Google DeepMind's AlphaProof Hits 84.1% on FrontierMath, Nearing Olympiad Silver Level

Google DeepMind has unveiled a major advance in automated mathematical reasoning with AlphaProof, a system that achieved 84.1% on the new FrontierMath benchmark and solved 83 of 100 problems from the 2024 International Mathematical Olympiad. The performance places the system at roughly silver medal level, highlighting how far AI has progressed on elite competition mathematics.

According to a Nature paper published today, AlphaProof combines large language models with reinforcement learning to tackle complex proof-based problems. The system is designed not just to generate answers, but to search for valid mathematical arguments and verify them through formal reasoning. That approach marks an important step toward automated theorem proving, a long-standing challenge in artificial intelligence.

FrontierMath was created to test systems on especially difficult problems drawn from advanced mathematical competition settings. AlphaProof’s score suggests that AI models are beginning to handle tasks once considered far beyond machine capability, though the benchmark remains highly demanding.

DeepMind’s result adds to a growing body of research showing that general-purpose AI techniques can be adapted for rigorous symbolic reasoning. While the system is not yet a replacement for human mathematicians, the breakthrough points to new possibilities for assisting research, education, and formal verification in mathematics.