NVIDIA Physis-Lang: Self-Evolving Physical Language

Video world models often generate visually convincing clips while failing fundamental physical laws. To bridge this gap, NVIDIA researchers have introduced Physis-Lang, an open self-evolving framework designed to embed structured physical reasoning directly into video captions.

Conventional captions typically describe surface events without explaining underlying mechanics, such as heat transfer or gravity. Physis-Lang addresses this limitation by appending a dedicated physics_reasoning field to base captions, explicitly detailing entities, causes, interactions, governing principles, temporal evolution, and effects.

Stock photo for illustration only, not from the actual event

Additionally, the pipeline formulates a scene-specific physics_negative_prompt for each clip. This text explicitly details implausible outcomes, such as a stone floating on water, functioning as negative conditioning during the inference stage to prevent unrealistic visual artifacts.

"A video world model can make a convincing clip and still get the physics wrong. Our researchers just released Physis-Lang, an open self-evolving framework that adds physics reasoning to video captions."

NVIDIA Researchers

The system's optimization loop freezes the captioner while evolving its instructions. A GPT-5.5 captioner generates descriptions for a development set, evaluated by a physics-aware critic model powered by Gemini-3.1-Pro, while an evolution agent refines prompts based on claim-level failures tracked across PhysCapBench.