Research2026-09-02

Nvidia researchers trained two Nemotron models to solve competitive programming problems. They used 22,000 curated problems, supervised fine-tuning and reinforcement learning. They also added GenCorrect, a method that generates several solutions and refines them using test feedback. The smaller model rose from 130 points to 291 after training, then to 468 with GenCorrect. That passed the gold medal threshold at IOI 2025. The larger model was then entered live at IOI 2026 under the same time, internet and submission limits as human contestants. It scored 535.4 out of 600, above the gold line and above the best human score. The authors say this is the first AI system to outscore the top human contestant on an IOI problem set.

What changed

No AI system had beaten the highest-scoring human on an IOI problem set.

What it unlocks

A recipe for specialising open models to solve competition-level programming problems.

  • 535.4 of 600 at IOI 2026
  • gold threshold 361.12
  • top human score 498.27
  • 22,000 curated problems

What you need to act on it

  • large-scale training and reinforcement learning infrastructure

Sources

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.