Source-linked AI summary
Post-Training Language Models for Gold-Medal Performance in Coding Competitions
Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar, Boris Ginsburg
TL;DR
Competitive programming tests whether language models can reason, synthesize algorithms, and produce implementations under strict constraints. This paper combines data curation, post-training, and feedback-driven inference, culminating in a live IOI 2026 score that exceeded the highest human score.
Problem
Competitive programming provides a challenging evaluation of language-model reasoning because solutions must handle complex constraints and pass hidden tests under time and memory limits.
Method
The paper develops an end-to-end pipeline using curated problems, synthetic reasoning data, SFT, RL, and GenCorrect for feedback-driven solution refinement.
Results
535.4 out of 600 on IOI 2026 exceeded the top human contestant by 37.1 points in a live prospective run.
Takeaways & Limitations
The results report gold-medal performance and an AI system exceeding the highest-scoring human contestant on an IOI problem set.
Takeaways & Limitations
The live result is a system-level comparison rather than an equal-resource comparison with humans, and the findings may not generalize beyond competitive programming.
Abstract
from arXiv · showhide
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.
1. Introduction
The paper presents an end-to-end competitive-programming specialization pipeline and evaluates its progression from post-training through feedback-driven inference on IOI benchmarks. The resulting systems reach gold-level performance on IOI 2025 and exceed the highest human score in an unofficial prospective IOI 2026 evaluation.
- Pipeline: The pipeline combines problem curation, synthetic reasoning data, SFT, RL, and iterative test-time refinement for competitive programming.It uses 22,000 curated problems and trains Nano-CC with SFT and RL, while Ultra-CC uses SFT alone.
- Pipeline: GenCorrect iteratively generates diverse solutions, evaluates submissions, and refines later solutions using feedback under a constrained submission budget.Both models use GenCorrect at inference time.
- IOI 2025 results: 468 points: Nano-CC rises from 130 to 468 on IOI 2025 after post-training and five GenCorrect rounds, exceeding the 438.3 gold threshold.Its intermediate scores are 280 after SFT and 291 after RL.
- IOI 2025 results: 502.0: Ultra-CC reaches this IOI 2025 score after five GenCorrect rounds, compared with 304 before iterative refinement.The model uses supervised fine-tuning without code-specific RL.
- IOI 2026 results: 535.4/600: the competition-specific Ultra-CC system exceeds both the IOI 2026 gold threshold of 361.12 and the top human score of 498.27.The evaluation used matched competition time, internet-access, and submission constraints, but was unofficial and unsupervised.
2. Competition Settings
IOI and ICPC are demanding competitive-programming settings with different scoring, participation, and submission structures. IOI awards partial credit by subtask, whereas ICPC uses binary problem-level scoring and team-based ranking.
- International Olympiad in Informatics: IOI contestants compete individually over two contest days, solving six problems worth up to 100 points each.The maximum combined score is 600.
- International Olympiad in Informatics: IOI submissions earn partial credit for subtasks whose hidden tests pass, allowing partially correct or less efficient algorithms to score.Contestants may submit at most 50 solutions per problem.
- International Olympiad in Informatics: IOI medal thresholds are based on the final score distribution, with approximately the top 1/12 receiving gold.The next 1/6 receive silver and the next 1/4 receive bronze.
- International Collegiate Programming Contest: ICPC World Finals teams of three compete for five hours using one shared computer across 12 algorithmic problems.The description specifies team-based participation and a single shared computing environment.
- International Collegiate Programming Contest: ICPC uses binary scoring: a problem counts only when one submission passes all hidden tests, with rankings primarily determined by solved-problem count and penalty time.Penalty time includes elapsed time to acceptance plus 20 minutes for each preceding incorrect submission.
3. Method
The pipeline combines curated executable problem data, synthetic reasoning traces, SFT, RL, and iterative test-time refinement. GenCorrect generates diverse candidates, selects representatives for evaluation, and uses subtask feedback to guide later rounds under submission constraints.
- 3.1. Data Curation: 22,000 problems are curated from 16 regional and international competition families spanning two decades, alongside online-platform problems.Each problem is packaged into an executable environment with statements, constraints, test cases, auxiliary files, and reference solutions.
- 3.1. Data Curation: Evaluation problems are excluded from training, and the corpus is deduplicated against IOI 2025, ICPC 2025, and LiveCodeBench Pro.IOI 2026 is evaluated prospectively before its problems are publicly released.
- 3.2. Supervised Fine-Tuning: Synthetic SFT data comprises 1.2 million reasoning traces for Nano and 477,642 traces for Ultra, including teacher self-improvement traces.The traces expose models to iterative refinement behavior used by GenCorrect.
