- Published on
NVIDIA Researchers Introduce Physis-Lang: Self-Evolving Physical Language That Lifts Cosmos 3 Past Veo 3.1 on Physics...
- Authors
- Name
- DeepHorus
NVIDIA researchers, in collaboration with scientists from MIT and the University of Oxford, have introduced Physis-Lang, a self-evolving physical language framework that substantially improves the physical plausibility of generated video. The technique treats physics-grounded language as a shared, optimizable representation spanning data curation, training, and inference. On the Physics-IQ Verified leaderboard snapshot dated September 29, 2026, Physis-Lang on Cosmos3-Super ranks first with a score of 48.2 Ā± 1.4, while the Cosmos3-Nano variant places second at 43.3 Ā± 1.5. Both positions sit ahead of Google's Veo 3.1, signaling a notable shift in how physical reasoning can be injected into video world models.
The core problem Physis-Lang addresses is that conventional video captions describe events without explaining the underlying physics. A caption reading "butter melts as the temperature rises" says nothing about heat transfer or gravity, leaving models free to render butter that spreads like paint or balls that pass through walls. Physis-Lang augments each base caption with a physics_reasoning field that spells out entities, causes, interactions, governing principles, temporal evolution, and effects. It also generates a scene-specific physics_negative_prompt describing likely implausible outcomes, which serves as negative conditioning during inference. Together, these textual signals give the model explicit guidance on what should and should not happen.
The framework's self-evolving caption loop keeps the captioner model frozen while iteratively refining only its instruction prompt. A GPT-5.5 captioner produces captions for a fixed 20-video development set containing 273 human-verified assertions, and a Gemini-3.1-Pro physics-aware critic scores them on precision and recall. An evolution agent then reads the claim-level failures and rewrites the prompt. Each revised prompt is validated on PhysCapBench, a new benchmark of 246 videos and 3,794 verified assertions, where caption F1 climbed from 78.64 at iteration 1 to 87.82 at iteration 9. The path was not linear; iteration 2 made captions overly cautious and briefly dropped F1 to 76.28.
Language-guided data curation forms the second pillar of the system. A GPT-5.5 diagnosis agent maps generated-video failures to physics categories such as rigid-body motion, collision, and fluid dynamics, then matches those deficiency profiles against physics tags on a large video gallery. The final training set contains 183K videos: 71K filtered from WISA-80K plus 112K retrieved clips. Retrieval alone added 3.01 points on average across three benchmarks, with chemical and thermal processes on VideoPhy-2 each gaining 8.00 points.
Benchmarks show Physis-Lang on Cosmos3-Nano outperforming Veo 3.1 across most metrics: 71.04 vs 65.63 on PhyGenBench, 43.41 vs 34.99 on Physics-IQ Verified, and 62.36 vs 58.43 on the VideoPhy-2 Hard split. Gains hold across multiple backbones, from Wan2.1-14B to Cosmos3-Super-64B, while general video quality on VBench-I2V improved slightly from 88.32 to 88.69. Even prompting alone helps, lifting a frozen Cosmos3-Nano from 61.67 to 67.29 on PhyGenBench.
Critically, the team demonstrated that commercial APIs are not required. They distilled the GPT pipeline into two Qwen3-VL-4B-Instruct models, PhysThinker-C and PhysThinker-U, reducing per-run API costs from roughly $24,120 to about $120 while retaining +6.76 of the original +7.05 gain on Wan2.1-14B. A fully local setup still delivers +4.76 at zero cost, making the approach accessible to independent researchers and smaller labs. Looking ahead, Physis-Lang suggests that language-centric physical reasoning could become a standard component of video generation pipelines, complementing rather than replacing architectural and objective improvements.
