Preprint reports mathematical results from open-world AI agents

A preprint reports that an open-world, multi-agent environment called the Station produced results the authors describe as novel relative to prior literature on five mathematical problems. Across 12 AlphaEvolve construction problems and two case studies, the authors report new finite-field Kakeya sets, 604-point kissing configurations in dimension 11, records in two problems, and an improved Erdős minimum-overlap lower bound. The project released raw agent dialogues, proofs and verification code.