---
title: "Gold-medalist Performance in Solving Olympiad Geometry with AlphaGeometry2"
authors:
  - "Y. Chervonyi"
  - "T. H. Trinh"
  - "M. Olsak"
  - "X. Yang"
  - "Y. Wu"
  - "N. Hoang"
  - "A. Ho"
  - "T. Luong"
  - "Google DeepMind"
year: 2025
venue: "arXiv preprint"
arxiv: "2502.03544"
doi: null
source: "https://arxiv.org/abs/2502.03544"
topics:
  - "mathematical-reasoning"
  - "theorem-proving"
  - "imo"
seed_rank: 1222
seed_batch: "alpoge-math-2026-08-24"
reviewed: "2026-08-24"
pool: "agents"
relevance_score: 10
lineage: ai-mathematical-reasoning
cites:
  - title: "Solving olympiad geometry without human demonstrations"
    url: "https://arxiv.org/abs/2312.12272"
    year: 2024
    arxiv: "2312.12272"
    doi: "10.1038/s41586-023-06747-5"
see:
  - "1221-solving-olympiad-geometry-without-human-demonstrations"
---

# Gold-medalist Performance in Solving Olympiad Geometry with AlphaGeometry2

## One-sentence takeaway

AG2: Gemini LM plus expanded domain language plus SKEST search-tree sharing — 84% of IMO 2000-2024 geometry (was 54%), past average gold, and the geometry half of IMO 2024 silver.

## Why it matters here

The named upgrade. Coverage 66% to 88%; IMO-AG-50 solve 42/50.

## Problem

AG1 language missed motions, linear angle/ratio/distance equations, non-constructive problems.

## Design

Gemini architecture, faster DDAR, SKEST (shared knowledge ensemble of search trees), more synthetic data. Progress toward natural-language input.

## Evidence

arXiv:2502.03544, Feb 2025. Code: google-deepmind/alphageometry2. IMO 2024 P4.

## Limitations

Still geometry. Combinatorics half of IMO 2024 stayed open (see AlphaProof).

## Implications for Broadside

See 1221 and 1223. One card for AG2, not a blog duplicate.

## Bottom line

Mint 2502.03544.

## Links

- arXiv: [2502.03544](https://arxiv.org/abs/2502.03544)