AI Co-Mathematician (Google DeepMind)¶
external · status: active · focus: end-to-end · discipline: general · started: 2026
Project page: https://arxiv.org/abs/2605.06651
Source: projects/landscape/ai-co-mathematician.yml
Positioning¶
A closed, agentic multi-agent workbench (arXiv:2605.06651) built on Gemini 3.1 for open-ended mathematics research. A hierarchy of specialized agents runs under a top-level project coordinator across asynchronous, parallel workstreams — ideation, literature search, computational exploration, theorem proving, theory building — managing uncertainty, tracking failed hypotheses, and producing LaTeX write-ups with margin annotations and provenance notes. Explicitly modelled on agentic coding environments (e.g. Claude Code), it spans the full pipeline from research intent to a written mathematical result rather than optimizing a single stage.
Distinctive contribution¶
One of the first agentic systems to demonstrate open-ended mathematical discovery rather than benchmark problem-solving alone: it scored 48% (23/48) on FrontierMath Tier 4 — versus 19% for the Gemini 3.1 Pro base model — and Oxford mathematician Marc Lackenby used it to resolve Problem 21.10 of the Kourovka Notebook, a group-theory question open since 1965. The RISE skills catalog's theorist-toolbox (co-math skills) explicitly emulates this system's coordinator + workstream + reviewer-gate design in an open, skills-as-Markdown form.
Evaluation scores¶
| Dimension | Score (0–3) | Note |
|---|---|---|
| Lifecycle coverage | 2 | Seven stages from refining research intent through proving to LaTeX write-up; no dedicated referee-simulation or dissemination stage. |
| Autonomy level | 2 | Designed as a collaborative workbench (mathematician sets direction and reviews artifacts); async workstreams run without per-step approval, and it solved FrontierMath Tier 4 problems autonomously. |
| Architectural transparency | 1 | Paper describes the hierarchical agent topology and coordinator/workstream design, but no code, prompts, or configs are released. |
| Inputs supported | 2 | Accepts both well-posed open problems and open-ended research directions, with literature search and computational-exploration tooling. |
| Outputs / reproducibility | 2 | Persists LaTeX write-ups plus code from computational exploration with provenance notes; closed system and LLM nondeterminism preclude external end-to-end reproduction. |
| Internal evaluation | 2 | Systematic benchmark result (48% on FrontierMath Tier 4) plus an expert-verified new result (Kourovka Notebook 21.10); still an arXiv preprint, not peer-reviewed. |
| Openness | 0 | Closed research system; no code, weights, or public access at scoring date. |
| Maturity / traction | 1 | DeepMind research prototype (arXiv, May 2026), not a product; demonstrated real use by an external mathematician but no general availability. |
| Cross-family policy | 0 | Single model family — all agents run on Gemini 3.1. |
| Runtime assurance | 2 | In-pipeline review cycles, failed-hypothesis tracking, provenance annotations, and proof/computational verification act as moderate runtime gates. |
| Cross-platform portability | 0 | Closed system locked to DeepMind infrastructure and Gemini; not deployable elsewhere. |
Scored on 2026-07-23. See the evaluation rubric.
Tags¶
Pipeline stages: rq-formulation hypothesis-generation literature-discovery literature-synthesis formal-modeling code-generation paper-drafting
Architectural features: multi-agent human-in-loop tool-use persistent-memory iterative-loop artifact-versioning
Inputs: open-problem research-direction
Outputs: proofs latex-writeups computational-results literature-references
Knowledge sources: mathematical-literature
Limitations¶
- Closed system: no code, prompts, or public access — architecture is known only from the paper, and the FrontierMath and Kourovka results cannot be independently reproduced.
- Benchmarked and demonstrated in pure mathematics; transfer to empirical or applied research disciplines is unshown.
- Autonomous benchmark performance still leaves the majority of Tier 4 problems unsolved, and headline discoveries required an expert mathematician in the loop.
Related projects in this catalog¶
Papers describing this project¶
- AI co-mathematician: Accelerating mathematicians with agentic AI — Zheng, D., von Glehn, I., Zwols, Y., Beloshapka, I., Buesing, L., Roy, D. M., et al. (2026). arXiv (Google DeepMind). arXiv:2605.06651