AlphaGeometry, a new AI system, has achieved a significant milestone by solving complex geometry problems at a level comparable to a human International Mathematical Olympiad (IMO) gold medalist. Developed by researchers Trieu Trinh and Thang Luong, AlphaGeometry successfully tackled 25 out of 30 Olympiad geometry problems within the standard time limit, nearly matching the average human gold medalist score of 25.9 and far surpassing the previous best AI system, which solved 10.
The system employs a neuro-symbolic approach, combining a neural language model with a symbolic deduction engine. The language model generates intuitive ideas, while the symbolic engine applies formal logic to deduce solutions, mimicking the “thinking, fast and slow” concept. This collaboration allows the AI to predict and add useful geometric constructs to diagrams, facilitating the deduction process.
AlphaGeometry’s training involved generating and filtering a vast dataset of 100 million unique synthetic geometry examples, enabling it to learn without human demonstrations. Its solutions are verifiable, clean, and utilize classical geometry rules, making them human-readable.
This breakthrough not only showcases AI’s growing logical reasoning capabilities but also opens up new possibilities for AI applications in mathematics and science. The system’s code and model have been open-sourced, encouraging further advancements in the field.
Read more at Google DeepMind…