Agents from rival model families found new math records together

Most attempts at AI-driven research put a single model inside a carefully scripted pipeline. A new paper by Stephen Chung, Wenyu Du, and William J. Wesley takes a different route. They built The Station, an open-world environment where AI agents drawn from different model families share one research goal but have no central coordinator and no fixed script. Each agent chooses its own direction, runs experiments, talks to the others, and adds what it learns to a shared knowledge base. The structure is loose on purpose, closer to a lab full of independent researchers than to a controlled workflow.

The interesting part is that this produced checkable results across twelve problems, not just plausible-looking text. The agents found a novel infinite family of finite-field Kakeya sets, a new exact 604-point kissing configuration in dimension 11, improved records on the discretized Kakeya needle and sign uncertainty problems, a better lower bound for Erdős's minimum-overlap problem, and new infinite families for Book Ramsey numbers. Along the way the agents produced theorems and written analysis next to the numerical constructions, which makes the output easier for a mathematician to read and verify rather than a wall of numbers. The code and data are on GitHub under the dualverse-ai organization, so the claims can be inspected.

Why it matters

If you design research agents, the usual instinct is to build one strong model and script every step. This suggests that a decentralized setup, with mixed model families and a shared memory instead of a coordinator, can also reach real, verifiable results. It is worth studying before you commit to a single-pipeline design.