Station’s autonomous AI agents report new results on five mathematical construction problems
- Station, an open-world multi-agent environment without a central coordinator or scripted pipeline, reports results new to prior literature on five problems across 12 AlphaEvolve construction problems and two case studies.
- Agents found a new infinite family of finite-field Kakeya sets, exact 604-point kissing configurations in dimension 11, and new records for the discretized Kakeya needle and sign uncertainty problems.
- The system also reports a substantially improved lower bound for Erdős's minimum-overlap problem and novel infinite families for Book Ramsey numbers.
- Rather than returning only numerical constructions, agents produced theorems and analyses intended to explain the constructions and support later mathematical work.
- The authors released the raw agent dialogues, proofs, verification code, and project source code, providing artifacts for checking how the reported discoveries emerged.
Hacker News 의견들
I think objections that this is not "original" miss the professional mathematicians' point. Vast accessible memory plus rapid testing and recombination of known approaches will outdo much of the mathematical community; that may count as a form of creativity, though perhaps not an exhaustive one.
I'd like to see Station make some of its reward system endogenous. Keep the final mathematical evaluator external, but let agents create prizes, peer-review rules, journals, reputation systems, or compute-allocation rules, then compare that with the architect-defined system. It might create useful specialization, or just herding.
I'm curious how agent-generated status signals work when any single agent is short-lived because of limited context. Humans who built status systems usually have much longer continuity.
I think the review loop is the important part: different models critique each other's work and reach consensus. You need adversarial work alongside creative work.
I'm torn on anthropomorphizing systems with terms like "thinking" and "holidays." It distorts expectations, but repeated use may also strip away the unjustified magic people assign to human reasoning and creativity.
After working on custom harnesses, I found terminology matters because these concepts are new. I dropped "agent" in favor of "thread" for that reason.
The paper describes the open-source Station project at github.com/dualverse-ai/station.