2025-10-24 · Google

Gemini achieves gold-medal level at the International Collegiate Programming Contest World Finals

agentsmodels

read at source ↗ deepmind.google

Gemini achieves gold-medal level at the International Collegiate Programming Contest World Finals

Source: DeepMind Date: 2025-10-24 URL: https://deepmind.google/blog/gemini-achieves-gold-medal-level-at-the-international-collegiate-programming-contest-world-finals/

Summary

An advanced Gemini 2.5 Deep Think variant solved 10 of 12 ICPC World Finals 2025 problems, completing 8 in 45 minutes and all 10 within 3 hours. Would rank 2nd overall among participating university teams. Notably solved Problem C — a resource distribution optimization challenge — which no human team solved during the competition. The system used parallel agent approaches, iterative testing, and multi-step verification.

Implications

Solving a problem no human team solved is the signal. ICPC gold-level performance is impressive, but solving an unsolved-by-humans problem in a live competition moves this from “matches human experts” to “exceeds human experts in specific problem classes.” That’s a different capability claim and deserves verification.

ICPC complements IMO gold. Math olympiad (theorem proving) + ICPC (algorithmic programming) are the two canonical hard computation benchmarks. Gemini claiming gold in both within the same calendar year is a systematic statement about reasoning and algorithmic capability, not cherry-picked performance.

Parallel agents + iterative testing is the methodology. The approach described — multiple parallel agents exploring solutions, iterative testing, multi-step verification — is the same multi-agent architecture that’s increasingly used in coding tools like Devin. This is not a single-pass completion; it’s agentic problem-solving applied to competitive programming.

Watch:

  • Independent verification that the ICPC result holds under the same conditions human teams face (no hidden test case advantage)
  • Whether the parallel-agent approach is made available in Gemini’s coding API or remains internal to benchmark systems
  • OpenAI o3-series ICPC performance — this is the direct competitive comparison to make

← all signals