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

2609.02849 ↗ · 2026-09-02 · cs.LG, cs.AI, cs.CL, cs.MA, cs.SE

Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar, Boris Ginsburg

The finding

Headline claim

An NVIDIA post-training pipeline (SFT, RL, and an iterative test-time refinement method called GenCorrect) lets a 30B-A3B model exceed the IOI 2025 gold threshold and a 550B-A55B SFT-only system score 535.4/600 on IOI 2026, above the top human contestant's 498.27, claimed as the first AI system to outscore the best human on an IOI problem set.

The test

Evaluation on IOI 2025, ICPC 2025 World Finals, and LiveCodeBench Pro, with evaluation problems excluded and deduplicated from training data (22,000 curated problems, 1.2M synthetic traces for Nano, 477,642 for Ultra). Final results averaged over 1,000 runs; intermediate checkpoints over 50 runs; IOI Score@200 over five runs; LCB Pro over eight runs. The IOI 2026 run was prospective, under human time, internet, and 50-submission constraints, but unofficial and unsupervised by IOI.

Effect size & scope

Nano-CC IOI 2025 Score@1: 130 to 291 after SFT+RL (21.7% to 48.5%), 468 with five GenCorrect rounds versus gold threshold of 438.3. Ultra-CC reaches 502 on IOI 2025 and 535.4/600 on IOI 2026 versus gold threshold 361.12 and top human score 498.27. SFT alone drives most gains (Nano IOI 21.7% to 47.3%; ICPC 16.9% to 46.7%; LCB Pro 17.6% to 70.7%); RL adds only 1 to 4 points; GenCorrect adds 107.6 (Nano) to 158.1 (Ultra) points. Nano-CC at 48.5% IOI still trails GLM-5.2 (66.0%), DeepSeek-V4-Pro (56.8%), and DeepSeek-V4-Flash (55.3%) at Score@1.

Limitations

The authors concede the IOI 2026 run was not an official contestant and was unsupervised and unranked by IOI. RL gains are modest, attributed to sparse terminal rewards and long-horizon credit assignment. Ultra's SFT uses a teacher-generated dataset from DeepSeek-V4-Flash, raising dependency questions. Unstated: competition-specific adaptations for IOI 2026 (expanded final-round selection, NVFP4 quantization, GLM-5.2-based Ultra SFT) mean the live system differed from the general pipeline; contamination risk from an online-problem corpus is reduced but not provably zero; compute cost of 200 candidate generations per round and 1,000-run averaging is not reported; results are leaderboard-shaped, with no measurement of real-world software engineering transfer.

Counterpoint

A skeptic would note this is competitive-programming benchmark chasing with heavy test-time compute (200 samples, five feedback rounds, 50 submissions), far from single-shot capability, and that scoring 535.4 in an unofficial unsupervised run against human students, on the authors' own harness, is partly a PR result dressed as science. The marginal contribution over prior gold-medal systems (OpenAI, DeepSeek) is a pipeline replication, and Nano-CC's Score@1 is well behind frontier models. Prompt and harness sensitivity, and the use of a different teacher model for the IOI 2026 system, complicate clean attribution of which component produced the headline result.

Economy implication

Little or nothing yet for enterprise economics in the next 12 to 24 months: the immediate signal is that verifiable-reward domains with executable feedback are now cheaply saturatable by small-model post-training plus test-time compute, which modestly supports falling cost-per-capability in coding, but no enterprise ROI, revenue, or labor data is presented.

Who this matters for

Technical Capability: Shows that a 3B-active-parameter model can reach gold-medal competitive-programming performance through data curation, SFT, and feedback-driven test-time compute, evidence that capability at a fixed benchmark is increasingly a training-recipe problem rather than a frontier-scale one.

Labor & Knowledge Work: Elite algorithmic problem-solving, long a credential signaling top-tier software talent, is now above the best human level, which bears on how much weight olympiad-style coding skill should carry in hiring and evaluation.

Bears on (advisory · papers never write evidence)

[1.1] Frontier benchmark scores kept rising gen-over-gen through 2025–26.

[1.4] Inference-time compute yields gains that don't saturate quickly on hard reasoning.

Abstract

Competitive programming has become a key test of large language model reasoning, with international competitions such as IOI and ICPC representing its most challenging settings. We present an end-to-end specialization pipeline combining large-scale problem curation, synthetic reasoning traces, supervised fine-tuning (SFT), and reinforcement learning (RL). Using 22,000 curated problems, we train Nemotron-3-Nano-CC (30B-A3B) with SFT and RL and Nemotron-3-Ultra-CC (550B-A55B) with SFT alone. We further introduce GenCorrect, a feedback-driven test-time compute strategy that iteratively generates, evaluates, and refines diverse solutions. On IOI 2025, Nano-CC improves from 130 points to 291 after post-training and to 468 with GenCorrect, exceeding the gold threshold of 438.3 while Ultra-CC reaches 502. Guided by these results, we develop a competition-specific Ultra-CC system and evaluate it prospectively during IOI 2026. Under the same time, internet-access, and submission constraints as human contestants, it scores 535.4 out of 600, exceeding both the gold threshold of 361.12 and the top human score of 498.27. To our knowledge, this is the first AI system to outscore the highest-scoring human contestant on an IOI problem set.