LLM agents obey the physics law of detailed balance

LLM agents obey the law of detailed balance, learning global potential functions that guide states toward goals. This marks a shift from black-box models to predictable science.