PhysLoop harness uses LLM physics knowledge to self-improve video generators

Researchers have published PhysLoop, a framework that draws on large language models' physical knowledge to push video generation models toward more physically consistent output. The harness guides and constrains LLMs during generation, selection and correction, converting their judgments into training data and into repairs of flawed clips at inference time. Physical behavior was also assessed on the external Physics-IQ Verified benchmark. The authors frame the work as a step toward recursive self-improvement in video generation with limited human intervention.