AI vs. Humans: Google's AI outperforms Math Olympiad gold medalists in geometry problems

Google's AI Outperforms Math Olympiad Gold Medalists in Geometry. (Image: Canva)

Google DeepMind has developed an artificial intelligence system that can solve complex geometry problems better than top human minds. Named AlphaGeometry2 (AG2), the AI has outshined International Mathematical Olympiad (IMO) gold medallists, marking a major step in AI's problem-solving abilities.

AlphaGeometry2: Smarter Than Gold Medalists
AG2 solved 84% of IMO geometry problems, surpassing the 81.8% solved by gold medalists. Designed by Google DeepMind, AG2 goes beyond simple pattern recognition. It blends logic with creativity, a rare trait in AI systems.

The findings were shared on February 7 in the arXiv preprint database. This comes just a month after Microsoft launched its AI math system, rStar-Math, which uses small language models (SMLs) to solve varied math problems.

Read Also: Mini galaxy at Solar System’s edge? NASA supercomputer uncovers twisting secrets of Oort Cloud

While rStar-Math aims for broad mathematical reasoning, AG2 focuses on mastering complex geometry. This makes AG2 unique, as geometry requires both visual thinking and logic, unlike algebra or calculus.

The updated version of AG2 shows a 30% improvement over its earlier model, first released in January 2024. Despite this success, experts caution against calling AG2 a step toward Artificial General Intelligence (AGI).

"AlphaGeometry2 shows intelligence, but it’s not human-like," said John Bates, CEO of SER Group. "Human intelligence involves invention, not just applying knowledge."

How AG2 Solves Complex Math Problems
AG2 merges neural language models with symbolic engines. The language model reads geometry problems and suggests possible constructions. The symbolic engine then tests these ideas, filtering out incorrect ones.

If one approach fails, the system tries another, working in parallel until it reaches a solution. This continuous back-and-forth mirrors how human mathematicians solve complex puzzles.

AG2 is faster and more accurate than its predecessor due to an upgraded data set and a quicker symbolic engine. It also uses a specialised algorithm to search for geometric proofs.

However, AG2 isn’t without flaws. It struggles with 3D geometry, non-linear equations, and problems involving variable or infinite points. Another major drawback is its inability to explain its solutions in human language.

The Future of AI in Mathematics
DeepMind aims to further refine AG2, improving speed, accuracy, and its ability to solve even tougher problems. The goal is to fully automate geometry problem-solving without errors.

While AG2 focuses on geometry, its technology could benefit fields like engineering, robotics, pharmaceuticals, and genomic research.

For now, AG2 remains a specialised tool. But its success hints at AI’s growing potential to solve complex human challenges—one equation at a time.