AlphaGeometry, developed by Google DeepMind, is a theorem-proving system that combines symbolic reasoning with deep learning to solve challenging geometry problems, such as those found in mathematical Olympiads. The repository provides the full implementation of DDAR (Deductive Difference and Abductive Reasoning) and AlphaGeometry, two automated geometry solvers described in the 2024 Nature paper “Solving Olympiad Geometry without Human Demonstrations.” AlphaGeometry integrates a symbolic deduction engine with a transformer-based language model to propose and validate geometric constructions in a stepwise proof process. The DDAR solver focuses purely on rule-based reasoning, while AlphaGeometry enhances this by using a learned model to suggest auxiliary constructions when logical reasoning alone is insufficient. The repository includes pre-trained weights, vocabulary files, and detailed configuration options for reproducing experiments.
Features
- Implements DDAR and AlphaGeometry, two state-of-the-art geometry theorem provers
- Combines symbolic logic and transformer-based language modeling for geometric proofs
- Includes pre-trained model weights and vocabulary files for reproducibility
- Provides complete examples for solving IMO-style geometry problems
- Modular Python code with explicit definitions, rules, and reasoning utilities
- Fully scriptable environment for testing, benchmarking, and extending theorem solvers
Project Samples
