Physis-Lang · Cosmos3-Super
48.2 ± 1.4
Physis-LangSelf-Evolving Language as a Physical
Representation for Video World Model
1 NVIDIA 2 MIT 3 University of Oxford † Equal contribution * Corresponding authors
Language can do more than describe a scene.
It can represent how the physical world evolves.
Physics-IQ Verified
First and second on the public leaderboard.
Physis-Lang · Cosmos3-Nano
43.3 ± 1.5
Super improves on the previous leading score of 42.7 by 5.5 percentage points. Nano also surpasses the original Cosmos3-Super by mean score.
View the live leaderboard
I2V Verified · Ranked by mean score · September 29, 2026
The Idea
A shared language for physical intelligence.
Visually convincing videos can still break the laws of physics. Physis-Lang makes causes, interactions, governing principles, and effects explicit in language—then improves that language through feedback. The same representation connects data curation, model training, and inference.
Representation
Beyond what happens. Explain why it happens.
“Butter melts as the temperature rises.”
Conventional captionCause
Rising temperature supplies heat to the butter.
Physical law
Heat drives melting; gravity makes the butter slump.
Effect
The solid shrinks as a shallow liquid pool expands.
Framework
One representation. Three connected stages.
Evolve the language
A physics-aware critic diagnoses missing or unsupported claims. An agent refines shared captioning guidelines while the captioner stays fixed.
Retrieve missing physics
Model failures become physics-domain queries. Language-guided retrieval finds visually diverse videos that address those deficiencies.
Train and generate
The evolved language re-captions training videos and expands inference prompts, with scene-specific negative descriptions to discourage implausible dynamics.
PhysCapBench
A benchmark for evaluating the richness of physical detail in video captions.
246Reviewed physical videos
3,794Human-verified assertions
15.4Assertions per video
Is each claim supported?
Decompose the caption into atomic claims and verify each against the video as correct, incorrect, or uncertain.
Is the physical meaning covered?
Check whether the caption explicitly states or entails each complete human-curated physical assertion.
F1 combines precision and recall. PhysCapBench also provides validation for selecting the best evolved guidelines.
Self-Evolution
Better guidelines. Better physical descriptions. Better physical generation.
PhysCapBench F1
78.64 → 87.82
PhyGenBench
64.17 → 67.29
Release paper, Figure 4. Downstream comparison keeps pretrained Cosmos3-Nano fixed and changes only the inference captions from guideline iterations 1, 4, 8, and 9. Caption F1 is non-monotonic; iteration 9 is selected.
Language-Guided Data Curation
Find the physics the model is missing.
VideoPhy-2 gains by physical category
Improvement after language-guided retrieval
Cloth deformation (167)
+7.19
Fracture mechanics (94)
+7.45
Elasticity (81)
+1.23
Chemical processes (50)
+8.00
Rigid-body motion (33)
+6.06
Contact / collision (29)
+3.45
Soft-body motion (27)
+7.41
Thermal processes (25)
+8.00
Joint-score gain (percentage points)
Results
Stronger physical generation across benchmarks and backbones.
Cosmos3-NanoVeo 3.1Physis-Lang (Cosmos3-Nano)
Release paper, Tables 1–4. Scores use a 100-point scale with the same evaluation protocol per benchmark. VideoPhy-2 above uses the Hard split; on the full set, Physis-Lang scores 68.02 vs. Veo 3.1’s 68.87. PhyGenBench and VideoPhy-2 use GPT-5.5 evaluation.
View benchmark scores and backbone gains
| Model | PhyGenBench | Physics-IQ Verified | VideoPhy-2 All / Hard | PhyGround |
|---|---|---|---|---|
| Cosmos3-Nano | 61.67 | 40.23 | 60.41 / 48.31 | 65.18 |
| Veo 3.1 | 65.63 | 34.99 | 68.87 / 58.43 | 69.24 |
| Physis-Lang (Cosmos3-Nano) | 71.04 | 43.41 | 68.02 / 62.36 | 69.90 |
Gains across model families and scales
Wan
Wan2.114B+7.05
Cosmos
Cosmos3-Edge4B+3.24
Cosmos3-Nano16B+6.22
Cosmos3-Super64B+5.02
Release paper, Table 5. Mean gains are positive for every backbone; individual benchmark gains can vary.
Video Gallery
Physical processes, brought to life.
Contact and collision
The blocks topple in sequence.
Wetting and deformation
The paper settles and deforms in the liquid.
Tearing under tension
The tear opens as the grips pull apart.
Motion and shadows
The shadow changes as the mug rotates.
Deformation under load
The pillow compresses beneath the weight.
Generated outputs, not retrieved training videos. Examples illustrate individual phenomena rather than aggregate benchmark performance.
Citation
Build on this work.
@misc{lu2026physislang,
title={Physis-Lang: Self-Evolving Language as a Physical
Representation for Video World Model},
author={Liming Lu and Xianzheng Ma and Wenkun He and Guanqi Zhan
and Yilin Zhao and Junyu Chen and Mengyao Xu and Jiaojiao Fan
and Wenhang Ge and Yuchao Gu and Yunze Liu and Boyi Li
and Zhen Dong and Victor Prisacariu and Ming-Yu Liu
and Song Han and Han Cai},
year={2026},
institution={NVIDIA}
}