AI agents operating in an open‑world multi‑agent environment named the Station independently performed mathematical discovery. They generated novel constructions and accompanying theorems across multiple problem sets.

The Station was created to allow AI agents from different model families to pursue a shared research goal without a central coordinator or scripted pipeline. Agents select research directions, conduct experiments, collaborate, and contribute to a shared literature.

The agents addressed 12 construction problems from the AlphaEvolve catalogue and two additional case studies. Their results included an infinite family of finite‑field Kakeya sets, an exact configuration of 604 points on a kissing graph in dimension 11, new records for the discretized Kakeya needle problem and the sign uncertainty problem, and a lower bound for Erdős’s minimum‑overlap problem that exceeds the previous record. They also discovered infinite families for Book Ramsey numbers.

In addition to the constructions, the agents produced theorems and analyses that explain the underlying mechanisms, enhancing interpretability for mathematicians. The authors released raw agent dialogues, proofs, and verification code, providing a transparent record of the discovery process.

The study demonstrates a potential pathway for autonomous scientific inquiry, while challenges remain in scaling the approach to broader mathematical domains.