Google’s Gemini AI cracks olympiad gold standard in math

Google DeepMind’s advanced AI system has achieved gold-medal standard performance at the International Mathematical Olympiad (IMO), solving problems once considered far beyond the reach of machines.

The announcement this week marks a breakthrough in AI’s ability to reason through complex symbolic tasks, often seen as the final frontier of human intelligence.

The feat was accomplished using an advanced version of Gemini, Google’s multimodal AI model, in conjunction with DeepMind’s mathematical research team.

The system solved five out of six problems from past IMO papers, scoring 35 out of 42, surpassing last year’s silver-level score of 28 when DeepMind used AlphaGeometry 2 and AlphaProof in tandem.

The IMO is widely regarded as the most challenging math competition for pre-university students, drawing elite talent from over 100 countries.

Achieving gold typically places a contestant in the top 8% globally.

AI models performing at this level signal a dramatic leap in formal reasoning, going beyond language prediction and pattern recognition.

In a blog post, researchers Thang Luong and Edward Lockhart said the Gemini-based system was trained using a new framework that combined symbolic logic, formal theorem proving and language-based learning, allowing it to “understand and prove” mathematical concepts rather than just pattern-match them.

This development comes amid growing debate over whether today’s AI systems truly understand the tasks they perform.

Google’s results suggest that at least in formal mathematics, systems like Gemini are beginning to approach human-level insight.

While the achievement won’t immediately impact everyday AI products, it could have long-term implications for scientific research, software verification, and education.

Google is positioning its success as a benchmark moment for “AI for science” initiatives, following earlier work on protein folding and weather prediction.

DeepMind researchers said their next goal was to train AI that can assist mathematicians in making new discoveries, not just solve old problems.