The standard framing of AI-assisted mathematics is collaborative: a human knows what to prove, feeds the problem to a model, and the model helps search the proof space. That’s a real and useful mode. What Stephen Chung, Wenyu Du, and William Wesley describe in a paper submitted to arXiv last Monday is different: an environment where AI agents from different model families independently select which problems to work on, run experiments, share findings with each other, and produce results with no centralized coordination directing any of it.

They call it the Station. The key design choice is the open-world setup: agents aren’t assigned a problem and told to solve it. They operate in an environment stocked with mathematical construction problems and choose where to allocate effort. This means the system produces not just solutions but research prioritization — agents deciding that one direction looks more tractable than another, pivoting when a line of attack stalls, building on each other’s partial results.

The outputs are not trivial. Across 12 construction problems and two case studies, the Station produced novel results on five of them:

  • A new infinite family of finite-field Kakeya sets. The Kakeya conjecture (in its various finite-field and Euclidean forms) has attracted serious attention for decades; a new infinite family is not a benchmark score, it’s a mathematical contribution.
  • New 604-point kissing configurations in dimension 11. Kissing number problems — how many non-overlapping unit spheres can touch a central sphere in nn dimensions — are hard and well-studied.
  • Improved records for the discretized Kakeya needle and sign uncertainty problems.
  • A substantially improved lower bound for Erdős’s minimum-overlap problem.
  • Novel infinite families for Book Ramsey numbers.

What makes these results interesting beyond the raw novelty is that the agents generated proofs and analyses alongside the numerical constructions. This matters for mathematicians trying to build on the work — a construction accompanied by a theorem explaining why it works is far more useful than a black-box numerical answer. The paper’s release package includes complete agent conversations, proofs, and verification code.

The multi-agent, multi-model design is worth noting. Rather than running many instances of one model, the Station involves agents from different families, which means different biases, different blind spots, and different exploratory tendencies. When one agent’s approach to a problem stalls, a different model’s prior on what to try next may be different enough to find a path forward. This is speculative reasoning about why it works — the paper’s analysis is what it is — but it’s at least plausible that heterogeneity in the agent pool contributes to the breadth of problems where forward progress happened.

The honest question is how to evaluate claims like these. “Novel relative to the prior literature” is the standard the authors apply, and they’ve released the verification code, so the mathematical community can check. But multi-agent systems produce a lot of output, and some fraction of “discoveries” in any such system will be rediscoveries of known results that weren’t in the training data or the comparison corpus. The release of agent conversations is the right call here — it makes the discovery process auditable rather than asking readers to trust a benchmark table.

What this points at, if the results hold up, is a shift in how AI math tools get used. The current paradigm treats AI as a proof-search accelerator for human-directed goals. A system that can autonomously identify tractable problems, allocate exploration effort, and generate interpretable results starts to look like a research collaborator rather than a search engine. Whether the Station generalizes beyond the specific problem classes tested is the next question — combinatorial construction problems may have structural properties that make them more amenable to automated search than other areas of mathematics. But the output is specific enough to be falsifiable, and that’s already more than most “AI for math” announcements offer.