Google DeepMind develops AI capable of solving geometry problems

AI-generated picture [Image source: DALL·E]

Google DeepMind, in collaboration with researchers from New York University, has successfully developed an artificial intelligence (AI) capable of solving geometry problems from the International Mathematical Olympiad (IMO). The achievement is seen as a significant advancement in AI, particularly in tackling complex math problems that were traditionally challenging for AI systems.

The joint research team unveiled AlphaGeometry, an AI designed to solve geometry problems in the IMO. The research findings were published in a Nature journal on Wednesday.

Solving math problems has long been considered a formidable challenge for AI. Traditional AI models rely on recognizing patterns and predicting the next steps based on that pattern recognition. However, mathematical problem-solving demands rigorous inference and logic, making it difficult for conventional AI models to excel in this domain.

To address this, the Google research team developed a language model specifically for geometry-related proofs. The model was trained with geometric rules, such as “If line A intersects with line B, a line parallel to B will also intersect with A.”

Google applied AlphaGeometry to solve 30 geometry problems from the IMO. The AI correctly answered 25 of these problems, a substantial achievement compared to previous AI systems that struggled to solve even 10 problems.

The research team described AlphaGeometry as an AI capable of achieving a level equivalent to winning a gold medal in the IMO.

However, the researchers emphasized that AlphaGeometry is not yet capable of tackling highly advanced problems taught at the university level.