PhysTrace: Interpretable Temporal Evaluation of Physical Consistency in Video Generative Models

Taekoan Yoo1 Wonkyung Jung1 Kyunghun Kim1 Kyeongbo Kong2,†
1Independent Researcher, Korea 2Pusan National University, Korea

Abstract

World action models increasingly rely on pretrained diffusion/flow-based video generators assumed to capture physical dynamics for prediction, planning, and action generation. Existing evaluations assess plausibility, probe representations, or measure predictive surprise, yet physical consistency is history-dependent: preceding events constrain what can physically occur later. We introduce PhysTrace, a training-free framework that uses MLLMs to formulate history-conditioned physical expectations and tracks their agreement with observed video windows through temporal video–language alignment, rather than relying on direct plausibility judgments or task-specific probes. On IntPhys2, PhysTrace achieves 67.2% possible–impossible discrimination accuracy, outperforming the evaluated direct MLLM baselines. We further introduce reference-guided masked denoising for controlled reconstruction from mixed possible–impossible video latents. To support these reconstructions and evaluate temporal localization as an additional capability, we introduce IntPhys2-Timing, adding violation intervals to IntPhys2, with PhysTrace achieving a Hit Rate of up to 77.4%. On these reconstructions, human evaluations reveal model-dependent differences in physical consistency and fewer physically impossible outcomes at higher initial noise levels, trends broadly captured by PhysTrace.

Representation probing and MLLM judgment compared with PhysTrace, which aligns history-conditioned expected states with each video window to localize physical violations.

1. Temporal Localization of Physical Violations on IntPhys2-Timing

Method Overview

Alignment score of each window

Qualitative Results

PhysTrace on the original IntPhys2 videos. Each scene comes as a Possible video and its Impossible twin, which differs only in the violation.

Possible
Impossible

2. Physical Tracing in Reference-Guided Reconstructions

Keeping the shared pre-violation context, each model denoises only the violation interval from the noised midpoint of the Possible / Impossible latents, guided toward the closer reference. PhysTrace then evaluates the reconstruction.