ProofCouncil: An LLM Agent for Solving Open Mathematical Problems (arxiv.org)

A groundbreaking development in AI, ProofCouncil emerges as a large language model (LLM) agent designed to tackle open mathematical problems with enhanced efficiency. Utilizing an author-critic architecture, ProofCouncil submitted solutions to the FirstProof challenge, successfully addressing 6 out of 10 complex issues with feedback indicating correctness after minor revisions. This achievement positions ProofCouncil at the forefront of LLM performance in mathematical problem-solving, showcasing its potential to blend AI with real-world mathematical practice. Significantly, ProofCouncil addresses the limitations of traditional AI approaches by employing agentic workflows tailored to the intricacies of mathematical reasoning. In evaluations beyond the challenge, it provided viable solutions for 30 additional open problems, with 5 solutions validated as completely correct and others recognized for their promising contributions. The open-source release of the agent-building library used to create ProofCouncil invites collaboration within the AI/ML community, paving the way for further innovation in using LLMs to advance mathematical inquiry and problem-solving.