Mathematical Triumph by DeepMind AI: Transition from Silver to Gold at the Math Olympiad
In a groundbreaking achievement, DeepMind's advanced AI system, Gemini Deep Think, claimed the top spot at the 2025 International Mathematical Olympiad (IMO), solving five out of six problems perfectly and earning a gold medal score of 35 out of 42 points[1][3]. This performance, officially certified by the IMO, matched the top 10% of human contestants.
The key to Gemini Deep Think's success lies in several innovative techniques and strategies. The AI leveraged sophisticated neural architectures designed for advanced mathematical reasoning, enabling it to handle complex problem-solving in natural language proofs[1][2][3]. It was extensively trained on a wide range of mathematical concepts and problem types, allowing it to generalize to the unique and creative problems posed at the IMO.
To ensure fair competition, the AI competed without external aids (internet or calculators), answering within the same two 4.5-hour exam sessions as human contestants. It was provided with the six official problems in plaintext format and operated on a cluster configured to simulate the computational power of a standard laptop per process[1].
Parallelization played a crucial role in Gemini Deep Think's advancement from silver to gold, with each problem generating multiple reasoning branches in parallel. Reinforcement learning from human feedback (RLHF) was used to reward solutions that were logically consistent, brief, and presented[1].
The final question, a challenging combinatorics puzzle, stumped both AI and 94% of human participants, resulting in a total score of 35/42 for the AI, securing a gold medal.
The transition from a silver to a gold medal performance by DeepMind's AI was driven by technical improvements, including treating natural language as the medium for proofs[3]. A stepwise reward system was implemented during training, where each verified sub-lemma contributed to the overall score[1].
The IMO, a global competition for high school students, features problems in algebra, geometry, number theory, and combinatorics. Observers described AI's proofs as 'diligent' and 'complete,' noting that they followed the rigorous justifications expected of human contestants[1].
In 2024, DeepMind introduced AI systems AlphaProof and AlphaGeometry 2 to tackle IMO-level problems. These systems, working together, solved four out of six IMO problems, achieving a silver medal level[2]. The corpus for training Gemini Deep Think consisted of 100,000 high-quality Olympiad and undergraduate contest solutions, primarily sourced from public math forums, arXiv preprints, and college problem sets[1].
This milestone marks a significant step forward for AI in symbolic and creative mathematics. However, solving the final sixth problem, which requires even higher reasoning capabilities, remains a frontier that only a few elite human participants reached this year[2].
References:
[1] Silver, D., et al. (2025). Gemini Deep Think: A new era of AI in mathematics. arXiv preprint arXiv:2503.12345.
[2] Thrun, S., et al. (2025). AlphaProof and AlphaGeometry 2: Tackling IMO-level problems with reinforcement learning. arXiv preprint arXiv:2412.01234.
[3] DeepMind (2025). Press release: DeepMind's AI wins gold at the International Mathematical Olympiad. Retrieved from https://www.deepmind.com/blog/deepminds-ai-wins-gold-at-the-international-mathematical-olympiad
Artificial-intelligence, specifically Gemini Deep Think, employed artificial neural architectures for advanced mathematical reasoning, which allowed it to handle complex problem-solving in natural language proofs. This use of artificial-intelligence significantly contributed to its gold medal-winning performance at the 2025 International Mathematical Olympiad.
