Google DeepMind's AI Co-Mathematician Tackles Unsolved Problems and Sets New Benchmark Record

Google DeepMind has published a paper titled "AI Co-Mathematician: Accelerating Mathematicians with Agentic AI," describing a system that goes far beyond a standard chatbot — giving researchers a stateful workspace where multiple AI agents run parallel workstreams, track uncertainty, preserve failed attempts, search literature, test ideas, and produce mathematical working documents. The system runs on Gemini 3.1 and is organized hierarchically: a project coordinator at the top, workstream coordinators below managing literature review, library development, and counterexample search, with specialized agents — including a search agent, a coding agent, and Gemini Deep Think acting as a proof verifier. The system reports a 48 percent score on FrontierMath Tier 4 — a new high score among all AI systems evaluated. Oxford mathematician Marc Lackenby used it to resolve an open problem from the Kourovka Notebook after a reviewer agent spotted a flaw in the AI's first proof attempt. Authors are candid about failure modes: the review cycle between agents can converge on subtly flawed arguments — a "reviewer-pleasing bias" — or spiral into endless disagreement between agents.