
PhysTrace: Interpretable Temporal Evaluation of Physical Consistency in Video Generative Models
PhysTrace is a training-free framework for evaluating physical consistency in videos, relevant to video-based world models used for robot prediction and planning. It aligns multimodal-LLM-generated, history-conditioned expectations with observed video windows to localize physical violations over time and describe them. On the new IntPhys2-Timing benchmark, it achieves up to 70.7% hit rate across physical principles; on IntPhys2, it outperforms compared methods in overall possible–impossible discrimination accuracy.

