OpenAI's Math Prover Scores 41.2% on Tough Olympiad Benchmark

OpenAI has trained a neural network that can tackle real high-school math competition problems, the kind you see on the AMC12 and AIME, and even adapted problems from the International Mathematical Olympiad. They didn’t just build a calculator. They built a system that writes formal proofs in the Lean proof assistant, using a language model to generate the step-by-step reasoning. The raw numbers are striking: their prover hit 41.2% on the miniF2F benchmark, a curated set of tough problems. That blows the previous state-of-the-art of 29.3% out of the water.

The secret sauce is something they call ‘statement curriculum learning.’ It’s a neat feedback loop. You start with a language model that’s pretty weak at proving things and a collection of statements with no proofs, organized from easy to hard. The model searches for proofs, and every time it stumbles onto a correct one, that proof gets fed back into its training data. Then you retrain. The model gets a little smarter. Repeat that cycle eight times, and you end up with a prover that can handle problems it couldn’t touch before. It learns to generate the genuinely creative bits—like pulling the term ‘n+1’ out of thin air as a solution for a proof step—which is something you can’t do with a simple script.

The team frames formal math as a game-like domain, but one with a nasty twist. Unlike Go or chess, the ‘action space’ here is infinite. At any step, the model isn’t picking from a fixed set of moves; it has to generate a complex tactic and possibly invent original mathematical terms as arguments. The other problem is there’s no opponent. Self-play, which made AlphaGo a killer, falls apart when the challenge isn’t an adversary but a static, impossibly hard statement. Their workaround uses the language model’s generative chops to navigate that infinite space, and uses the iterative curriculum learning to simulate the kind of progression that self-play would otherwise provide. The prover has successfully chained together multiple non-trivial reasoning steps, like using contraposition and then conjuring a specific witness to finish a proof.

What’s genuinely new here is the emergent behavior. The model isn’t just pattern-matching. It’s starting to propose original mathematical objects as solutions, a skill that feels less like retrieval and more like a flicker of insight. Closing problems from the AMC12 and IMO isn’t just a party trick—it suggests these systems can grapple with the kind of creative reasoning that defines high-level math. It’s still a long way from tackling open problems in a research paper, but the jump from 29% to 41% on a tough benchmark is the kind of concrete progress that makes you pay attention.