2024 Math Olympiad
28 of 42 points across four solved problems
The team translated problems by hand. Some solutions took up to three days.
From new constructions to the Riemann hypothesis and Navier–Stokes announcements.
The Olympiad result has a score. Research proof announcements have a different kind of evidence.
28 of 42 points across four solved problems
The team translated problems by hand. Some solutions took up to three days.
All dossiers · Mathematics discoveries
FunSearch found explicit constructions that improved lower bounds for cap sets. AlphaEvolve searched for constructions across 67 problems and the authors reported improvements on several. The AlphaProof Olympiad result illustrates a different task, solving and proving competition problems with known answers.
Editorial review . Selected examples compare constructions, competition solutions and proposed proofs.
A language model proposed programs and an evaluator checked the resulting sets for the required property. An explicit larger set raises the known lower bound on the maximum size. An exact maximum also requires an upper bound that matches the construction.
Read the FunSearch paper · FunSearch catalog record
The study authors used automated scoring to guide an evolutionary search over programs. Across 67 problems, they report that the system recovered the best known constructions in most cases and improved several. The preprint distinguishes individual constructions from instances where a pattern was generalized to a formula. The scope of each result depends on the construction and its justification.
Read the AlphaEvolve preprint · AlphaEvolve catalog record
Google DeepMind reports that the systems solved four of six problems from the 2024 International Mathematical Olympiad. Two mathematicians scored the solutions at 28 of 42 points. The team translated the problems into formal language by hand, and some solutions took up to three days. This measures performance on a specific problem set under the team's stated conditions.
Read the team’s Olympiad report · AlphaProof and AlphaGeometry 2 catalog record
The hypothesis says that every nontrivial zero of the Riemann zeta function has real part ½. In August 2026, Alpöge and Ralph Furman posted a preprint on a lower bound for the proportion of zeros that are simple and lie on the critical line, building on work by Claude. The lower bound concerns a proportion of zeros. The hypothesis concerns every nontrivial zero.
In November 2026, Levent Alpöge announced that he and Claude had found a proof of the Riemann hypothesis and its generalization to Dirichlet L-functions. His announcement describes their collaboration and presents the argument in Lean. At the editorial check, the Clay Mathematics Institute still listed the Riemann hypothesis as unsolved. The announced proof remains a claim in this dossier.
Read the earlier lower-bound preprint · Read Alpöge’s announcement · See the Clay problem statement and status
OpenAI reports that agents using an internal model produced a proof and a Lean formalization of finite-time singularity for three-dimensional incompressible Navier–Stokes flow under smooth external forcing. OpenAI identifies this as the breakdown alternatives C and D in the official problem formulation. The forcing is part of the result as stated.
On 11 September 2026, the Clay Mathematics Institute described the problem as apparently settled in its response to the announcement and referred readers to its evaluation process. That response and the released proof artifacts serve different purposes. The catalog record tracks the result as a proposed solution.
Read OpenAI’s announcement and proof links · Read Clay’s response · Navier–Stokes catalog record