- 3.3. Reinforcement Learning: RL is applied only to Nano using 3,219 reliable, sufficiently fast executable problems split into 2,847 training and 372 validation problems.Generated C++17 solutions are compiled and executed for executable rewards.
- 3.4. Test-Time Compute: GenCorrect runs for up to five rounds, generating candidate solutions, selecting diverse representatives, executing them, and refining later generations with evaluator feedback.For IOI, five rounds of 10 submissions match the official 50-submission limit per problem.
- 3.4. Test-Time Compute: Diversity selection chooses up to 10 centers, assigns candidates to clusters, and submits the highest-scoring representative from each cluster.Filtering and selection occur before current-round evaluator scores are observed.
- 3.4. Test-Time Compute: Each subsequent round is conditioned on accumulated per-subtask scores and three references that preserve solved subtasks, target remaining gaps, and maintain diversity.The first round uses only the problem statement; later rounds also use prior solutions and evaluator feedback.
4. Experiments
Experiments evaluate the specialized models across IOI 2025, ICPC 2025, and LiveCodeBench Pro, measuring single-sample, parallel-sampling, reinforcement-learning, and iterative-refinement performance. SFT and RL improve Nano-CC across benchmarks, while Ultra-CC achieves the strongest single-sample results and GenCorrect substantially increases scores.
- Main Results: 51.0% ICPC 2025 Pass@1 and 71.6% LCB Pro Pass@1 are achieved by Nano-CC, versus 16.9% and 17.6% for its base model.The gains transfer beyond IOI, and Nano-CC also exceeds several evaluated baselines on LCB Pro.
- Main Results: 50.7% IOI, 57.4% ICPC, and 74.5% LCB Pro are achieved by Ultra-CC after SFT, improving from 45.5%, 54.0%, and 72.6%.Ultra-CC uses SFT without the CC RL stage and achieves the strongest results among the authors’ models on all three benchmarks.
- Reinforcement Learning: 1.8 percentage points is the IOI gain from RL after third-epoch SFT, while ICPC and LCB Pro increase by 3.7 and 0.9 points.RL raises Nano-CC from 46.7% to 48.5% on IOI, 47.3% to 51.0% on ICPC, and 70.7% to 71.6% on LCB Pro.
- Test-Time Compute: 107.6 points is Nano-CC’s five-round GenCorrect gain on IOI 2025, increasing its mean score from 360.6 to 468.2.GenCorrect uses evaluator feedback to refine solutions across rounds while concentrating 200 generations into 10 submissions per round.
- Test-Time Compute: 158.1 points is Ultra-CC’s five-round GenCorrect gain on IOI 2025, increasing its mean score from 343.9 to 502.0.Ultra-CC’s larger gain follows its stronger performance under parallel sampling, where it reaches 505 compared with Nano-CC’s 461.
- Test-Time Compute: 9.6 problems solved is Ultra-CC’s ICPC 2025 result after two GenCorrect rounds, maintained through round five.Nano-CC increases from 8.6 to 9.4 solved problems and begins to plateau after the third round.
5. IOI 2026
The competition-specific Ultra-CC system was evaluated live on IOI 2026 under human-matched constraints, combining adapted training, expanded final-round search, and execution-based selection. It scored 535.4/600, above both the gold threshold and the top human score, while general GenCorrect independently averaged 521.72.
- 5. IOI 2026: IOI 2025 served as the development benchmark for selecting competition-specific training and inference adaptations for the live IOI 2026 system.The system was designed to maximize one live run within official time and submission limits.
- 5. IOI 2026: Ultra fine-tuning used SFT rather than RL because SFT provided most post-training gains, while RL at Ultra scale exceeded the available compute budget.The live system used GLM-5.2 data for Ultra SFT.
- 5. IOI 2026: 1,000 final-round solutions were generated instead of the standard 200, while only 10 candidates were submitted per problem.The expanded pool was motivated by limited expected benefit from repeating the standard fifth-round procedure.
- 5. IOI 2026: Execution-based selection ranked candidates using generated test inputs and a scoring script before submitting the 10 highest-ranked solutions.The procedure generated and validated 100 test inputs, executed every compiled candidate, and ranked them by estimated subtask performance.
- 5. IOI 2026: 535.4/600 was scored by Ultra-CC in a single prospective IOI 2026 run under the same time, internet-access, and submission constraints as human contestants.The live result exceeded the gold-medal threshold by 174.3 points and the top human contestant by 37.1 points.
