What is AlphaGeometry2
AlphaGeometry2 is an advanced artificial intelligence system launched by Google DeepMind, specifically designed to solve geometry problems in the International Mathematical Olympiad (IMO). It combines a neuro-symbolic approach, coordinating the Google Gemini series language models with a symbolic engine, predicting geometric constructions through neural networks and performing logical reasoning with the symbolic engine. AlphaGeometry2 has achieved an 84% solution rate on IMO geometry problems from the past 25 years, surpassing the average level of gold medalists. Core upgrades include an extended domain-specific language, a more powerful symbolic reasoning engine DDAR2, a brand-new search algorithm SKEST, and a more powerful language model. It can handle more complex geometry problems, including those involving object motion and angle equations. AlphaGeometry2's training data is generated by DeepMind itself, containing more than 300 million theorems and proofs of varying complexity.
AlphaGeometry2's main features
- Solving complex geometry problems: AlphaGeometry2 can solve 84% of geometry problems in the IMO over the past 25 years, surpassing the level of the average gold medalist.
- Predicting geometric constructions: The system combines the Google Gemini series language models with a symbolic engine. The Gemini model can predict the geometric constructions needed to solve the problem (such as points, lines, and circles), while the symbolic engine performs deductions based on mathematical rules.
- Expanded problem coverage: The system extends the original AlphaGeometry language and can handle harder problems involving object motion as well as linear equations of angles, ratios, and distances.
AlphaGeometry2's technical principles
- Neuro-symbolic approach: AlphaGeometry2 adopts a neuro-symbolic approach, combining the Google Gemini series language models with a symbolic engine. The Gemini model predicts the geometric constructions needed to solve the problem (such as points, lines, and circles) through a neural network architecture, while the symbolic engine performs reasoning and proof based on strict mathematical rules.
- Symbolic reasoning engine DDAR2: The core of the symbolic engine is DDAR2 (Deductive Database Arithmetic Reasoning), an algorithm for computing deductive closure. DDAR2 can start from a set of initial facts and, through fixed inference rules, gradually derive all possible facts until no further derivation is possible. DDAR2's improvements include the ability to handle coincident points, a faster algorithm implementation (changed from Python to C++, with a 300x speed increase), and more efficient rule application.
- Parallel search algorithm SKEST: AlphaGeometry2 uses the SKEST (Search Tree with Knowledge Sharing Ensemble) search algorithm. The algorithm runs multiple search trees with different configurations in parallel, where each node corresponds to an auxiliary construction attempt and a run of the symbolic engine. If an attempt succeeds, all search trees terminate; if it fails, the successfully proven facts are recorded in a shared fact database for use by other nodes.
- Synthetic data training: To address the lack of training data for solving geometry problems, DeepMind generated more than 300 million theorems and proofs of varying complexity for training.
AlphaGeometry2's project address
- arXiv technical paper：https://arxiv.org/pdf/2502.03544
AlphaGeometry2's application scenarios
- Math competitions: AlphaGeometry2 solves geometry problems in the International Mathematical Olympiad (IMO). It becomes a powerful tool for researching and training math competition contestants.
- Mathematics education: AlphaGeometry2 can serve as an auxiliary tool in mathematics education, helping students and teachers better understand and solve complex geometry problems. By demonstrating the problem-solving process and logical reasoning, it can provide examples for students to learn from and practice with.
- Mathematics Research: The technology of AlphaGeometry2 can be extended to the field of mathematics research, in scenarios requiring complex geometric reasoning and proofs. It can provide mathematicians with new ideas and methods to help solve unsolved geometry problems.
- Formal Mathematical Reasoning: Combined with other AI models such as AlphaProof, AlphaGeometry2 can be used for formal mathematical reasoning.
- Scientific and Engineering Computing: The technology of AlphaGeometry2 can be extended to the fields of science and engineering, for example, providing geometric reasoning support in complex engineering computations.