Video generation models create realistic moving images but often break the laws of physics. Objects vanish, multiply, or fail to react on impact. Fixing this requires physically accurate training data, which is very hard to produce at scale under controlled conditions.
PhysLoop connects a Large Language Model (LLM) to an internal simulator. The LLM knows physics from text, but the simulator enforces spatial and causal constraints. The system guides the LLM to generate, test, and correct video scenes automatically.
The main finding is that this loop produces physically consistent videos that improve the video generation model's own output quality. It provides a practical path toward , where models learn to fix their own limitations with limited human intervention.