Gemini AI solves coding problem that stumped 139 human teams at ICPC World Finals (arstechnica.com)

Google says its Gemini 2.5 “Deep Think” model competed in the 2025 ICPC World Finals — connected to an ICPC-approved remote environment with human teams given a 10-minute head start — and solved 10 of 12 algorithmic problems within the five-hour contest to earn a gold medal. Google emphasizes that this was the same general Gemini 2.5 model used in other products (not a freshly trained, contest-specific model); it was only “enhanced” to run extended chain-of-thought or “thinking” tokens over the full five hours to explore solution strategies and produce verified code. Only four of 139 human teams matched that 10/12 result. The result is significant because it showcases a generative model’s ability to perform sustained algorithmic reasoning, plan over long horizons, and generate correct competitive programming solutions without task-specific fine-tuning — capabilities relevant to software engineering, automated theorem proving, and complex decision tasks. Practically, it raises questions about evaluation and contest design (fairness, reproducibility, compute budgets, and what counts as human vs. tool achievement) and strengthens Google’s argument that large models are approaching more general problem-solving competencies. ICPC leadership called the performance a milestone for defining future academic and AI standards, while the ML community will watch for technical details on inference-time compute, verification methods, and how the model handled tricky edge cases.