AlphaGeometry2 - Google DeepMind launches AI system for solving complex geometric problems

What is AlphaGeometry2?

AlphaGeometry2 is an advanced artificial intelligence system developed by Google DeepMind, specifically designed to solve geometry problems in the International Mathematical Olympiad (IMO). It combines a neurosymbolic approach, integrating Google's Gemini language models with a symbolic engine. Neural networks predict geometric structures, while the symbolic engine performs logical reasoning. AlphaGeometry2 has achieved an 84% success rate in solving IMO geometry problems over the past 25 years, surpassing the average level of gold medalists. Core upgrades include an expanded domain-specific language, a more powerful symbolic reasoning engine (DDAR2), a new search algorithm (SKEST), and a more robust 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 and contains over 300 million theorems and proofs of varying complexity.

Main functions of AlphaGeometry2

• Solving complex geometric problemsAlphaGeometry2 can solve 84% of the geometry problems in the IMO over the past 25 years, surpassing the average gold medalist's performance.
• Predictive geometryThe system combines Google's Gemini language model and symbolic engine. The Gemini model can predict the geometric structures needed to solve problems (such as points, lines, and circles), while the symbolic engine derives them based on mathematical rules.
• Expanding the scope of the problemThe system extends the original AlphaGeometry language to handle more complex problems involving object motion and linear equations involving angles, scales, and distances.

The technical principle of AlphaGeometry2

• Neural symbolic methodAlphaGeometry2 employs a neurosymbolic approach, combining Google's Gemini language models and symbolic engine. The Gemini model uses a neural network architecture to predict the geometric structures (such as points, lines, and circles) needed to solve problems, while the symbolic engine performs reasoning and proof based on strict mathematical rules.
• Symbolic reasoning engine DDAR2The core of the symbolic engine is DDAR2 (Deductive Database Arithmetic Reasoning), an algorithm for calculating deductive closure. DDAR2 starts from an initial set of facts and gradually derives all possible facts through fixed reasoning rules until no further derivation is possible. Improvements to DDAR2 include the ability to handle overlapping points, a faster algorithm implementation (switching from Python to C++, resulting in a 300x speedup), and more efficient rule application.
• Parallel Search Algorithm SKESTAlphaGeometry2 uses the SKEST (Knowledge-Shared Ensemble-Based Search Tree) search algorithm. The algorithm runs in parallel through multiple search trees with different configurations, with each node corresponding to one auxiliary construction attempt and the execution of the symbol engine. If the attempt succeeds, all search trees terminate; if it fails, the successfully proven fact is recorded in a shared fact base for use by other nodes.
• Synthetic data trainingTo address the lack of training data for geometry problems, DeepMind generated over 300 million theorems and proofs of varying complexity for training.

AlphaGeometry2 project address

• arXiv technical paper:https://arxiv.org/pdf/2502.03544

Application scenarios of AlphaGeometry2

• Mathematics CompetitionAlphaGeometry2 solves geometry problems from the International Mathematical Olympiad (IMO). It has become a powerful tool for researching and training mathematics competition participants.
• Mathematics EducationAlphaGeometry2 can serve as a supplementary tool in mathematics education, helping students and teachers better understand and solve complex geometric problems. By demonstrating problem-solving processes and logical reasoning, it provides students with examples for learning and practice.
• Mathematical researchAlphaGeometry2's technology can be extended to mathematical research, particularly in scenarios requiring complex geometric reasoning and proof. It can provide mathematicians with new ideas and methods to help solve unsolved geometric problems.
• Formal Mathematical ReasoningCombined with other AI models such as AlphaProof, AlphaGeometry2 can be used for formal mathematical reasoning.
• Scientific and engineering computingAlphaGeometry2's technology can be extended to scientific and engineering fields, such as providing geometric reasoning support in complex engineering calculations.