Video World Models · 2026-10-23
PhysTrace: Interpretable Temporal Evaluation of Physical Consistency in Video Generative Models
Taekoan Yoo, Wonkyung Jung, Kyunghun Kim, Kyeongbo Kong
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
PhysTrace is a training-free framework for temporally localizing physical consistency violations in video generative models. It repeatedly aligns history-conditioned, MLLM-generated expected states with observed video windows in a shared multimodal embedding space. The paper introduces IntPhys2-Timing with violation intervals, and evaluates PhysTrace on reconstructions along diffusion/flow-based generative trajectories.
Why it matters
The authors argue that physical consistency is temporal and that prior evaluations often miss where expected and observed dynamics diverge. PhysTrace aims to localize physical violations over time without task-specific probes/classifiers, and provides interpretable descriptions of which history-conditioned physical expectation conflicts with observations.
Method
- For each temporal window, an MLLM generates a current-state description (no history) and a history-conditioned expected prediction for the same window; an alignment scorer computes cosine similarity between the prediction text and the video window to produce an alignment trace over time.
- PhysTrace uses selective state retention via a MLLM-produced Trusted indicator to decide whether each newly observed state is added to history for subsequent predictions.
- The paper constructs IntPhys2-Timing by adding temporal violation annotations and evaluates temporal localization using Minimum Alignment Score (MAS) and Hit Rate based on whether the MAS-timing window overlaps the annotated violation interval.
Limitation
The paper notes that IntPhys2-Timing excludes categories with ambiguous object identities, substantial occlusion, cluttered scenes, or atypical structures that make temporal violation annotation difficult.
Abstract (from arXiv)
World action models increasingly rely on pretrained video models assumed to capture physical dynamics for prediction, planning, and action generation. Existing evaluations assess plausibility, probe representations, or measure predictive surprise, but rarely identify which history-conditioned physical expectations conflict with observations. We introduce PhysTrace, a training-free framework aligning MLLM-generated, history-conditioned expectations with observed video windows to temporally localize physical violations and provide interpretable descriptions without task-specific probes or classifiers. We also introduce IntPhys2-Timing, adding violation intervals to IntPhys2. PhysTrace achieves a Hit Rate of up to 70.7% across physical principles on IntPhys2-Timing and outperforms the compared methods in overall possible-impossible discrimination accuracy on IntPhys2. Applied to reconstructed videos along the trajectories of pretrained diffusion/flow-based video models, PhysTrace reveals attribute- and noise-dependent patterns in physical violation localization, enabling interpretable physical consistency analysis across timesteps.
Related papers
- Time-JEPA: Grounding Video World Models with Physical Time Series
- PVBench: Quantifying Physical Plausibility in Video Generation with Object-Level Physical Measurements
- Physis-Lang: Self-Evolving Language as a Physical Representation for Video World Model
- World Models Need a Physical Registry: Learning Internal Predictive States for Physically Plausible World Models
- STRIKE: Learning Visual State Transitions for Physical World Modeling
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-22 · Indexed by Pub-AI: 2026-10-23 UTC
PhysTrace: Interpretable Temporal Evaluation of Physical Consistency in Video Generative Models
Preprint posted: 2026-10-22 · Original record ↗
- arxiv
- 2610.31429
Explicit source versions (1)
- v1 · 2026-10-22
- Preprint available
Source date: 2026-10-22 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
World action models increasingly rely on pretrained video models assumed to capture physical dynamics for prediction, planning, and action generation. Existing evaluations assess plausibility, probe representations, or measure predictive surprise, but rarely identify which history-conditioned physical expectations conflict with observations. We introduce PhysTrace, a training-free framework aligning MLLM-generated, history-conditioned expectations with observed video windows to temporally localize physical violations and provide interpretable descriptions without task-specific probes or classifiers. We also introduce IntPhys2-Timing, adding violation intervals to IntPhys2. PhysTrace achieves a Hit Rate of up to 70.7% across physical principles on IntPhys2-Timing and outperforms the compared methods in overall possible-impossible discrimination accuracy on IntPhys2. Applied to reconstructed videos along the trajectories of pretrained diffusion/flow-based video models, PhysTrace reveals attribute- and noise-dependent patterns in physical violation localization, enabling interpretable physical consistency analysis across timesteps.