Google DeepMind team has recently officially launched a new AI framework called "AlphaProof Nexus". This system, with its powerful artificial intelligence reasoning and formal verification capabilities, has made milestone breakthroughs in the field of mathematical research, successfully solving two Erdős mathematical problems that had remained unsolved for 56 years.

image.png

Original Four-Level Agent Architecture Collaboratively Breaks Through

AlphaProof Nexus features an innovative architecture composed of four AI agents with increasing complexity. The system starts with the basic Gemini model and Lean compiler interaction, gradually introduces code completion snippet mechanisms and evolution mechanisms similar to AlphaEvolve, and finally enables multiple proof drafts to be shared, rated, and sorted through the most complete agent.

In autonomous testing on 353 open Erdős problems, the system not only solved 9 problems but also proved 44 open conjectures in the Online Encyclopedia of Integer Sequences. What is astonishing is that AlphaProof Nexus only requires hundreds of dollars to perform reasoning on a single complex problem.

Compiler Feedback Anchoring Drives AI Evolution

Aside from achieving significant results in combinatorial mathematics and number theory, the system also successfully solved a Hilbert function problem that had existed for 15 years and significantly improved known bounds in convex optimization. When analyzing experimental data, the research team found that even the simplest base agent could solve some problems with the "anchoring" effect of the powerful Lean compiler feedback.