AI agents running unsupervised in an open-world research environment produced results novel to the mathematical literature on five of twelve construction problems taken from the AlphaEvolve catalogue, according to a paper from the team behind the Station. The agents also independently rediscovered a counterexample to the Jacobian conjecture within a day, and the team published every raw agent dialogue, proof and verification script alongside the claims.
Key facts
- Across 12 construction problems from the AlphaEvolve catalogue plus two case studies, the Station produced results novel relative to prior literature on five problems.
- The novel results include a new infinite family of finite-field Kakeya sets, new 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 Erdos's minimum-overlap problem.
- The agents come from different model families and work with no central coordinator and no scripted pipeline.
- Primary source: Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment, arXiv 2608.23691, with code at dualverse-ai/station.
Most multi-agent research systems are pipelines wearing a costume: a planner hands work to a coder who hands results to a critic, and the interesting decisions were made by the person who drew the diagram. The Station is built the other way. Agents from different model families are dropped into a shared environment, choose their own research directions, run their own experiments, collaborate when they want to, and write into a shared scientific literature the others can read. There is no coordinator deciding who works on what.
The design constraint the team is explicit about is worth noting for anyone tempted to copy it: the Station suits tasks that are scorable, meaning each run can be evaluated with a clear number, and fast, meaning each run finishes in roughly two hours. Mathematical constructions fit perfectly. You are hunting for an object -- a set, a configuration, a bound -- and whether you found one is not a matter of taste.
The kissing-number result is the easiest to picture. Ask how many identical balls can touch one central ball without overlapping. In two dimensions the answer is six, and you can check it with coins on a table. In eleven dimensions nobody knows, and progress comes from explicitly constructing arrangements that push the known lower bound up. The Station's public log shows that bound climbing over months -- 600 touching balls in June, then 604 -- with the construction notebook published each time.
The claim that separates this from a search script is about explanation. "Agents also discovered novel infinite families for Book Ramsey numbers," the authors write, and note that the agents "produced not only numerical constructions but also theorems and analyses explaining how those constructions work, making the results more interpretable and easier for mathematicians to build upon." A brute-force search returns an object. A collaborator returns an object plus an argument for why the pattern continues, and only the second is something a mathematician can extend.
Why it matters: this is the strongest current evidence that multi-agent systems can be more than an expensive way to run one model several times, and it lands the same week as two AI-assisted elliptic-curve rank records and follows the Station's earlier Jacobian-conjecture counterexample. Mathematics keeps being the proving ground because the verification is free and merciless.
The honest caveats: the AlphaEvolve catalogue is a curated set of construction problems chosen because they are amenable to machine search, so five out of twelve is a score on a friendly board rather than a claim about mathematics generally. Five novel results also means seven that were not, and the paper's own accounting includes problems where the agents did worse than the published state of the art. Running the Station requires API keys for commercial model providers and the OpenAI Codex CLI, so the compute bill is real and unpublished. The mitigating factor is transparency: the v2 data viewer and data repository put the full research trail in the open, which is more than most agent papers offer.
The Station has a public track record worth checking rather than taking on faith. Its news log shows the eleven-dimensional kissing-number bound moving from 600 in June, alongside a novel algebraic family for a book-Ramsey task, to 604 later that month, each with a published construction notebook. The v1 system was described in an earlier paper in November 2025. Watching a lower bound tick upward over months in public, with the artifacts attached each time, is a very different kind of evidence from a single announcement claiming a breakthrough.
Originally published on Ground Truth, where every claim is checked against the primary source.
