Gemini achieves gold-medal level at the Programming Contest World Finals (deepmind.google)

An advanced Gemini 2.5 Deep Think model achieved gold‑medal level performance at the 2025 ICPC World Finals, solving 10 of 12 problems under the same five‑hour constraint (it started 10 minutes after human teams). Gemini produced eight correct solutions within 45 minutes and two more within three hours; its combined penalty time (677 minutes) would place it 2nd among the 139 competing university teams. The ICPC — the premier global collegiate algorithmic contest — enforces exact correctness and strict time penalties, making this a high‑bar demonstration of automated competitive programming. Technically, the win builds on innovations across pretraining, post‑training, novel reinforcement learning, multi‑step reasoning and parallel “multi‑agent” solution search where agents propose, execute and iteratively improve code via terminals. A standout example: Problem C, unsolved by any human team, was tackled by reducing an infinite‑configuration flow optimization to a parameterized dynamic programming problem via reservoir “priority” values; Gemini applied the minimax theorem and nested ternary searches in a convex search space to find optimal priorities efficiently. Google notes this milestone as progress toward richer AI problem‑solving and collaboration (a lightweight Deep Think is available to Google AI Ultra users), while ICPC has accepted the submitted solutions but did not validate the underlying system or processes.