Google DeepMind’s AlphaProof Nexus solved open math problems and went through peer review

Google DeepMind’s AlphaProof Nexus solved dozens of open math problems, according to a study published in Science. The reported results include nine of 353 open Erdős problems, with two described as unsolved for 56 years. The key takeaway for readers: this is being presented not just as AI-assisted writing, but as formal proof work that passed through review before publication.

The system is built around Gemini 3.1 Pro, which drafts and revises proofs in Lean, a formal language used to verify mathematical arguments. Lean checks each draft after edits, and error messages guide the next attempt. For harder sub-problems, AlphaProof Nexus can call on AlphaProof, though The Decoder reports it solved the nine Erdős problems even without AlphaProof.

The report lands during a broader dispute over how AI-generated math proofs should enter the field. The Decoder contrasts DeepMind’s peer-reviewed route with OpenAI’s publication of hundreds of AI proofs on GitHub, which prompted a boycott call from the Association for Human Mathematics led by Terence Tao. For researchers and readers, the practical distinction is process: formal verification can help check proof syntax and logic, while peer review adds external scrutiny before claims are published.

AlphaProof Nexus also proved 44 conjectures from the OEIS integer sequence database, solved an open question in algebraic geometry, and improved a known bound in min-max optimization. Those details make the work relevant beyond a single problem set, especially for people tracking AI systems that can interact with formal mathematical tools. The next question is how mathematicians, journals, and AI labs standardize disclosure, verification, and review for machine-generated proofs.