Nvidia and partners launch Physis-Lang: physics-enhanced language model surpasses veo 3.1 in benchmarks

NVIDIA, in collaboration with MIT and the University of Oxford, has introduced Physis-Lang, a self-evolving framework that integrates physics reasoning into video captions. This system enhances video world models by adding explanations of physical phenomena, improving accuracy on physics benchmarks. On September 29, 2026, Physis-Lang achieved a score of 48.2 ± 1.4 on the Physics-IQ Verified leaderboard, outperforming Google's Veo 3.1. The Cosmos3-Nano version ranked second with 43.3 ± 1.5. Physis-Lang uses a pipeline that generates captions with physics reasoning and negative prompts to avoid implausible outcomes. The system evolves its instruction prompts through feedback from a physics-aware critic, improving caption quality over iterations. On PhyGenBench, the model lifted a frozen Cosmos3-Nano from 61.67 to 67.29. The framework also leverages a large video gallery to refine training data, achieving gains across multiple benchmarks. Fine-tuning with LoRA on attention projections improved performance without altering the model architecture.