DeepMind has introduced AlphaProof Nexus, a multi-agent framework powered by Gemini 3.1 Pro designed to solve complex mathematical problems. The system uses independent subagents that iterate through reasoning loops, with results verified by automated proof checkers.
The framework successfully resolved 9 out of 353 Erdős problems, including two that remained unsolved for 56 years. Additionally, it proved 44 out of 492 conjectures from the Online Encyclopedia of Integer Sequences. The computational cost for these results is reported at a few hundred dollars per problem.
A separate system, Aletheia, built on Gemini Deep Think, has been active since December 2025. It has evaluated approximately 700 Erdős problems, identifying or creating solutions for 13 of them, four of which were novel autonomous resolutions.
Gemini Deep Think also demonstrated performance at the 2025 International Mathematical Olympiad, scoring 35 out of 42 points and solving 5 out of 6 problems within standard time limits. DeepMind researchers suggest these agentic methods could be applied to fields like combinatorics, quantum optics, and algebraic geometry.
