First-hand information for everyone
This paper describes a focused post‑training process that made language models much stronger at competitive programming. Competitive programming refers to timed algorithm and coding contests such as the International Olympiad in Informatics (IOI) and the International Collegiate Programming Contest (ICPC). These contests are used by researchers as tough tests of problem solving and reasoning for large language models.
The team built an end‑to‑end specialization pipeline. They curated 22,000 programming problems and created synthetic “reasoning traces,” which are step‑by‑step solution sketches used to teach the models how to think through problems. The pipeline then uses supervised fine‑tuning (SFT) — training the model on example input–output pairs — and, for one model, reinforcement learning (RL) — training that rewards correct behavior. Two models were reported: Nemotron‑3‑Nano‑CC (30 billion parameter variant, named 30B‑A3B) trained with SFT and RL, and Nemotron‑3‑Ultra‑CC (550 billion parameter variant, named 550B‑A55B) trained with SFT alone.
They also introduce GenCorrect, a test‑time compute strategy that repeatedly generates solution candidates, runs them, checks which ones pass tests, and refines the code. In plain terms, GenCorrect makes the system try many different answers and fix mistakes before submitting. This gives the models extra problem‑solving power during the contest by using more computation and iterative feedback at test time.
The quantitative results are concrete. On the IOI 2025 problem set, Nano‑CC started at 130 points before post‑training, rose to 291 after the post‑training pipeline, and reached 468 points when GenCorrect was used. The gold medal threshold on that set was 438.3, so Nano‑CC exceeded it with GenCorrect. Ultra‑CC reached 502 on the same set. Building on these results, the authors created a competition‑specific Ultra‑CC system and evaluated it prospectively at IOI 2026. Under the same time limits, access to documentation, and submission rules as human contestants, the system scored 535.4 out of 600. That score exceeds the IOI gold threshold of 361.12 and also exceeds the top human contestant score of 498.27. The authors report this as the first AI system to outscore the highest‑scoring human contestant on an IOI problem set.
The paper includes a limitations section and the authors are careful to present these outcomes in the context of the competition settings they studied. The reported gains come from a large curated problem set, synthetic traces, and substantial compute both during training and at test time with GenCorrect. The paper does not claim that the same approach will automatically solve other kinds of reasoning tasks, and broader generalization is not demonstrated here. Readers should also note that one model used RL while the larger model used only SFT, and that the test‑time strategy relies on extra iterative computation.