AlphaGeometry
Neuro-symbolic geometry solver that solved 25 of 30 olympiad problems, versus 10 for the previous best system.
A language model proposes auxiliary constructions and a symbolic deduction engine verifies them. Trained on 100 million synthetic examples derived from one billion random diagrams, with no human proofs. Code and model released.
- Date
- Wednesday, 17 January 2024
- Lab
- Google DeepMind
- Kind
- paper
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| IMO geometry problems solved (30-problem benchmark) | 25/30 The previous best was 10 | company |
Published in Nature 2024-01-17. Gemini-based AlphaGeometry 2 followed (see IMO 2024 record).
Sources
- deepmind.google/blog/alphageometry-an-olympiad-level-ai-system-for-geometry/
- en.wikipedia.org/wiki/AlphaGeometry
This record was checked against its sources on 6 October 2026. How we check