NVIDIA researchers have introduced Physis-Lang, a self-evolving framework designed to integrate physics reasoning into video generation. The system adds a physics_reasoning field to captions, detailing causal interactions and governing principles, alongside a scene-specific physics_negative_prompt to prevent implausible outcomes during inference.
The development process utilized a GPT-5.5 captioner and Gemini-3.1-Pro as a critic to refine instructions over nine iterations. The final training set comprises 183 000 videos, including 71 000 from WISA-80K and 112 000 retrieved clips. Performance validation on PhysCapBench showed an increase in caption F1 scores from 78.64 to 87.82.
On the Physics-IQ Verified leaderboard as of September 29, 2026, the Cosmos3-Super model with Physis-Lang ranked first with a score of 48.2, followed by Cosmos3-Nano at 43.3. The framework improved performance across various backbones, including a 7.05 point gain on Wan2.1-14B and 5.02 on Cosmos3-Super-64B.
The team distilled the pipeline into two Qwen3-VL-4B-Instruct models: PhysThinker-C and PhysThinker-U. Using these local models reduced API costs from $24.12K to $0.12K while maintaining significant performance gains. The research was officially released by September 30, 2026.
