Make sure you check out our new project - extraordinary.technology!
Post-Training Language Models for Gold-Medal Performance in Coding Competitions
Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar, Boris Ginsburg · Read the paper
Hook: The reported results suggest that post-training and iterative test-time solution refinement can substantially improve language-model performance on difficult coding competitions.
Summary: The paper presents a post-training pipeline for competitive-programming language models that combines curated problems, synthetic reasoning traces, supervised fine-tuning, and reinforcement learning. The authors report that their systems exceeded the gold threshold at IOI 2025, while a competition-specific system scored 535.4 out of 600 at IOI 2026, above the highest human score reported for that contest.
Industry impact: For organizations evaluating coding models, the paper highlights problem curation, feedback-driven generation, and post-training as important components beyond base-model scale. Its results also provide a benchmark for systems operating under the same time, internet-access, and submission constraints as human contestants.
Potential implications: The findings indicate that competitive-programming performance may depend strongly on how models generate, test, and refine multiple candidate solutions, not only on their initial capabilities. However, the reported conclusions are specific to the described IOI evaluations and do not by themselves establish broader software-engineering performance.
