NVIDIA 等提出 Physis-Lang，让 Cosmos 3 在物理基准上超过 Veo 3.1

Video world models can render convincing clips that still break physics. Butter spreads like paint. Balls pass through walls. A team from NVIDIA, MIT and the University of Oxford argues the fix can come from language itself, not from extra visual, latent or numerical signals.

Their framework, Physis-Lang , treats physical language as a shared, optimizable representation. The same text drives data curation, model training and inference. On the public Physics-IQ Verified leaderboard snapshot dated September 29, 2026, Physis-Lang on Cosmos3-Super ranks first at 48.2 ± 1.4. The Cosmos3-Nano version ranks second at 43.3 ± 1.5.

A video world model can make a convincing clip and still get the physics wrong. Our researchers just released Physis-Lang, an open self-evolving framework that adds physics reasoning to video captions. The captions explain why and how a scene unfolds. We use them to fine-tune… pic.twitter.com/agIHuIB6N2

Conventional captions describe what happens, not why. ‘Butter melts as the temperature rises’ says nothing about heat transfer or gravity. Physis-Lang adds a physics_reasoning field to each base caption. It spells out entities, causes, interactions, governing principles, temporal evolution and effects.

The pipeline also writes a scene-specific physics_negative_prompt . This text describes likely implausible outcomes, such as a stone floating on water. It acts as negative conditioning at inference time.

The loop keeps the captioner frozen and evolves only its instruction. A GPT-5.5 captioner writes captions for a fixed 20-video development set with 273 human-verified assertions. Gemini-3.1-Pro acts as a physics-aware critic. An evolution agent reads the scores and claim-level failures, then rewrites the prompt.

The critic scores 2 dimensions:

Every revised prompt is validated on PhysCapBench , a new benchmark of 246 videos and 3,794 human-verified assertions. Caption F1 rose from 78.64 at iteration 1 to 87.82 at iteration 9. The path was not smooth. Iteration 2 made captions overly cautious and dropped F1 to 76.28. Iteration 9 required every visible causal step and raised frame sampling from 2 fps to 4 fps.

A GPT-5.5 diagnosis agent maps generated-video failures to physics categories like rigid-body motion, collision and fluid dynamics. That deficiency profile is matched against physics tags on a large video gallery. Retrieval targets physical content, not visual appearance.

The final training set holds 183K videos: 71K filtered from WISA-80K plus 112K retrieved clips. Retrieval alone added 3.01 points on average across 3 benchmarks. On VideoPhy-2, chemical and thermal processes each gained 8.00 points.

Fine-tuning uses LoRA on attention projections, with no architecture or objective change. Physis-Lang on Cosmos3-Nano versus Google's Veo 3.1:

Gains hold across backbones: +7.05 on Wan2.1-14B, +3.24 on Cosmos3-Edge-4B, +6.22 on Cosmos3-Nano-16B and +5.02 on Cosmos3-Super-64B. General quality held steady on VBench-I2V, where Cosmos3-Nano moved from 88.32 to 88.69.

Prompting alone also helps. Physics reasoning plus negative prompts lifted a frozen Cosmos3-Nano on PhyGenBench from 61.67 to 67.29.

The research team distilled the GPT pipeline into 2 Qwen3-VL-4B-Instruct models: PhysThinker-C for captioning and PhysThinker-U for prompt upsampling. On Wan2.1-14B, the commercial pipeline gave +7.05 at about $24.12K in API cost. Swapping in PhysThinker-C kept +6.76 at about $0.12K. A fully local setup cost $0 and still added +4.76.

Scores are from the Physis-Lang paper , Tables 1 to 4, run under one protocol per benchmark (PhyGenBench and VideoPhy-2 use a GPT-5.5 evaluator). Physis-Lang numbers use the Cosmos3-Nano backbone. n/r = not reported in that comparison. Release status checked September 30, 2026.