AI co-mathematician: Accelerating mathematicians with agentic AI

Summary

The AI co-mathematician is a Google DeepMind workbench through which mathematicians interactively use AI agents for open-ended research. It is designed to support the exploratory, iterative reality of mathematical work — ideation, literature search, computational exploration, theorem proving and theory building — through an asynchronous, stateful workspace that manages uncertainty, refines the user's intent, tracks failed hypotheses and outputs native mathematical artifacts, thereby mirroring human collaborative workflows. In early tests the system helped researchers solve open problems, identify new directions and recover overlooked literature references. It also reports state-of-the-art results on hard problem-solving benchmarks, including 48% on FrontierMath Tier 4, which the authors describe as a new high among all AI systems evaluated.

Contribution

Claimed: a highly interactive paradigm for AI-assisted mathematical discovery, plus benchmark leadership. What the abstract supports: the 48% FrontierMath Tier 4 figure and a qualitative account of early use. The "helped researchers solve open problems" claim is anecdotal in the abstract — no count, no protocol, no statement of how much of each result was the system's versus the mathematician's. The abstract says nothing about the agent architecture; the catalog entry's description of a coordinator with parallel workstreams on Gemini 3.1 comes from the paper body, not the abstract.

Method

As far as the abstract states: a system description, early tests with researchers, and evaluation on problem-solving benchmarks including FrontierMath Tier 4. The abstract does not specify the underlying models, the agent topology, the number of users or problems in the early tests, the evaluation protocol for FrontierMath (attempts, compute, verification), or which other AI systems it was compared against. The arXiv listing (v2, May 2026, "23 pages; several citations added") carries no journal reference.

Relevance to RISE

The abstract's own list of supported activities maps onto rq-formulation ("refines user intent"), hypothesis-generation, literature-discovery, formal-modeling, code-generation ("computational exploration") and paper-drafting ("native mathematical artifacts"), consistent with the catalog's tags. Catalog slug: ai-co-mathematician; the catalog also notes that RISE's own theorist-toolbox skills emulate its coordinator-plus-workstream design in open form. Compared with google-co-scientist, whose abstract foregrounds a tournament-evolution process for hypotheses, this abstract foregrounds interactivity and persistent state rather than any competitive aggregation. For the ISR question of structure and epistemic quality, the only structural mechanism the abstract itself names is a stateful workspace that "tracks failed hypotheses" — a memory-based rather than adversarial mechanism — while the catalog entry (scored from the full paper) records in-pipeline review cycles; FrontierMath Tier 4 is attractive as an outcome measure for structure comparisons because correctness is checkable, unlike open-ended discovery claims.

Critique / open questions

The system is closed (catalog entry: no code, prompts or public access), so nothing in the abstract can be independently reproduced. "Early tests" are unquantified, and the abstract cannot tell us how much of the headline discoveries depended on the expert in the loop. The 48% figure leaves most Tier 4 problems unsolved, and "among all AI systems evaluated" leaves the comparison set unspecified. The abstract concedes no limitations. Transfer beyond pure mathematics to empirical disciplines is not addressed.

Key quotes

"We introduce the AI co-mathematician, a workbench for mathematicians to interactively leverage AI agents to pursue open-ended research." (abstract)

"By providing an asynchronous, stateful workspace that manages uncertainty, refines user intent, tracks failed hypotheses, and outputs native mathematical artifacts, the system mirrors human collaborative workflows." (abstract)

"Besides demonstrating a highly interactive paradigm for AI-assisted mathematical discovery, the AI co-mathematician also achieves state of the art results on hard problem-solving benchmarks, including scoring 48% on FrontierMath Tier 4, a new high score among all AI systems evaluated." (abstract)