AI solves math problems on par with international Olympic gold medalists

AlphaGeometry2, an artificial intelligence (AI) math problem solver, has solved 42 out of 50 geometry problems in the International Mathematical Olympiad (IMO) over the past 25 years, while the average score of gold medalists is 40.9.

The information was published in Nature - the world's leading scientific journal, last weekend. The AlphaGeometry2 author team works at Google DeepMind, consisting of 10 people, including 4 Vietnamese.

This is an upgraded version of AlphaGeometry - which solved 53% of problems when it was first released.

In 2024, research teams in India and China used other approaches to get AI to achieve IMO gold medal-level results in geometry, but not as many tests as AlphaGeometry2.

AlphaGeometry combines a specialized language model and a “symbolic engine” that specializes in logical reasoning. The model is trained to express standard mathematical language, ensuring logical rigor and eliminating incoherent or incorrect statements that AI robots are prone to make.

AlphaGeometry2 has several improvements, integrating Google's state-of-the-art large language model - Gemini. This model supports a "symbolic engine", which uses mathematical rules to deduce solutions to problems, giving possible proofs for a given geometric theorem.

Geometry Olympiad problems are based on diagrams that require additional “structures” before they can be solved, such as points, lines, or circles. AlphaGeometry2's Gemini model predicts what structures might be useful to add to a diagram, which the tool references to make inferences.

Future improvements will include solving problems involving inequalities and non-linear equations, aiming to "solve the entire geometric problem," the team said.

Buzzard assesses that AI still has many other challenges to overcome before solving research-level problems.