Google's latest version of Gemini 2.5 Deep Think made history in the 2025 International Collegiate Programming Contest (ICPC) World Finals, not only achieving a gold medal but also solving a problem that no human team could crack, marking a major breakthrough in artificial intelligence's ability to tackle complex problems.
Competition Performance Exceeding Human Players
The ICPC World Finals, held on September 4th in Baku, Azerbaijan, brought together elite competitors from over 3,000 universities across 100 countries. A total of 139 teams qualified for the competition. In this top-tier event, participants were required to solve 12 algorithmic problems within 5 hours, and Gemini Deep Think demonstrated astonishing strength.
The AI system participated online according to official rules, starting 10 minutes after the human teams. Notably, it solved 8 questions within 45 minutes and completed another 2 within 3 hours, totaling 677 minutes to solve 10 questions, an achievement that would rank second among human competitors.
The most remarkable achievement was Gemini Deep Think successfully solving Problem C — a complex question about maximizing the efficiency of a liquid distribution system using configurable pipelines and reservoirs. No human team could conquer this challenge. This task required the system to find the fastest filling strategy among almost infinite possibilities, showcasing a breakthrough in AI's abstract reasoning capabilities.
Behind the Technical Breakthrough
According to Google DeepMind, this success resulted from comprehensive advancements in pre-training, post-training, advanced reinforcement learning methods, multi-step logical reasoning, and parallel problem-solving. During the reinforcement learning process, the model was trained on extremely difficult programming tasks, with multiple Gemini agents generating different candidate solutions, testing them in a virtual terminal, and continuously iterating and improving.
This version was directly built upon Gemini 2.5 Deep Think, which recently won a gold medal at the International Mathematical Olympiad (IMO), demonstrating Google's rapid iteration and improvement in AI reasoning capabilities.
Industrial Significance and Future Prospects
Dr. Bill Poucher, Executive Director of ICPC, called this a turning point: "Gemini's successful entry into this field and achieving gold-level results marks a critical moment in defining the next generation of artificial intelligence tools and academic standards." He emphasized that the ability to break down complex problems, design multi-step strategies, and implement them correctly is not only important in programming, but also in drug discovery, chip design, and broader research areas.
Google DeepMind stated that Gemini can serve as a partner for developers. Theoretically, combining the best solutions from humans and AI could solve all 12 competition problems. Currently, the lightweight version of Gemini 2.5 Deep Think is available through the Gemini app to users of the Google AI Ultra program. Future versions are expected to develop into more powerful programming assistants, applied in software development, logistics, and scientific research fields.
The White-Hot AI Olympic Competition
This breakthrough followed a major success by OpenAI. In August 2025, the OpenAI system won a gold medal at the International Olympiad in Informatics (IOI), outperforming only 5 of the 330 human participants. Similar to Gemini, this result also came from a general reasoning model rather than a highly customized system, showing the rapid development of AI reasoning capabilities.
