Gemini 2.5 Deep Think ICPC World Finals Performance

Gemini 2.5 Deep Think achieves gold-medal level at the 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. This milestone demonstrates a significant leap in abstract problem-solving and follows a gold-medal standard achievement at the International Mathematical Olympiad (IMO) two months prior.

Performance Metrics and Ranking

Gemini 2.5 Deep Think solved 10 out of 12 problems during the competition, operating under the same five-hour time constraint as human contestants (starting 10 minutes after the start).

Key performance statistics include:

Problem Solving Rate: 10/12 problems solved correctly.
Speed: Eight problems were solved within the first 45 minutes, and two additional problems were solved within three hours.
Total Time: The model used a combined total of 677 minutes to solve the 10 problems.
Comparative Rank: Based on these results, Gemini 2.5 Deep Think would have ranked 2nd place overall among the university teams in the competition.

Solving Unsolved Problems: The Case of Problem C

Gemini 2.5 Deep Think successfully solved Problem C, a task that no university team in the contest was able to solve. Problem C involved optimizing the distribution of liquid through a network of interconnected ducts to reservoirs to find the fastest filling configuration.

To solve this, the model employed a sophisticated reasoning chain:

Priority Value Assumption: It assumed each reservoir had a "priority value" representing its relative favorability.
Dynamic Programming: It used a dynamic programming algorithm to find the best duct configuration for a given set of priority values.
Minimax Theorem: It applied the minimax theorem to find priority values that made the resulting flow most constrained.
Nested Ternary Searches: It utilized nested ternary searches to navigate the bowl-like convex solution space and identify the optimal priority values.

Technical Foundations of the Breakthrough

The model's performance is the result of integrated advances in pretraining, post-training, novel reinforcement learning (RL) techniques, multi-step reasoning, and parallel thinking.

To enhance its capabilities, Google DeepMind trained Gemini to reason and generate code for the most difficult programming problems, allowing the model to learn from feedback on results and evolve its approach. The system utilizes multiple Gemini agents that propose solutions, execute code and tests via terminals, and iterate on those solutions based on the results of all attempts.

Internal studies indicate that this version of Gemini 2.5 Deep Think would also achieve gold-medal level performance in the 2023 and 2024 ICPC World Finals, performing on par with the world's top 20 competitive coders.

Implications for Software Engineering and AGI

This achievement indicates that AI can serve as a high-level problem-solving partner for programmers. Google DeepMind notes that if the best AI and human solutions from the competition were combined, all 12 problems would have been solved correctly, suggesting a complementary relationship between AI and human expertise.

Beyond coding, the ability to understand complex problems, devise multi-step logical plans, and implement them flawlessly is applicable to other scientific and engineering fields, such as microchip design and drug discovery. This progress marks a step toward artificial general intelligence (AGI) by moving AI from information processing to solving complex reasoning problems.

Gemini successfully joining this arena, and achieving gold-level results, marks a key moment in defining the AI tools and academic standards needed for the next generation.