跳转至

被动物体状态世界模型中运动学、接触及物体恒存场的事件条件诊断

文章背景与核心概要

现代世界模型虽然能够成功预测未来的物理状态，但高预测准确率本身并不能完全解释物理信息是如何在其内部动力学中被组织和利用的。为了解决这一问题，作者引入了一种受控诊断协议，用于研究被动物体状态世界模型中由事件条件的隐式物理结构。

该研究使用了一个包含自由运动、碰撞和遮挡事件的平衡、受控生成器数据集，并在固定视界预测设置下评估了循环模型、基于注意力的模型以及隐状态空间转移模型。核心发现表明：隐藏表征能够可靠地编码事件状态信息；事件上下文能够系统性地重新加权物理场读数（如自由运动侧重运动学，碰撞结合运动学与接触结构，遮挡融合运动与物体恒存结构）；时间对齐与方向一致性分析揭示了物理场侧重点随阶段的变化；而因果场效应分析则表明抑制特定对齐方向会损害事件相关预测。需要注意的是，这些结果支持隐式物理场的事件条件组织和固定视界功能敏感性，但并不意味着存在显式的物理模块、隔离的因果电路或上下文不变的滑动窗口泛化能力。

Abstract Summary / 摘要概要

While modern world models can successfully predict future physical states, high prediction accuracy alone does not explain how physical information is internally organized and utilized within their latent dynamics.

While modern world models can successfully predict future physical states, high prediction accuracy alone does not explain how physical information is internally organized and utilized within their latent dynamics.

To address this, the authors introduce a controlled diagnostic protocol to study event-conditioned latent physical structures in passive object-state world models. Using a balanced, controlled-generator dataset featuring free-motion, collision, and occlusion events, the study evaluates recurrent, attention-based, and latent state-space transition models under a fixed-horizon forecasting setup.

To address this, the authors introduce a controlled diagnostic protocol to study event-conditioned latent physical structures in passive object-state world models. Using a balanced, controlled-generator dataset featuring free-motion, collision, and occlusion events, the study evaluates recurrent, attention-based, and latent state-space transition models under a fixed-horizon forecasting setup.

Key Findings / 核心发现:

  1. Event-Regime Readout: Hidden representations reliably encode event-regime information.
  2. Contextual Reweighting: Event contexts systematically reweight physical field readouts:
  3. Free motion is predominantly kinematic.
  4. Collision combines kinematic and contact structures.
  5. Occlusion merges motion-related and object-permanence structures.
  6. Phase-Related Shifts: Time-aligned and directional-consistency analyses demonstrate phase-dependent shifts in field emphasis.
  7. Causal Field Effects (CFE): Fixed-horizon projection analyses indicate that suppressing field-aligned directions can degrade event-relevant predictions. The strongest evidence surfaces for contact-aligned structure during collision-contact windows, and more qualified evidence for object-permanence-aligned structure in hard-occlusion hidden windows.

Key Findings:

  1. Event-Regime Readout: Hidden representations reliably encode event-regime information.
  2. Contextual Reweighting: Event contexts systematically reweight physical field readouts:
  3. Free motion is predominantly kinematic.
  4. Collision combines kinematic and contact structures.
  5. Occlusion merges motion-related and object-permanence structures.
  6. Phase-Related Shifts: Time-aligned and directional-consistency analyses demonstrate phase-dependent shifts in field emphasis.
  7. Causal Field Effects (CFE): Fixed-horizon projection analyses indicate that suppressing field-aligned directions can degrade event-relevant predictions. The strongest evidence surfaces for contact-aligned structure during collision-contact windows, and more qualified evidence for object-permanence-aligned structure in hard-occlusion hidden windows.

Note: These results support event-conditioned organization and fixed-horizon functional sensitivity of latent physical fields, but do not imply the existence of explicit physical modules, isolated causal circuits, or context-invariant sliding-window generalization.

Note: These results support event-conditioned organization and fixed-horizon functional sensitivity of latent physical fields, but do not imply the existence of explicit physical modules, isolated causal circuits, or context-invariant sliding-window generalization.

Access & Resources / 访问与资源