Google’s DeepMind AI Can Solve Math Problems on Par with Top Human Solvers

Google’s AI can have the easiest smarts of mathematics

By Davide Castelvekchi & Nature Magazine

A year ago Alphageometry, the artificial problem created by Google Deepmind (AI) solved the world, the world In the silver level level of the Olympics of the international mathematics (Imo), a famous competition that brings together high competition competition.

The DeepMind Group now has the operation of its innovated system, Alphageometry2, has passed the level of the golden average dominist. The results are described in an advanced ARXIV.

“I think representations are not being marked by the computers imora,” Kevin Buzzard says, a mathematician in Imperial College London.

To help Science Journalism

If you enjoy this article, consider entering award-winning journalism Subscribe. By purchasing subscription, you are helping to ensure the future of stories about the discoveries and ideas that are conformed to today.

Solving Euclidean geometry problems is one of the four issues covered in IMO problems. The others cover the theory of numbers, algebra and combinatorial branches. The geometry requires specific Air skills, as competitors must provide strict evidence of statements about geometric objects on the aircraft. In July, Alphageometry2 made his public debut with the newly submitted system, Alphaproof, developed by Deepind to solve geometry questions developed in depth.

Mathematical language

Alphageometry is a combination of components that include specialized language models and neuro-symbolic system. The symbolic neural network is not trained, learning from data like networks, but humans have an abstract reasoning. The group trained the language model to speak a formal mathematical language, which allows its automatic departure to the logical rigor – and ‘hallucinations’, to be consistent or false expressions that tend to make AI Chatbots.

For Alphageometry2, the group improvement several improvements, including the integration of the great leading language model of Google, Gemini. The team also has the ability to reason about moving geometric objects around the plan. Move a point in a line in a line to change the height of a triangle and solve linear equations.

The system was able to resolve 84% of all Geometry issues in IMOS over the last 25 years compared to 54% of the first Alfageometry. (Groups in India and China used different perspectives last year, to get the gold medal level in geometry, but imo geometry in a lower subset of problems.)

Deepmind Paper’s authors write the improvements of future alfageometry to deal with the problems involving differences and nonlinear equations; They will be asked to “repair geometry to fully”.

Fast progress

To obtain a gold medal score for the general test, the US $ 5 million award can be called AI Olympiad Mathematical Award – even if this competition requires open sources, which is not the case of Deepmind.

Buzzard says he does not surprise Deepmind and quick progress made by Indian and Chinese teams. However, despite hard problems, the subject is still conceptually simple, and there are other challenges that are able to solve problems before air at the level of research.

AI researchers will be eagerly waiting for IMO’s next iteration, July, Australia, Australia, July. When his problems allow human participants to resolve, AI-based systems are also solved. (Alas agents do not participate in the competition and therefore cannot earn medals right. “Fresh problems is the most reliable test for machine learning systems.

This article reproduces with permission and has been First posted On February 7, 2025.