Google DeepMind's AlphaProof Nexus solved dozens of open math problems independently, according to a peer-reviewed study published in Science. The system successfully cracked nine of 353 open Erdős problems, including two that had remained unsolved for 56 years, after the team initially shared the solutions online.
The AI system 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. Built around Gemini 3.1 Pro, AlphaProof Nexus drafts and revises proofs in the formal language Lean, which checks each draft and uses error messages to guide subsequent attempts. For more difficult sub-problems, it leverages the specialized proof tool AlphaProof.
This peer-reviewed publication contrasts with recent industry approaches, such as OpenAI releasing hundreds of AI proofs on GitHub rather than submitting them to academic journals. Google DeepMind chose the traditional peer-review route for its findings. This mathematical milestone demonstrates significant progress in automated reasoning and formal verification for complex problem-solving.