From a recent arXiv preprint,
We introduce FrontierMath, a benchmark of hundreds of original, exceptionally challenging mathematics problems crafted and vetted by expert mathematicians. The questions cover most major branches of modern mathematics -- from computationally intensive problems in number theory and real analysis to abstract questions in algebraic geometry and category theory. Solving a typical problem requires multiple hours of effort from a researcher in the relevant branch of mathematics, and for the upper end questions, multiple days. FrontierMath uses new, unpublished problems and automated verification to reliably evaluate models while minimizing risk of data contamination. Current state-of-the-art AI models solve under 2% of problems, revealing a vast gap between AI capabilities and the prowess of the mathematical community. As AI systems advance toward expert-level mathematical abilities, FrontierMath offers a rigorous testbed that quantifies their progress.
This question resolves to YES if the score that AlphaProof or some later version of it achieves on the FrontierMath benchmark, as reported prior to midnight, January 1st 2026 Pacific Time, is above 30.0%. Credible reports include but are not limited to blog posts, arXiv preprints, and papers. Otherwise, this question resolves to NO.
I will use my discretion in determining whether a result should be considered valid. Obvious cheating, such as including the test set in the training data, does not count.
As for what "AlphaProof or some later version of it" refers to:
the intent is roughly to catch any AI build before 2026 that has a similar enough architecture to AlphaProof that it's taken to be a successor, and was developed by GDM
credit goes to @mathvc :
See also these markets:
/MatthewBarnett/will-an-ai-achieve-85-performance-o
1,000
3.00Arg, I regret having put this up to 40% (and think people rightly traded in the other direction again). Somehow I didn't read this properly and thought it asked about *any* system doing it before 2026. AlphaProof is a good contender, but it might just get one more version before 2026 or similar, so not many attempts.
The title was meant to hold. Has anyone traded thinking the description was meant to hold? If so I'll N/A and make it again. I guess I'll @traders
@RyanGreenblatt good question. the intent was roughly to catch any AI build before 2026 that has a similar enough architecture to AlphaProof that it's taken to be a successor, and was developed by GDM. I'm not sure how to define that more precisely, or whether once announced that AI's architecture would be known; I'm open to suggestions. AlphaProof 2 would count, but the intent is to also be robust to arbitrary naming choices.
on the other market you brought up that AlphaProof might not be equipped to do stuff like find a numerical solution effectively, and needs questions formalized in Lean. I only know the basic idea behind Lean and this seems like reasonable reasons to think this market would resolve NO, but presumably we'd want the spirit of the market to be specified into something like, we allow humans to translate the problems into Lean before letting the AI go at it? Or alternatively we'd be relying on some future variation on AlphaProof that is able to make that translation. I made the market not considering those alternatives, but if you think either or some third alternative is best lmk.