, we measure action sensitivity with Gram matrices at three perturbation scales. Across nearly a thousand states from Reacher and OGBench Cube, the most sensitive action directions differ by median angles of about $36^\circ$ and $42^\circ$, respectively, at the smallest scale.

摘要：我们研究了预训练的 LeWorldModel (LeWM) 在多大程度上重现了微小动作变化的影响。通过匹配模拟器和模型的响应，我们在三个扰动尺度上使用 Gram 矩阵测量了动作敏感度。在来自 Reacher 和 OGBench Cube 的近一千个状态中，在最小尺度下，最敏感的动作方向的中位数角度差异分别约为 $36^\circ$ 和 $42^\circ$。

Increasing the perturbation tenfold reduces median normalized Gram discrepancy by roughly 70% on Reacher and 47% on Cube. Geometry measured in pixels and embeddings varies more across scales than the chosen physical-state geometry. On held-out directions, environment Grams achieve median embedding response errors of 17% and 28%, compared with 26% and 38% for an isotropic reference with the same trace.

将扰动增加十倍后，Reacher 和 Cube 上的中位数归一化 Gram 差异分别减少了约 70% 和 47%。在不同尺度下，以像素和嵌入（embeddings）测量的几何结构比所选的物理状态几何结构变化更大。在留出的方向上，环境 Gram 矩阵实现了 17% 和 28% 的中位数嵌入响应误差，而具有相同迹（trace）的各向同性参考模型的误差分别为 26% 和 38%。

LeWM’s errors are higher at this scale. These findings support scale-specific evaluation of action sensitivity; its effect on control remains an open question.

LeWM 在此尺度下的误差更高。这些发现支持对动作敏感度进行特定尺度的评估；其对控制的影响仍是一个悬而未决的问题。