Post-Training Language Models for Gold-Medal Performance in Coding Competitions
Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar, Boris Ginsburg
cs.LG · cs.AI · cs.CL · cs.MA · cs.SE
Quick verdict
NVIDIA's SFT+RL+GenCorrect pipeline reaches IOI gold; live unsupervised IOI 2026 run scores 535.4/600, beating top human — credible but rests on one run.
Summary
An NVIDIA team post-trains two Nemotron models for competitive programming: Nemotron-3-Nano-CC (30B-A3B), refined with SFT on 1.2M DeepSeek-V4-Flash reasoning traces plus GRPO-based RL, and Nemotron-3-Ultra-CC (550B-A55B), trained with SFT alone (477,642 traces). Traces derive from 22,000 curated competition problems with evaluation-set exclusion and deduplication. GenCorrect, a five-round test-time strategy that samples 200 solutions per round, submits 10 diversity-selected candidates, and conditions later rounds on subtask feedback, lifts Nano-CC from 291 to 468 on IOI 2025 and Ultra-CC to 502, above the 438.3 gold threshold. A competition-specific Ultra-CC variant (GLM-5.2 SFT data, NVFP4 quantization) scored 535.4/600 in an unsupervised live run during IOI 2026, exceeding gold threshold 361.12 and the top human score of 498.27; the authors claim this is the first AI system to outscore the highest-scoring IOI contestant. Limitations: the live run was not IOI-supervised and rests on a single attempt whose five-run post-competition range (495.0–545.8) straddles the human margin; no confidence intervals are reported; the full training corpus cannot be released.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- SFT on curated synthetic reasoning traces provides the largest single-sample gains (Nano-CC IOI 2025 Score@1 21.7%→46.7%), with GRPO RL adding smaller gains (→48.5%).
- GenCorrect raises Nano-CC to 468 and Ultra-CC to 502/600 on IOI 2025 within the official 50-submission limit, exceeding the gold threshold of 438.3.
- A competition-specific Ultra-CC (GLM-5.2 SFT data, NVFP4 quantization, expanded final round) scored 535.4/600 in a live IOI 2026 run, above gold threshold 361.12 and the top human score of 498.27 — claimed as the first AI to outscore the highest-scoring IOI contestant.
- When model size and inference cost are not constraints, adapting a stronger base model with limited SFT outperforms extensive post-training of a smaller model.
Strengths
Prospective live evaluation at IOI 2026
The competition-specific system was run during the official contest before problems were public (§5.1), eliminating retrospective contamination concerns that affect nearly all other IOI/ICPC evaluations cited in §6.
Explicit contamination control
All IOI 2025, ICPC 2025, and LiveCodeBench Pro problems are excluded and deduplicated from SFT/RL data (§3.1, Appendix A), with environment validation via reference solutions and gpt-oss-120b-generated candidates.
Unusually complete ablation suite
Honest scoping of the headline claim
Full method transparency
Weaknesses
Headline claim rests on n=1 live run with comparable variance
No uncertainty quantification anywhere
Live-model configuration complicates attribution
GenCorrect selection procedure not ablated
Partial reproducibility
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review: Post-Training Language Models for Gold-Medal Performance in Coding Competitions
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Did an AI beat the best human at IOI 2026?
According to this NVIDIA paper, its competition-specific Nemotron-3-Ultra-CC system scored 535.4 out of 600 in a live run during IOI 2026, above the gold-medal threshold (361.12) and the top official human score (498.27). The run was unofficial and not supervised by IOI, so it does not appear in official rankings.
What is GenCorrect?
A five-round test-time compute strategy for competitive programming: each round generates up to 200 candidate solutions, selects 10 diverse representatives via token-shingle clustering, submits them for subtask-level scores, and conditions the next round on accumulated feedback — all within the official 50-submission limit per problem.
Which models were trained and how large are they?
Nemotron-3-Nano-CC (30B total, 3B active parameters; SFT plus GRPO reinforcement learning) and Nemotron-3-Ultra-CC (550B total, 55B active; SFT only), trained on 22,000 curated competition problems using teacher traces from DeepSeek-V4-Flash (and GLM-5.2 for the competition-specific variant).
Is the code or model available?
The authors plan to release the competition checkpoint together with inference and evaluation recipes in NeMo-Skills; the full training corpus cannot be released due to third-party redistribution restrictions.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- No truncation detected: the extracted text runs through the Appendix C.1 prompt templates and appears complete.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.