MAC: Multi-Agent Causal Discovery Framework Using LLMs

A new framework called MAC (Multi-Agent Causal Discovery) combines large language models with traditional statistical causal discovery methods. It uses a multi-agent debate system to improve causal graph accuracy. The framework includes a Debate-Coding Module that selects and executes the best SCD algorithm, and a Meta-Debate Module that refines the graph through adversarial debate. A Meta Fusion mechanism bridges these modules. The approach addresses limitations of both purely statistical methods and single-agent LLM approaches.