AlphaProof Nexus
#7 v Reasoning modelygoogle-deepmind · od 21. Mai 2026 (arXiv-Preprint 2605.22763v1) · 17× · naposledy 30. 6. 2026
AlphaProof Nexus is a research framework released by Google DeepMind (not a commercial product) that pairs the Gemini 3.1 Pro LLM with the Lean proof assistant to autonomously generate and machine-verify mathematical proofs. It uses a multi-tier agent design (four variants ranging from a basic LLM-compiler loop to evolutionary search with Elo rating) and can optionally call the earlier RL-based AlphaProof system as a subagent. According to the accompanying arXiv preprint (May 21, 2026), the system autonomously solved 9 of 353 open Erdős problems and 44 of 492 open OEIS conjectures, at an inference cost of a few hundred dollars per problem. All generated Lean proofs are publicly available on GitHub under the Apache 2.0 license.
Vlastnosti
| Key Benchmark (%) | 9/353 open Erdős problems solved (~2.5%); 44/492 OEIS conjectures proven (~9%) |
| License | Apache License 2.0 (results repository on GitHub) |
| Multimodality | Not multimodal – text-based Lean proof generation and verification |
| Platform | Research framework: Gemini 3.1 Pro + Lean compiler (sandboxed), optional AlphaProof subagent; code/proofs on GitHub |
| Price per 1M Tokens | Not specified; only total inference cost of 'a few hundred dollars per problem' documented |
| Release Date | May 21, 2026 (arXiv preprint 2605.22763v1 published) |