Google DeepMind said on the 17th (local time) that its artificial intelligence (AI) model "Gemini 2.5 DeepSeek" delivered a gold medal–level performance at the 2025 International Collegiate Programming Contest (ICPC), the world's most prestigious computer programming competition. It came just two months after the model showed gold medal–level performance at the International Mathematical Olympiad (IMO).
"Gemini Deep Seek" is a research AI model specialized for high-level logic and computation tasks such as mathematics, algorithms, and abstract problem-solving.
The ICPC is the largest collegiate algorithm programming competition in the world. It is a step above the high school–level International Mathematical Olympiad (IMO).
At this year's ICPC, held on 4th in Baku, Azerbaijan, 139 teams selected from more than 3,000 universities in 103 countries around the world competed.
Teams must solve difficult algorithm problems over five hours, only perfectly correct answers receive scores, and submission times are factored into the rankings. Only the top four teams earn gold medals.
In this competition, the advanced version of "Gemini 2.5 Deep Seek" participated remotely online and solved 10 out of 12 problems, achieving a gold medal–level result equivalent to second place overall. The model solved eight problems within 45 minutes and another two within an additional three hours.
It drew attention especially for solving one problem that none of the collegiate teams could solve. The problem required optimizing a complex liquid flow distribution and was a hard challenge that had to consider countless combinations of pipelines. Gemini introduced an original concept called "reservoir priority value," simplified the problem via dynamic programming, and combined it with a ternary search technique to find the optimal solution.
Google DeepMind said, "This outcome is the result of combining various innovations, including pretraining, post-training, a new reinforcement learning method, multi-step reasoning, and parallel thinking," adding, "Gemini solved problems by having multiple agents each generate, execute, and verify code, and continuously improve."
It added, "This proves that AI can function as a true problem-solving partner," and "if the competition had combined AI and human solutions, all 12 problems could have been fully solved."
Google DeepMind said this result suggests that AI can contribute to abstract problem-solving beyond programming, including complex scientific and engineering fields such as new drug development and semiconductor design, and that it has entered a phase where it can tangibly help solve humanity's grand challenges.