Mean-Field Dynamics Framework for Chain-of-Thought Reasoning

This framework models LLM reasoning as a guided discovery process on a clue graph using mean-field approximation. By deriving a one-dimensional ordinary differential equation for clue discovery, the method identifies reproducible statistical regularities in reasoning chains without altering model architecture.