Oriented on making people closer

Google's Gemini DeepThink artificial intelligence system has joined the distinguished list of systems to achieve the "gold" standard at the prestigious 2025 International Mathematical Olympiad (IMO). The model demonstrated an equivalently outstanding performance, solving 5 out of 6 problems and scoring 35 out of a possible 42 points. The solutions generated by the AI are available for review via the provided link.

Unlike OpenAI's experimental model, which has not yet disclosed the specifics of its approach to solving the 2025 IMO problems, Gemini DeepThink participated in the Olympiad as an "invited guest". The organizers of the 2025 IMO officially verified and confirmed all of Google's solutions, commending their clarity, accuracy, and comprehensibility. It is important to note that the infrastructure of Google's model itself was not independently audited by the organizing committee.

There is also a distinction in the release plans for these AI models. While OpenAI has emphasized that their model is experimental and will not be publicly available in the coming months, Gemini DeepThink is based on the existing Gemini 2.5 Pro platform. To prepare for the Olympiad, developers applied a novel reinforcement learning method and a specialized dataset of Olympiad problems. Google has announced that Gemini DeepThink will soon be available for testing to a select group of mathematicians, followed by all Gemini Ultra subscribers.

Despite the breakthrough in AI's computational capabilities, the absolute winners of the 2025 IMO remain human. Five students out of 641 participants (from 112 countries) successfully solved all six problems, achieving the maximum score of 42 points. This underscores that humans still hold the pinnacle in mathematical problem-solving. Google DeepMind continues to advance AI, striving to create systems capable of solving even more complex mathematical challenges.

ORIENT

Photo: deepmind.google