Original title: Post-Training Language Models for Gold-Medal Performance in Coding Competitions
Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar +1 · September 2026
Top coding contests are a tough test for AI reasoning. Researchers created a specialized training pipeline to tackle them. They gathered 22,000 coding problems and generated synthetic step-by-step thinking paths. They then trained their models using supervised practice and reinforcement learning, which rewards good answers. At test time, they used GenCorrect, a method where the AI creates many different solutions, tests them, and fixes its mistakes using feedback. This approach worked well. On the IOI 2025 contest, their smaller model beat the gold-medal cutoff, and their larger model scored even higher. Tested live in IOI 2026 under strict human rules, their system scored 535.4 out of 600, surpassing the top human score of 498.27.
Imagine practicing for a spelling bee by studying word lists and breaking down tricky spellings, but during the bee, you face brand-new words and must think on your feet. Now, if you had a system that not only learned from past words but also, during the bee, quickly tried different spellings, checked them against rules, and refined the best attempt, it might beat the champion. In this paper, the AI system is like that: it trains on thousands of coding problems with detailed steps, then during competition, it generates and refines answers in real-time to outperform the top human.
Competitive programming tests advanced logic and problem-solving. When AI beats top humans here, it shows AI can handle complex reasoning tasks that were once uniquely human, potentially changing how we build programming tools or teach computer science.
The results are specific to competitive programming under IOI contest rules, with time and internet access constraints similar to humans. The abstract does not address performance on non-competitive coding tasks or real-world applications outside this setting.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Post-Training Language Models for Gold-Medal Performance in Coding Competitions — Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.