NVIDIA Researchers Introduce Physis-Lang: Self-Evolving Physical Language That Lifts Cosmos 3 Past Veo 3.1 on Physics Benchmarks

NVIDIA researchers have introduced Physis-Lang, a framework designed to improve the physical accuracy of video world models by using language-based guidance rather than additional visual or numerical data. By integrating this self-evolving physical language, the team’s Cosmos 3 model outperformed Veo 3.1 on established physics benchmarks. This approach aims to address the common issue where generative models produce realistic clips that fail to adhere to basic physical principles.