Gemini 2.5 Achieves Gold Medal Performance at ICPC World Finals

An advanced version of Gemini 2.5 Deep Think has achieved gold-medal level performance at the 2025 International Collegiate Programming Contest (ICPC) World Finals,, solving 10 out of 12 problems including one no human team could crack.

The ICPC World Finals this year was held in Baku, Azerbaijan on September 4 with top university teams from over 100 countries competing in person. Across five hours competitors faced twelve taxing algorithm problems where correct solutions and speed are both crucial. Only four teams earned gold medals among 139 entrants.

In a remote online setting under the same ICPC rules, DeepMind’s advanced version of Gemini 2.5 Deep Think began 10 minutes after the human teams yet solved 10 out of the 12 problems correctly within the time limit. This performance qualifies as gold-level under the ICPC grading system.

Gemini solved 8 problems in about 45 minutes and then 2 more problems by the 3-hour mark, leveraging complex data structures and advanced algorithms. One problem in particular stood out as no university team managed to solve it during the competition but Gemini did.

Known as Problem C, the challenge involved distributing liquid through a network of ducts and reservoirs where each duct could be open closed or partially open. There are infinitely many possible configurations. Gemini addressed it by assigning priority values to each reservoir, then applying dynamic programming nested search methods, and minimax reasoning to select optimal priority values in a convex search space.

The company credits the result to several innovations including reinforcement learning novel training for multi-step reasoning, parallel solution generation, evaluation and iteration of code before final submission.

DeepMind says this accomplishment builds on Gemini’s earlier gold level performance at the International Mathematical Olympiad (IMO) just two months prior.