Gemini 2.5 Deep Think Wins Gold at ICPC 2025

Key Takeaways

1. The ICPC is a prestigious competition for university-level algorithmic programming, with the 2025 World Championship held in Baku featuring top teams from over 100 nations.
2. Gemini 2.5 Deep Think, an AI developed by Google, performed exceptionally well by solving ten out of twelve problems in the competition, ranking second among human teams.
3. Problem C, which no human team solved, was tackled successfully by Gemini using advanced techniques like dynamic programming and nested ternary searches.
4. The performance of Gemini reflects advancements in AI, including pretraining, innovative reinforcement learning, and multi-step reasoning capabilities.
5. The collaboration between AI and human teams can enhance problem-solving in various fields, offering new opportunities in software development and complex engineering tasks.

The International Collegiate Programming Contest (ICPC) is widely recognized as the earliest and most esteemed competition for university-level algorithmic programming. Google states that the top teams from more than 100 nations took part in the 2025 World Championship held in Baku, Azerbaijan. Participants were required to tackle intricate algorithmic challenges within a five-hour timeframe. Only those solutions that were error-free were awarded points. Among 139 competing teams, merely four achieved gold medals.

Gemini’s Impressive Performance

As per Google’s information, Gemini 2.5 Deep Think was activated just ten minutes after the human teams and managed to solve eight problems in 45 minutes, with two additional problems resolved within three hours. In total, it correctly answered ten out of twelve problems, all completed in 677 minutes, which would have ranked Gemini second among the university teams. Dr. Bill Poucher, the ICPC Global Executive Director, remarked:

“The successful entry of Gemini into this field, achieving gold-level results, signifies an important moment in delineating the AI tools and academic standards essential for the upcoming generation. Kudos to Google DeepMind; this effort will assist us in igniting a digital renaissance for everyone’s benefit.”

Unique Challenges Faced

Particularly significant was Problem C, which remained unsolved by any human team. The challenge involved setting up a network of pipes to fill several reservoirs as swiftly as possible. To tackle this, Gemini utilized dynamic programming, the minimax strategy, and nested ternary searches to determine priority values for optimal flow, successfully resolving the issue.

According to Google, the performance of Gemini hinges on breakthroughs in pretraining, post-training, innovative reinforcement learning methods, multi-step reasoning, and parallel thought processes. Various AI agents generate their own solutions, test them, run code, and refine their approach until the best answer is found. Google’s internal reviews suggest that Gemini would have also reached gold level in the 2023 and 2024 ICPC World Finals.

Collaboration Between AI and Humans

These outcomes highlight that AI can serve as a collaborative problem-solving partner for developers. Merging AI with human teams could facilitate the tackling of extensive and intricate tasks. This presents substantial opportunities in software development, scientific inquiry, and complex engineering projects. Google AI Ultra subscribers can already access a simplified version of Gemini 2.5 Deep Think through the Gemini app.

Source:
Link