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.
Figures & tables
Figure 2: Overview of PhysTrace. History-conditioned predictions are aligned with the current video window, while salient observations are selectively retained for subsequent reasoning. The alignment scorer can employ either a unified multimodal embedding model or separate contrastively aligned video and text encoders.
Model
Prediction
State
All Trusted
Ours
Guided
Guided
Qwen-VL (8B)
0.270
0.263
0.201
0.195
Qwen-VL (32B)
0.229
0.223
0.145
0.134
Cosmos-Reason2 (8B)
0.350
0.343
0.270
0.242
GLM-4.1V
0.268
0.263
0.222
0.220
Swapped Prompts
0.500
0.501
Table 1: Cosine distance( ↓ ) between MLLM-generated prompts and ground-truth captions using Qwen3-Embedding [ 47 ] . All Trusted: All true Trusted ; Guided: Guided History and Trusted.
Figure 3: Example of physical violation localization for permanence. State is observed from the current window, while Prediction describes how the window should appear given the trusted history.
Figure 4: Temporal alignment across diffusion timesteps. Alignment traces for a spatio-temporal continuity example at different noise levels σ , computed with Wan2.1-T2V-1.3B. Red dot : MAS within the violation interval. Orange dots : MAS outside the violation interval.
Figure 5: Effects of window and hop sizes. Broader temporal coverage makes the violation more salient, yielding a sharper drop within the ground truth (GT) interval for the solidity attribute.
Method
Attribute
Easy
Med.
Hard
Unk.
All
Ours
Continuity
62.5
60.0
69.2
58.3
63.2
Immutability
100.0
73.3
64.0
58.3
70.7
Permanence
50.0
53.3
61.5
58.3
56.6
Solidity
83.3
58.3
70.0
100.0
69.6
VoE
Continuity
50.0
53.3
73.1
66.7
61.8
Immutability
50.0
56.7
52.0
58.3
54.7
Table 2: Temporal localization hit rates on IntPhys2-Timing. Hit Rates (%) across physical attributes and difficulty categories. VoE: a V-JEPA2-based surprise baseline. Med. : Medium; Unk. : Unknown
Model
Easy
Med.
Hard
Unk.
Overall
Ours
59.6
63.0
56.5
59.3
59.9
GPT-4o
51.0
54.0
51.2
51.2
52.3
Qwen-VL 2.5 (72B)
51.9
49.3
50.0
48.3
49.6
Gemini-2.5 (Pro)
62.5
54.8
53.6
56.4
55.4
Gemini-2.5(Flash)
52.9
57.3
55.1
54.7
55.6
VideoMAEv2
42.3
52.0
48.2
51.2
49.6
Table 3: Possible-vs-Impossible accuracy (%) on the IntPhys2 benchmark across difficulty levels, with MAS-based predictions for our method; Med. : Medium; Unk. : Unknown.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Alignment Scoring Model
Rate (%)
Δ (%)
PE-AV-Large [ 37 ]
84.9
-18.9
PE-Core-G14-448 [ 10 ]
88.4
-15.3
PE-Core-L14-336 [ 10 ]
86.0
-15.2
InternVideo2-CLIP-6B [ 41 ]
88.4
-17.5
InternVideo2-Stage2-6B [ 41 ]
79.1
-12.1
VidVec embedding (Qwen3-VL-8B) [ 36 ]
79.1
-13.5
Appendix
Table 4: Caption substitution sensitivity on WebVid. Det.Rate : fraction of caption-element substitutions that lower the original alignment score. ΔDet. : mean signed score change normalized by the mean drop from unrelated-sentence replacements ( −100% = as large as the unrelated-sentence drop).
Figure 6: Per-clip examples of caption substitution sensitivity on WebVid. Each row replaces one caption element (subject, place, action, or attribute); the last row replaces the entire caption with an unrelated sentence. Columns denote alignment scoring models. Cells report ΔDet. , the signed score change relative to the original caption, normalized by the mean drop from unrelated-sentence replacements. Green : ΔDet.≤−25% ; Yellow : −25%<ΔDet.<0 ; Red : ΔDet.≥0 .
Figure 7: Qualitative examples of physical violation localization. Results are shown across different physical attributes, with time in seconds on the x-axis.
Figure 8: Qualitative examples of physical violation localization. Results are shown across different physical attributes, with time in seconds on the x-axis (continued).
Physical consistency in video generation studies how to make and evaluate videos whose geometry, motion, interactions, and state changes obey real-world laws.