Google DeepMind、数学研究を加速する「AI for Math Initiative」始動 Accelerating discovery with the AI for Math Initiative

Google DeepMind has unveiled the AI for Math Initiative, a new collaboration designed to accelerate mathematical discovery using its most advanced AI systems. The program brings together five leading research institutions: Imperial College London, the Institute for Advanced Study, the Simons Institute for the Theory of Computing, the Tata Institute of Fundamental Research, and the University of California, Berkeley.

Under the initiative, mathematicians at partner institutions will gain access to a suite of DeepMind's frontier tools. These include Gemini DeepThink, a reasoning-focused variant of Gemini designed for extended deliberation; AlphaProof, a system specialized in formal theorem proving; and AlphaEvolve, an evolutionary coding agent capable of searching for novel algorithms and mathematical constructions. The goal is to support researchers across pure and applied mathematics in formulating conjectures, exploring open problems, and constructing rigorous proofs.

The announcement builds on a striking trajectory of AI progress in mathematics. In 2024, DeepMind's AlphaProof and AlphaGeometry 2 together achieved a silver-medal score on problems from the International Mathematical Olympiad. In 2025, Gemini DeepThink reportedly reached gold-medal level performance on IMO problems, signaling that general-purpose reasoning models are catching up with bespoke math systems. AlphaEvolve, meanwhile, has already been credited with discovering improved algorithms for matrix multiplication and other long-studied problems.

The initiative also reflects a broader shift in the mathematical community toward AI-assisted research. Fields Medalist Terence Tao and others have publicly embraced tools like the Lean proof assistant, GitHub Copilot, and large language models as everyday collaborators. Formalization projects such as the Liquid Tensor Experiment and ongoing efforts to formalize parts of the Fermat's Last Theorem proof demonstrate how machine-checkable mathematics is becoming a genuine research methodology rather than a niche curiosity.

DeepMind is not alone in this space. OpenAI has invested heavily in math benchmarks like FrontierMath, Meta's research arm has explored neural theorem proving, and startups such as Harmonic are pursuing reliable mathematical reasoning as a core product. The competitive landscape suggests that mathematics, long viewed as the most rigorous test of machine intelligence, has become a key proving ground for the next generation of AI systems.

Questions remain about how AI-generated proofs should be validated, how credit should be assigned between human and machine collaborators, and whether opaque model outputs can truly contribute to mathematical understanding rather than just answers. Still, by pairing top mathematical institutions with cutting-edge models, the AI for Math Initiative could meaningfully shorten the path from conjecture to theorem, and may help establish norms for how AI is integrated into foundational scientific research going forward.