Use agents to probe open mathematical questions

A group built a math environment where agents chose their own leads, ran tests, and shared findings without a central planner.

Done with Custom agent· Advanced

What they did
The authors made Station, an open setting where AI agents from different model families worked on one common research aim. Rather than following a fixed sequence, each agent could pick a direction, try experiments, compare notes with others, and contribute to a shared pool of background work. They assessed it on 12 AlphaEvolve-catalogue construction problems and two case studies.
How it went
The system produced new-to-the-literature outcomes on five problems, including results tied to finite-field Kakeya questions, 604-point kissing configurations in dimension 11, improved records in discretized Kakeya needle and sign uncertainty, and a stronger Erdős minimum-overlap bound. The agents also found Book Ramsey families and wrote explanatory theorems.
Tips
This approach fits problems where broad search matters more than a single fixed plan. Because the source code, raw agent chats, proofs, and verification code were published, you can check how the results were assembled.

Try it yourself with Manus

You are working in a shared math-research environment with no central manager. Begin by listing [problem area] questions that seem worth testing, then choose a few directions and explain the first experiments or checks for each. Use the shared workspace to exchange observations with other agents and build on their findings. If a result would be irreversible, such as submitting a proof, publishing a conclusion, or deleting a line of work, stop and ask for approval first. If a path stops paying off, switch to another and note what changed.

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Read the original on arxiv.org ↗

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.