AlphaGeometry

• 1. System Architecture And Design

  ›

  DeepMind released an artificial intelligence program named AlphaGeometry in 2024. This system combines two distinct technologies to solve Euclidean geometry problems. A data-driven large language model works alongside a rule-based symbolic engine called Deductive Database Arithmetic Reasoning. Traditional programs rely exclusively on human-coded rules to generate rigorous proofs. Such rigid systems lack flexibility when facing unusual situations. The new design allows the symbolic engine to request help from the large language model. When the engine fails to find a formal proof, it solicits suggestions for geometric constructs. These suggestions allow the process to move forward toward a solution.

  13456

• 2. International Mathematical Olympiad Performance

  ›

  The program solved 25 out of 30 International Mathematical Olympiad geometry problems within competition time limits. This performance level is almost as good as the average human gold medallist. Previous AI attempts struggled significantly with these same challenges. An older program known as Wu's method managed to solve only 10 problems. The gap between the old and new results highlights a major leap in capability. DeepMind tested the system against questions that had been used in past competitions. The strict time constraints ensured the results reflected genuine problem-solving speed rather than unlimited calculation time.

  26

• 3. Evolution From Symbolic Engines

  ›

  Traditional geometry programs are symbolic engines that rely exclusively on human-coded rules. AlphaGeometry combines such an engine with a specialized large language model trained on synthetic data. The training set consists of geometrical proofs generated specifically for this purpose. When the symbolic engine does not manage to find a formal and rigorous proof on its own, it solicits the large language model. The model suggests a geometric construct to help the system proceed. Researchers note that it remains unclear how applicable this method is to other domains of mathematics. Symbolic engines rely on domain-specific rules which limit their broader transferability. The need for vast amounts of synthetic data also presents a significant hurdle for future applications.

  135

• 4. Publication In Nature Journal

  ›

  DeepMind published a paper about AlphaGeometry in the peer-reviewed journal Nature on the 17th of January 2024. The research detailed the architecture and performance of the system. MIT Technology Review featured the program on the same day as the publication. This timing coincided with the release of findings regarding the International Mathematical Olympiad results. The scientific community received the announcement with interest given the high difficulty of the problems solved. Peer review processes ensured the claims were verified before public dissemination. The document serves as the primary record of the system's initial capabilities and design choices.

  7

• 5. Introduction Of AlphaGeometry Two

  ›

  An improved version named AlphaGeometry 2 was published on the 5th of February 2025. Developers added more features to the representation language to describe complex geometry problems. These updates allow the system to handle movements of objects and linear equations involving angles. The team targeted IMO geometry questions from 2000 to 2024 during development. The expanded representation language allowed them to cover 88% of the questions in that range. It uses Gemini finetuned on a synthetically generated dataset of problems and solutions. The model assists by making auxiliary constructions like lines and points to help tree search algorithms. Autoformalization converts problems written in English into the specific representation language used internally.

  8