Post-Training Language Models for Gold-Medal Performance in Coding Competitions

Post-Training Language Models for Gold-Medal Performance in Coding Competitions. By Aleksander Ficek and four others, from NVIDIA. Published on the second of September. No code repository is linked to it.

Competitive programming has become a key test of language model reasoning, and the international olympiads are its hardest setting. Those competitions are useful precisely because they are adversarial to memorisation: the problems are new, the time is fixed, and the scoring is external. Reaching the top of them is a specialization problem rather than a general capability one.

The work is an end-to-end specialization pipeline rather than a single technique, combining large-scale problem curation, synthetic reasoning traces, supervised fine-tuning, and reinforcement learning. Alongside it comes a test-time strategy the authors call GenCorrect, which is feedback-driven: it iteratively generates, evaluates and refines a diverse set of solutions rather than committing to one.

Twenty-two thousand curated problems are used to train two models. A smaller one gets both supervised fine-tuning and reinforcement learning; a much larger one gets supervised fine-tuning alone. That asymmetry is itself informative about where each stage is thought to pay off.

The progression on the two thousand and twenty-five olympiad is reported in stages, which makes the contributions separable. The smaller model goes from one hundred and thirty points to two hundred and ninety-one after post-training, then to four hundred and sixty-eight with the test-time strategy, passing a gold threshold of four hundred and thirty-eight point three. The larger model reaches five hundred and two. Then the claim that carries the paper: a competition-specific system was evaluated prospectively during the following year's olympiad, under the same time, internet-access and submission constraints as human contestants, and scored five hundred and thirty-five point four out of six hundred. That exceeds the gold threshold and the top human score of four hundred and ninety-eight point two seven.

Prospective is the word doing the work. Evaluating during the competition, under the contestants' own constraints, is a much stronger claim than scoring a past problem set, and the authors state this is the first system to outscore the highest-scoring human on an olympiad problem set.

Post-Training Language Models for Gold-Medal Performance in Coding Competitions. By Aleksander Ficek and four others, from NVIDIA. Published on the second of September. This is the long version: the whole paper rather than the abstract.

Competitive programming is a good test of reasoning for a specific reason: unlike conventional coding benchmarks, it requires synthesizing novel algorithms, reasoning over complex constraints, and producing implementations that pass hidden tests under time and memory limits. Gold-medal performance at the international olympiads has already been reported by several systems. What this paper adds is not the first gold medal. It is the ability to see which component earned it.

That gap is the stated motivation, and it is a fair complaint. Existing systems are often closed, rely on specialized models, or change several things at once: training data, post-training, model scale and inference-time compute all move together, so the contributions of the components required to reach gold remain difficult to isolate. This pipeline is presented stage by stage precisely so they can be separated. Curation of twenty-two thousand problems, synthetic reasoning traces generated by a frontier model, supervised fine-tuning, reinforcement learning, and iterative test-time refinement. All evaluation problems are excluded and deduplicated from training, which is the hygiene step that any claim like this stands or falls on.

Two models are built from different starting points, and the asymmetry between them is informative. A compact mixture-of-experts model of thirty billion parameters with three billion active gets both supervised fine-tuning and reinforcement learning. A much larger model, five hundred and fifty billion parameters with fifty-five billion active, gets supervised fine-tuning without code-specific reinforcement learning. Both use the test-time strategy the authors call GenCorrect, which iteratively refines a diverse set of solutions using feedback from previous submissions rather than committing to one answer.

The larger model reaches three hundred and four without code-specific reinforcement learning and five hundred and two after five rounds of the same strategy. Then comes the result that distinguishes the paper from a retrospective evaluation. Using the earlier olympiad as a development benchmark to select training and inference adaptations, the authors built a competition-specific system and evaluated it prospectively during the following year's olympiad, live. Under the same time, internet-access and submission constraints as human contestants, it scored five hundred and thirty-five point four out of six hundred, against a gold threshold of three hundred and sixty-one point one two and a top human score of four hundred and ninety-eight point two seven. To the authors' knowledge that is the first system to outscore the highest-scoring human contestant on an olympiad problem set.

Their limitations section is direct about it. The approach requires substantial training and test-time compute, so the live result should be interpreted as a system-level comparison under the same time and submission limits, rather than an equal-resource comparison with human contestants. A human contestant is not running five rounds of fifty submissions from a diverse candidate pool. Compute constraints also prevented reinforcement learning on the larger model and exhaustive ablations across scales and stages, which is a real cost to the paper's own separation goal: the component contributions are isolated for one model and not the other. The authors also note their findings may not generalise beyond competitive programming. And on release, checkpoints and distributable inference components are planned, but the full training corpus cannot be released because of third-party redistribution restrictions, so the curation stage is documented rather than reproducible.

The word carrying the weight is prospective. Scoring above every human on a problem set nobody had seen is a much stronger claim than a leaderboard number, and stating plainly that it is not an equal-resource comparison is what keeps it an honest one.