Skip to content
ContentLora

    Tip: press / anywhere to search.

    technology

    AI for mathematics: olympiad proofs, Lean and new results

    Also known as AI mathematics, automated theorem proving, AI theorem provers, AlphaProof, AlphaEvolve, FunSearch

    AI systems now solve olympiad problems at gold-medal level: an advanced Gemini Deep Think scored 35 of 42 points at the 2025 International Mathematical Olympiad, working in natural language within the time limit.[1][2] Google DeepMind reports that search systems pairing language models with automatic checkers have improved some known results in combinatorics and geometry,[3][4] but research-level benchmarks were far from solved when they were released.[5]

    Editor reviewedUpdated Wiki15 claims · 100% primaryAI for science, Artificial intelligence, Science
    Key facts

    Olympiad problems

    At IMO 2024, Google DeepMind‘s AlphaProof and AlphaGeometry 2 solved four of six problems for 28 of 42 points, a silver-medal score just below that year’s gold threshold of 29.[6] The problems were first translated by hand into a formal language, and some took the systems up to three days, against 4.5-hour sessions for students.[7]

    A year later an advanced version of Gemini Deep Think solved five of six problems perfectly, scoring 35 points, a gold score graded and certified by IMO coordinators.[1] It worked end-to-end in natural language and finished within the competition time limit.[2] The International AI Safety Report 2026 lists such olympiad results among the capability gains of general-purpose AI.[8]

    Formal proofs and Lean

    Lean is a free, open-source proof assistant, and mathlib is a community effort to formalise a unified library of mathematics in it.[9] Google DeepMind explains why this matters for AI: formal proofs can be checked for correctness automatically, while natural-language reasoning can include plausible but incorrect steps.[10] AlphaProof trains itself to prove statements in Lean, pairing a pre-trained language model with the AlphaZero reinforcement learning algorithm.[11]

    Searching for new constructions

    A second approach uses language models to write programs and an automatic evaluator to score them. FunSearch, announced in December 2023, works this way, with the evaluator guarding against incorrect ideas.[12] In some settings it found the largest cap sets yet known, which Google DeepMind called the biggest increase in cap set size in 20 years.[3]

    Its successor, AlphaEvolve, was applied to more than 50 open problems in 2025. Google DeepMind reports that it matched the best known solutions in roughly 75% of cases and improved on them in 20%.[4] Examples include multiplying 4x4 complex-valued matrices with 48 scalar multiplications, improving on Strassen’s 1969 method in that setting,[13] and a new lower bound for the kissing number problem in 11 dimensions.[14]

    Research-level mathematics

    The FrontierMath benchmark uses hundreds of new, unpublished problems with automatic answer checking.[5] Its authors say a typical problem takes a specialist several hours and the hardest several days,[15] and when it was released in November 2024 the best AI models solved under 2%.[5] For how such benchmarks are used more broadly, see how AI capabilities are measured.

    This page covers AI systems built or tested for mathematics. Results described here are as reported by the developers or benchmark authors; this page did not independently check any proof.

    More in AI for science

    Developing in AI & Computing

    Questions readers ask

    Has AI won a gold medal at the International Mathematical Olympiad?

    An AI system has reached gold-medal standard. At IMO 2025 an advanced version of Google DeepMind's Gemini Deep Think scored 35 of 42 points, and its results were graded and certified by IMO coordinators.[1]

    What changed between 2024 and 2025?

    In 2024 the problems had to be translated by hand into a formal language and some took up to three days to solve. In 2025 the Gemini model worked in natural language within the 4.5-hour time limit.[7][2]

    Why do some AI maths systems use Lean?

    Proofs written in a formal language such as Lean can be checked automatically, while natural-language reasoning can contain plausible but wrong steps.[10][9]

    Has AI discovered new mathematics?

    It has improved some known bounds. Google DeepMind reports that FunSearch found larger cap sets in some settings, and that AlphaEvolve set a new lower bound for the kissing number problem in 11 dimensions.[3][14]

    Can AI solve research-level maths problems?

