NVIDIA team presents Physis-Lang for more physically accurate video generation
Researchers from NVIDIA, MIT and the University of Oxford have introduced Physis-Lang, a framework that uses physics-focused language to improve video world models. It adds explanations of entities, causes, interactions, governing principles, changes over time and effects to ordinary video captions. The system also creates scene-specific negative prompts describing physically implausible results, such as a stone floating on water.
The framework keeps its captioner fixed while an evolution agent revises the captioning instruction using feedback from Gemini-3.1-Pro. On PhysCapBench, which contains 246 videos and 3,794 human-verified assertions, caption F1 increased from 78.64 at iteration 1 to 87.82 at iteration 9. Language-guided retrieval produced a 183K-video training set and added an average 3.01 points across three benchmarks.
On the September 29, 2026 Physics-IQ Verified snapshot, Cosmos3-Super with Physis-Lang ranked first at 48.2 ± 1.4, while Cosmos3-Nano ranked second at 43.3 ± 1.5. In comparisons using Cosmos3-Nano, the system exceeded Veo 3.1 on three of four benchmarks, while general VBench-I2V quality remained nearly unchanged. The paper is released, but code and model weights were not public as of September 30, 2026.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.