Google DeepMind AI Reaches Silver-Medal Standard in Math Olympiad

Google DeepMind has developed two AI systems capable of solving complex mathematical problems at a competitive human level. The models, AlphaProof and AlphaGeometry 2, achieved a silver-medal standard by solving four out of six problems from the 2024 International Mathematical Olympiad (IMO).

The systems demonstrated high-level proficiency across algebra, number theory, and geometry. The two models worked in tandem to tackle these advanced reasoning tasks. A notable highlight of the performance was AlphaGeometry 2's efficiency; once the problem was formalized, the system solved IMO Problem 4 in just 19 seconds.

The Challenge of Formal Reasoning

For years, artificial intelligence has struggled with "system 2" thinking—the slow, deliberate, and methodical reasoning required to construct a formal mathematical proof. While standard large language models (LLMs) can often mimic the structure of a mathematical argument, they are prone to "hallucinations," where they invent logical steps that are mathematically unsound.

To overcome this, DeepMind integrated LLMs with reinforcement learning and formal verification languages, specifically Lean. This architecture ensures that every step of a proof is logically verified, preventing the errors common in purely generative AI and allowing the system to iterate until a mathematically sound solution is reached.

Implications for Science

This achievement signals a transition for AI, moving beyond simple pattern recognition toward genuine logical reasoning. The ability to verify and generate formal proofs suggests that AI could eventually move from solving existing problems to automating the discovery of entirely new mathematical theorems.

Such a capability has profound implications for the broader scientific community. If AI can reliably handle formal reasoning at this level, it may eventually be applied to solve long-standing conjectures in computer science and physics, accelerating the pace of theoretical discovery by handling the rigorous verification that typically takes human mathematicians years to complete.

The Path Forward

While the silver-medal performance is a significant leap, the systems did not solve all six problems, leaving a gap between current AI capabilities and the gold-medal standard. Future development will likely focus on increasing the breadth of problems the AI can formalize and improving its ability to tackle the most abstract reasoning tasks without human intervention. As these systems evolve, the integration of formal verification will remain the critical guardrail against the logical instability of generative models.