Multi-agent system autonomously discovers new math theorems and constructions

Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment

The paper places AI agents from different model families into an open-world environment called the Station, with no central coordinator. Agents choose their own directions, run experiments, and build a shared literature. Across 12 construction problems from the AlphaEvolve catalog, the agents produced results novel to prior literature on five problems: a new infinite family of finite-field Kakeya sets, exact 604-point kissing configurations in dimension 11, new records for the discretized Kakeya needle and sign uncertainty problems, and a substantially improved lower bound for Erdős's minimum-overlap problem. They also found novel infinite families for Book Ramsey numbers. The agents output not just numerical constructions but also theorems and analyses explaining how they work. All raw dialogues, proofs, and verification code are released.

Why it matters: Multi-agent system autonomously produces 5 novel math results in an open-world setting, hitting all three HKR axes. Score capped below 85 because it's a fresh arXiv preprint with no peer review yet, and pure math discovery has an unclear path to product impact.