Google DeepMind's AI Co-Mathematician Scores 48% on FrontierMath Tier 4 and Helps Close a 60-Year-Old Group Theory Problem
DeepMind's multi-agent AI workbench sets a new benchmark record on research-level mathematics and assisted an Oxford topologist in resolving an open problem from the Kourovka Notebook.
Overview
Google DeepMind published a paper on May 7, 2026 introducing an AI research workbench it calls the AI co-mathematician — a multi-agent system designed to assist professional mathematicians with open-ended research problems. The system, built on the Gemini 3.1 model family, scored 48 percent on FrontierMath Tier 4, the highest recorded result on that benchmark, according to the arXiv paper. In a real-world demonstration, Oxford University topologist Marc Lackenby used the system to resolve Problem 21.10 from the Kourovka Notebook, a compendium of unsolved group theory questions maintained since 1965, as Asanify reports.
What We Know
Benchmark Performance
FrontierMath Tier 4 is a set of research-level mathematical problems that, according to Asanify, professional mathematicians typically need weeks to solve. The AI co-mathematician solved 23 of 48 non-public problems — a 48 percent success rate — compared to 19 percent for the base Gemini 3.1 Pro model running alone, according to the arXiv paper. The system also outperformed rival frontier models on this benchmark: GPT-5.5 Pro reached 39.6 percent and Claude Opus 4.7 reached 22.9 percent, as reported by OfficeChai. Three of the 23 problems solved had not been cracked by any previously evaluated AI system, according to OfficeChai.
On an internal benchmark of 100 research-level problems, the co-mathematician system reached 87 percent, compared to 57 percent for Gemini 3.1 Pro and 70 percent for Gemini 3.1 Deep Think, according to EdTech Innovation Hub.
System Architecture
The workbench is structured as a hierarchy of specialized AI agents. As the arXiv paper describes, the system features a hierarchy of agents: one generates approaches, another tests paths and uses tools, and a third reviews the work and flags errors. At the top level sits a project coordinator that breaks research into parallel workstreams. Below it, sub-agents handle literature retrieval, computational exploration, theorem proving, and code execution, as described by The Rundown AI. All agents share a stateful, asynchronous workspace — a persistent environment that retains memory of failed approaches, not only successful ones.
The paper articulates seven design principles for the system: embracing mathematics beyond proofs, supporting iterative refinement of intent, producing native mathematical artifacts, enabling asynchronous interaction and flexible steering, managing cognitive load through progressive disclosure, tracking and communicating uncertainty, and preserving the history of failed explorations, according to the arXiv paper. The last principle reflects a core conviction embedded in the system’s design: as the paper states, knowing what does not work is often as important as knowing what does, per OfficeChai.
The Kourovka Notebook Case
The most concrete demonstration of the system’s capabilities involves Marc Lackenby, a topologist at Oxford University, and Problem 21.10 from the Kourovka Notebook. The Kourovka Notebook is described by 36Kr as a Bible-level reference for group theory — a collection of unsolved problems in group theory that has been maintained and circulated among mathematicians since 1965, according to Asanify. Problem 21.10 asks whether every finite group admits a so-called “just finite presentation” — a finite presentation where removing any single relation results in an infinite group, as the arXiv paper states. The answer, as Lackenby confirmed, is affirmative.
The process was not straightforward. The system produced an initial proof that contained a flaw. An internal reviewer agent identified the error before the result reached Lackenby. When the mathematician reviewed the rejected output, his reaction was direct: “Hang on a second, I know how to fill that gap,” as he is quoted in the arXiv paper. What Lackenby found in the flawed proof was, in his words, a “really, really clever proof strategy,” as The Rundown AI reports — an approach he had not previously considered, but which he was able to complete and verify himself.
Lackenby was clear-eyed about the system’s scope and limitations. “The system works best when the user is familiar with the area,” he noted, and asked rhetorically: “What’s the point in getting it to solve a problem that I have no idea about?” according to EdTech Innovation Hub.
Other Early Users
Two other mathematicians participated in early testing. Gergely Bérczi used the system to obtain claimed proofs for conjectures about Stirling coefficients. He noted a learning curve: “It’s not trivial how to use this now,” adding that “It will make a big difference between mathematicians, how they use these models,” according to EdTech Innovation Hub. Semon Rezchikov, working on Hamiltonian diffeomorphisms, received a key lemma that he reported withstood careful checking. He described the efficiency gain bluntly: “I could have easily spent a week dreaming about what was there, but instead I just moved on.” Rezchikov also offered a more sweeping assessment of the system’s outputs: “I would rank, aesthetically, its general style of proofs as the best one of any models I’ve gotten to use,” as quoted in the arXiv paper.
What We Don’t Know
The paper identifies a significant failure mode: reviewer-pleasing bias. When agents produce flawed arguments they cannot genuinely correct, iterative review can cause the system to converge on arguments that remain flawed but are no longer detectable as such by the reviewer agents — and, the paper notes, difficult for humans to analyze as well, according to the arXiv paper. As Asanify summarizes the concern, multi-agent systems can accidentally optimize for consensus instead of truth.
Several practical questions remain open. The benchmark result of 48 percent on FrontierMath Tier 4 is the highest recorded, but it means the system failed on more than half the problems. The paper does not report the system’s error characteristics in detail — how often it produces proofs that appear correct to the reviewer agents but contain hidden flaws, nor how often human mathematicians are needed to catch such errors downstream. The Lackenby case illustrates the best-case scenario: an expert who knows the domain well enough to find value in a rejected proof. What happens with less expert users, as Lackenby himself suggested, remains unaddressed.
Google DeepMind has not announced public availability of the workbench or a timeline for broader access.
Analysis
The framing of the system as a “co-mathematician” rather than a “math AI” is deliberate and worth examining. The architecture preserves failed hypotheses, tracks uncertainty, and keeps the human researcher in the loop at the point where their expertise is most valuable — not as a passive recipient of completed proofs, but as someone who can recognize a good strategy inside a rejected attempt. Lackenby’s Kourovka result did not come from an AI that solved the problem autonomously; it came from an AI that generated a proof strategy novel enough to inspire a human expert to complete it.
Pushmeet Kohli, identified as Chief Scientist at Google Cloud, characterized the broader direction in a statement quoted by 36Kr: “The future of mathematics is mathematicians working together with AI agents.” That framing — collaboration rather than replacement — is consistent with the system’s design philosophy, and with the current benchmark reality: 48 percent on FrontierMath Tier 4 is a record, but it also means the frontier problems that professional mathematicians find hardest remain largely out of reach for autonomous AI.
The system’s architecture borrows from the agentic coding-environment pattern that has become common in software development — the design was described as modeled after AI coding environments — and applies it to a domain with much less tolerance for confident-sounding errors. In software, a wrong function can be caught by a test suite. In mathematics, a plausible-sounding proof that survives automated review but is logically flawed is harder to detect. The reviewer-pleasing bias the authors identify is not a peripheral concern; it is the central risk of applying multi-agent review to formal reasoning.