    Not reliably as of late 2024. On the FrontierMath benchmark, whose problems take experts hours or days, the best models solved under 2% at release.[5][15]

    Sources

    · 15 claims from 4 publishers

    Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.

    1. [1]

      At IMO 2025 an advanced version of Google DeepMind's Gemini Deep Think solved five of six problems perfectly for 35 of 42 points, a gold-medal score officially graded and certified by IMO coordinators. confirmedas of 2025-07-21

    2. [2]

      Unlike in 2024, the 2025 Gemini model worked end-to-end in natural language, producing proofs from the official problem statements within the 4.5-hour competition time limit. confirmedas of 2025-07-21

    3. [3]

      Google DeepMind reported that FunSearch generated programs that in some settings found the largest cap sets ever found, which it called the largest increase in cap set size in 20 years. confirmedas of 2023-12-14

    4. [4]

      Google DeepMind reported in May 2025 that its AlphaEvolve coding agent, applied to more than 50 open problems, rediscovered the best known solutions in roughly 75% of cases and improved on them in 20%. confirmedas of 2025-05-14

    5. [5]

      FrontierMath, a benchmark released in November 2024, uses hundreds of new, unpublished research-level problems with automated verification, and at release state-of-the-art AI models solved under 2% of them. confirmedas of 2024-11-07

    6. [6]

      At the 2024 International Mathematical Olympiad, Google DeepMind's AlphaProof and AlphaGeometry 2 solved four of six problems for 28 of 42 points, the level of a silver medallist, one point below that year's gold threshold of 29. confirmedas of 2024-07-25

    7. [7]

      For IMO 2024 the problems were manually translated into formal mathematical language first, and the systems solved one problem within minutes but took up to three days on others, against the 4.5-hour sessions given to students. confirmedas of 2024-07-25

    8. [8]

      The International AI Safety Report 2026 found that general-purpose AI capabilities kept improving, especially in mathematics, coding and autonomous operation, including gold-medal performance on International Mathematical Olympiad questions. confirmedas of 2026-02-03

    9. [9]

      Lean is a free, open-source proof assistant, and mathlib is a community-driven effort to build a unified library of mathematics formalized in Lean. confirmedas of 2026-11-14

    10. [10]

      Google DeepMind notes that proofs in formal languages can be verified for correctness, whereas natural-language approaches can produce plausible but incorrect reasoning steps. confirmedas of 2024-07-25

    11. [11]

      Google DeepMind describes AlphaProof as a system that trains itself to prove mathematical statements in the formal language Lean, coupling a pre-trained language model with the AlphaZero reinforcement learning algorithm. confirmedas of 2024-07-25

    12. [12]

      FunSearch, announced by Google DeepMind in December 2023, pairs a pre-trained language model that proposes solutions as computer code with an automated evaluator that guards against incorrect ideas. confirmedas of 2023-12-14

    13. [13]

      Google DeepMind says AlphaEvolve found a way to multiply 4x4 complex-valued matrices using 48 scalar multiplications, improving on Strassen's 1969 algorithm in that setting. confirmedas of 2025-05-14

    14. [14]

      Google DeepMind says AlphaEvolve found a configuration of 593 outer spheres, establishing a new lower bound for the kissing number problem in 11 dimensions. confirmedas of 2025-05-14

    15. [15]

      The FrontierMath authors say a typical problem takes a researcher in the relevant field multiple hours, and the hardest ones multiple days. confirmedas of 2024-11-07

    Revision history (1)
    1. Page created.

    Created Nov 14, 2026. Last reviewed by an editor on Nov 14, 2026. Next scheduled review: May 14, 2027.

    Cite this page

    "AI for mathematics: olympiad proofs, Lean and new results." ContentLora, updated Nov 14, 2026. https://contentlora.com/wiki/ai-for-mathematics

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.