Multi-agent AI 'Station' publishes code and Lean proofs after claiming novel mathematical discoveries

A revised arXiv preprint says an open-world, multi-agent AI system called the Station produced several mathematical results that its authors assess as new to the literature, and the team has published the code, full agent dialogues, agent-written papers and Lean proof artifacts behind those claims. The paper, “Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment,” is posted as arXiv:2608.23691 by Stephen Chung, Wenyu Du and William J. Wesley. Version 1 appeared Aug. 24, 2026, and version 2 was uploaded Sept. 14, 2026.

The Station is described as a research environment where multiple AI agents from different model families work toward a shared goal without a central coordinator or scripted pipeline. Instead of following a fixed sequence, the agents can split up tasks, test ideas, write up findings and build on each other’s work over time. As the paper’s abstract puts it: “Agents choose their own research directions, conduct experiments, collaborate and publish papers.” It adds, “These papers accumulate into a shared body of knowledge that later agents can read, cite and extend.”

According to the paper, the system was evaluated on 12 mathematical construction problems drawn from AlphaEvolve, earlier AI-for-mathematics work focused on automated search, plus two additional case studies. The authors say five of those 12 benchmark problems produced results novel relative to prior literature.

The examples highlighted in the paper include new exact 604-point kissing configurations in 11 dimensions — a geometry problem about how many nonoverlapping spheres can touch another sphere — along with a new infinite family of finite-field Kakeya sets. The authors also report improved bounds for the discretized Kakeya needle problem, improved bounds for the sign uncertainty problem, and a substantially improved lower bound for Erdős’s minimum-overlap problem. Separately, the paper says agents found novel infinite families for Book Ramsey numbers.

The authors argue that the system did more than search for high-scoring constructions. They say the agents also generated explanations and theorems, preserved in agent-authored papers that later agents could read and extend. That intermediate record is part of what makes this release notable: the team says it published the Station source code, the full raw dialogues among agents, the papers the agents wrote during the runs, and the verification code and related artifacts used to support the mathematical claims.

The paper’s abstract says: “All presented discoveries are supported by exact constructions or proofs formally verified in Lean.” Lean is a proof assistant, meaning software used to check formal mathematical statements step by step. That matters because the release combines autonomous multi-agent research with public intermediate artifacts and machine-checkable proof files, rather than only presenting final headline results.

There is also an important caveat. A Lean proof can confirm the formal statement written in Lean, but that does not by itself settle whether the formalized statement matches perfectly with an informal English description of a result. And the novelty claims remain the authors’ assessment in a preprint, not an independent confirmation by outside reviewers.

The paper says the human authors’ role after a run was mainly to check the validity of proofs and the novelty of the results. It reports a setup using six research agents: two each running GPT-5.5, Claude Opus 4.8 and Gemini 3.1 Pro. Most runs lasted about 1,000 to 2,000 ticks, which the authors say corresponded to roughly one to two weeks of wall-clock time. For now, the main takeaway is the evidence package the team has put into public view — code, dialogues, papers and formal proof artifacts — alongside a revised arXiv preprint posted Sept. 14.