PhysTrace: Interpretable Temporal Evaluation of Physical Consistency in Video Generative Models
PhysTrace: Interpretable Temporal Evaluation of Physical Consistency in Video Generative Models
This paper introduces PhysTrace, a training-free framework that uses Multimodal Large Language Models (MLLMs) to evaluate physical consistency in video generative models by tracking agreement between history-conditioned physical expectations and observed video windows. The method achieves 67.2% accuracy on the IntPhys2 benchmark for discriminating possible versus impossible events, outperforming direct MLLM baselines. Additionally, the authors introduce IntPhys2-Timing to evaluate temporal localization capabilities, with PhysTrace reaching a Hit Rate of up to 77.4%. The study also presents reference-guided masked denoising for controlled reconstruction from mixed video latents and includes human evaluations confirming model-dependent differences in physical consistency.
PhysTrace is a training-free framework utilizing MLLMs to evaluate history-conditioned physical consistency in video generative models, achieving 67.2% discrimination accuracy on the IntPhys2 benchmark and up to 77.4% Hit Rate for temporal violation localization.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ally 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 ev
Open source S3
aluated 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 evalua
Open source S4
PhysTrace: Interpretable Temporal Evaluation of Physical Consistency in Video Generative Models
Open source S1
arXiv:2610.31429v2 Announce Type: replace-cross 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 physic
Open source S2
tions reveal model-dependent differences in physical consistency and fewer physically impossible outcomes at higher initial noise levels, trends broadly captured by PhysTrace.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org