AI Triumph in Mathematics: DeepMind's AI Emerges Victorious at the Mathematics Olympiad, Moving from Silver to Gold
DeepMind's AI system, named Gemini Deep Think, claimed a gold medal at the International Mathematical Olympiad (IMO) in 2025. The AI solved 5 out of 6 extremely difficult Olympiad problems under official contest conditions, scoring 35 out of 42 points – the cutoff for a gold medal [2][3].
Key Innovations in Gemini Deep Think
Several significant technical improvements drove the rapid advancement from DeepMind's silver to gold medal performance.
Massive Parallelization and Dynamic Scheduling
Gemini Deep Think generated multiple reasoning branches for each problem in parallel and dynamically shifted computational resources toward the most promising lines of reasoning. This approach resembled how human mathematicians explore auxiliary inequalities or lemmas before committing to a full solution step [2].
Frozen Model Weights Before Competition
To maintain fairness and prevent any data leakage from official IMO problems, DeepMind froze the model three weeks before the event and filtered out previously unpublished Olympiad question solutions from its training data [2]. During the competition, the AI received the problems in plaintext with no internet access or external tools, perfectly replicating the exam conditions [2].
Natural-Language Proof Generation
Gemini Deep Think produced full natural-language proofs comparable in quality to those written by top human contestants. Expert graders, including former IMO medalists, unanimously scored the proofs, confirming their rigor and completeness [1][2].
Computational Resource Simulation
During the exam, the AI was run on a cluster simulating the computational power equivalent to a standard laptop per process, adhering to constraints that made the achievement comparable to human effort within the exam’s time limit of roughly nine hours in two sessions [2].
The Road to Success
The training process for Gemini Deep Think spanned three months and utilized approximately 25 million TPU-hours. The training corpus was mainly collected from public math forums, arXiv preprints, and college problem sets [2].
In 2024, DeepMind introduced two AI systems, AlphaProof and AlphaGeometry 2, to tackle IMO-level problems. The training process involved fine-tuning the model to predict next steps using a corpus of 100,000 high-quality Olympiad and undergraduate contest solutions. Human mentors reviewed training examples to filter illogical or incomplete proofs [2].
Despite stumbling on the final question, which was a challenging combinatorics puzzle, AI finished with a total score of 35/42 to secure a gold medal [2].
Implications for AI and Mathematics
This gold medal performance showed a significant improvement from DeepMind's silver medal achievement in 2024, demonstrating rapid progress in AI’s mathematical reasoning capabilities [2]. It marks a milestone in AI research, indicating the potential of large language models combined with parallel reasoning techniques to tackle complex, abstract mathematical problems at a world-class level [1][3].
The IMO, established in 1959, is a premier mathematics competition for high school students. Solving IMO problems requires mathematical creativity, logical thinking, and the ability to construct elegant proofs. With Gemini Deep Think's success, it is clear that AI is making strides toward more human-like cognitive abilities and becoming invaluable tools for mathematicians.
The question of whether AI will replace or enhance human creativity remains unresolved, but the 2025 IMO is a clear indication that artificial intelligence has made significant strides in logical reasoning.
Technology and artificial-intelligence were integral to DeepMind's success at the International Mathematical Olympiad (IMO) in 2025, as demonstrated by the AI system, Gemini Deep Think. To produce full natural-language proofs comparable to top human contestants, Gemini Deep Think employed parallel reasoning techniques and large language models, showcasing the potential of these technologies in tackling complex, abstract mathematical problems at a world-class level.
