AI Research

DeepMind AlphaProof Nexus Solves Open Math Problems in Science

Oct 12, 2026·4 min read·By Rahish Saifi

$Cover: DeepMind AlphaProof Nexus Solves Open Math Problems in Science — AI with Rahish$

On October 8, 2026, the journal Science published Google DeepMind’s paper “Advancing mathematics research with AI-driven formal proof search” (doi 10.1126/science.aej2213). The work describes AlphaProof Nexus, a framework in which large language model agents search for proofs in the Lean proof assistant, Lean checks each draft, and failed attempts feed back into the loop. A preprint appeared on arXiv as 2605.22763 on May 21, 2026; the peer-reviewed release is what turned a strong technical result into a milestone for AI-assisted mathematics—and a sharp contrast with labs that publish huge batches of unverified manuscripts.

How AlphaProof Nexus works

According to the paper and an AAAS EurekAlert release dated Oct 8, 2026, the system is built around Gemini 3.1 Pro for drafting and revising Lean proofs. A basic agent loop can propose proof steps and learn from Lean’s error messages. An optional full-featured agent coordinates subagents through an evolutionary algorithm and can call AlphaProof—DeepMind’s specialized reinforcement-learning prover—as a tool.

That split matters for trust: every accepted step must pass Lean’s kernel. The model does not get credit for a convincing narrative alone. When a proof fails, the feedback is mechanical, not rhetorical.

What the team claims to have proved

The authors, led by George Tsoukalas and colleagues, with correspondence from Pushmeet Kohli and Swarat Chaudhuri, report several result classes:

Erdős problems: 9 of 353 open Erdős problems were solved autonomously, including two that had been open for more than 50 years (56, per The Decoder)—Erdős problem #12(i) and #12(ii), posed by Erdős and Sárközy in 1970.
OEIS conjectures: 44 of 492 open conjectures from the On-Line Encyclopedia of Integer Sequences were proved after Gemini autoformalized them; manual review judged the statements correctly formalized and previously unproven.
Other mathematics: The paper also describes an improved bound and new step-size schedule for Anchored Gradient Descent-Ascent in min-max optimization, plus research deployments in algebraic geometry, quantum optics, and graph theory.

Lean proofs and artifacts are public in google-deepmind/alphaproof-nexus-results on GitHub.

Peer review, perspectives, and the verification bar

The same Science issue includes a Perspective by Jeremy Avigad and Matthew Ballard titled “AI for research mathematics has arrived,” arguing that even unsuccessful AI proof attempts can help mathematicians understand problems, and an Expert Voices piece by Emily Riehl. That editorial packaging underscores a difference from preprint dumps: DeepMind’s claims went through journal review, not only GitHub visibility.

The Decoder (Oct 12, 2026) framed the story around peer review, contrasting it with OpenAI’s decision to post hundreds of AI proofs on GitHub instead of submitting them to journals.

Contrast with unverified manuscript releases

On October 6, OpenAI posted 722 AI-written math manuscripts on GitHub from an unreleased internal model. About 162 include Lean formalizations, but the release is explicitly preprint-scale: not journal peer review, and not all families are machine-checked end to end. This site covered that drop in depth. AlphaProof Nexus stands out in the current news cycle because its headline results are Lean-verified in the loop and the core paper is peer-reviewed in Science—a higher bar for “this counts as established” than a repository README.

That contrast is not a verdict on every OpenAI manuscript; independent mathematicians still have to read both corpora. It is a difference in release posture: verified proofs plus journal publication versus volume-first disclosure.

Caveats worth keeping straight

None of this means mathematics is “solved” or that we have reached AGI. The Erdős and OEIS numbers are counts of specific formalized targets the system could attack, not a claim that all of combinatorics is closed.

Overlap with other labs: Community sites such as erdosproblems.com note that Erdős problems #741 and #846 were also found independently by an internal OpenAI model. For #846, erdosproblems.com also notes that a counterexample follows from earlier work of Reiher, Rödl et al. Some Nexus solutions are variants or parts of larger problems, not always wholly new theorems in every informal sense.
Ablation signal: The paper reports that a basic agent without the full evolutionary stack still solved all nine Erdős problems in an ablation—at higher cost on the hardest instances—so the core model-and-feedback loop carries real capability, not only the orchestration layer.
Human role: Autoformalization and problem selection still involve human judgment and review, especially for OEIS statements.

Why formal proof search matters now

As frontier models draft longer chains of reasoning, the bottleneck shifts from “can it sound right?” to “can we check it?” Lean and similar proof assistants offer a rare property: if the kernel accepts a proof, the logical steps are valid relative to the stated axioms. AlphaProof Nexus is an industrial-scale attempt to pair that checker with agentic search—the same direction Olympiad-style systems like AlphaProof explored, now extended to open research questions.

For developers, the practical lesson is architectural: treat the proof checker (or test suite, or type system) as the authority, and let the LLM be a propose-and-revise worker. DeepMind’s Oct 8 paper is evidence that loop can reach questions humans left open for decades—when the problem is formalized sharply enough for Lean to judge success.