Physis-Lang: Self-Evolving Language as a Physical Representation for Video World Model
Physis-Lang improves physical video generation, and Cosmos3-Nano variants surpass proprietary Veo 3.1.
Physis-Lang represents physical processes in language covering entities, causes, interactions, principles, evolution, and effects, then optimizes that language across curation, training, and generation. PhysCapBench decomposes processes into atomic assertions and scores caption recall and precision, while an agentic loop fixes assertion errors and retrieves videos for missing processes. On four physical-video benchmarks, Wan and Cosmos backbones improve, and Cosmos3-Nano variants enhanced by Physis-Lang surpass proprietary Veo 3.1.
56