ProofCouncil LLM Agent Excels at Open Math Problems.

Large language models (LLMs) are showing increasing capability in tackling complex mathematical problems, but their performance can be further enhanced through specialized agentic workflows. Researchers have introduced ProofCouncil, an LLM agent specifically engineered to solve open mathematical problems using an author-critic architecture, mimicking human problem-solving processes. ProofCouncil participated in the FirstProof challenge, which involved 10 real-world mathematical problems requiring autonomous agent solutions. The agent's submissions for six of these problems were judged as correct, requiring at most minor revisions, positioning it as the top performer among participating teams. Beyond the challenge, ProofCouncil was evaluated on 30 additional open problems sourced from mathematical researchers. Out of 21 solutions that received human feedback, five were deemed completely correct, two were promising pending final verification, and eight provided useful partial progress. The underlying agent-building library used to develop ProofCouncil has been released as open source, enabling broader community use and development.