- 5. IOI 2026: 521.72 was the mean score of five independent postcompetition runs using standard GenCorrect, with an observed range of 495.0–545.8.The live result was 13.68 points above this mean and remained within the observed range.
6. Related Work
Prior work established competitive programming as a demanding language-model evaluation setting and progressively combined specialized training, sampling, refinement, and reranking. This paper builds on synthetic reasoning data, execution-based reinforcement learning, and test-time scaling approaches.
- 6. Related Work: AlphaCode established domain-specific training with large-scale sampling, filtering, and behavioral clustering for competitive programming.AlphaCode 2 extended the approach with stronger models, additional fine-tuning, and learned reranking.
- 6. Related Work: LiveCodeBench and LiveCodeBench Pro introduced temporally separated and olympiad-level evaluations for measuring coding ability.
- 6. Related Work: Recent systems reported medal-level performance on IOI and ICPC, including specialized reinforcement learning and hand-engineered inference pipelines.The cited work includes o1-ioi, o3, and systems reporting gold-medal performance at ICPC 2025.
- 6. Related Work: Synthetic reasoning traces, self-critique, iterative refinement, and execution-grounded selection have become complementary directions for competitive-programming code generation.The paper positions its approach alongside OpenCodeReasoning, OpenCodeReasoning-II, test-based refinement, and hybrid inference scaling.
7. Conclusion
The paper concludes that combining synthetic data, SFT, RL, and GenCorrect improves competitive-programming performance, and that competition-specific adaptations produced a record IOI result. The live system exceeded both the gold threshold and the highest human score.
- 7. Conclusion: SFT provides the largest single-sample gains, RL adds smaller improvements, and GenCorrect substantially improves performance through feedback-driven refinement.
- 7. Conclusion: 468 points was reached by Nano-CC on IOI 2025 after post-training and GenCorrect, exceeding the gold threshold with 3B active parameters.
- 7. Conclusion: 502 points was reached by Ultra-CC on IOI 2025, while the competition-specific system scored 535.4/600 on IOI 2026.
- 7. Conclusion: The IOI 2026 result exceeded both the gold threshold and the highest human score during a live run under the same time, submission, and internet restrictions.The authors identify this as the first instance of an AI system outscoring the highest-scoring human contestant on an IOI problem set.
Limitations
The approach requires substantial training and test-time compute, limiting the comparability and breadth of the reported evaluation. The live result is therefore framed as a system-level comparison rather than an equal-resource comparison with human contestants.
- Limitations: Substantial training and test-time compute are required, so the live IOI result is a system-level rather than equal-resource comparison with human contestants.Compute constraints also prevented Ultra-CC RL training and exhaustive ablations across model scales and training stages.
- Limitations: The findings may not generalize beyond competitive programming.The paper also cannot release the full training corpus because of third-party redistribution restrictions.
A. Data Curation Details
The paper builds a standardized, validated corpus of 22,000 competition problems and removes unreliable or contaminated evaluations before training.
- Problem collection: 22,000 problems from 16 competition families and online platforms are packaged with statements, limits, tests, auxiliary files, and reference solutions.An automated pipeline constructs standardized executable evaluation environments across two decades of competitions.
- Environment validation: Known correct solutions must receive expected scores and known incorrect solutions must fail during environment validation.Additional generated-solution checks detect malformed statements, missing files, compiler incompatibilities, and inconsistent test harnesses.
- Filtering: Problems with inconsistent verdicts, failing reference solutions, or noncompiling provided solutions are removed, along with RL problems requiring over 300 seconds sequentially.These filters target evaluation reliability and practical execution constraints.
- Evaluation contamination: IOI 2025, ICPC 2025, and LiveCodeBench Pro problems are excluded and deduplicated against both SFT and RL corpora.IOI 2026 required no retrospective filtering because evaluation occurred before public release.
B. Training Details
Training combines synthetic reasoning traces, supervised fine-tuning, and reinforcement learning for Nano-CC, while GenCorrect-related behaviors and diverse candidate selection are incorporated into the training and inference pipeline.
- Synthetic data: 1.2 million traces train Nano-CC and 477,642 traces train Ultra-CC, with more generations allocated to harder problems.The traces are generated by DeepSeek-V4-Flash and the training problems are categorized as easy, medium, or hard.
- Supervised fine-tuning: Self-improvement traces expose student models to iterative solution refinement used by GenCorrect at inference time.The teacher improves a previously generated solution conditioned on the problem.
- Reinforcement learning: RL is applied only to Nano-CC using 3,219 executable problems split into 2,847 training and 372 validation problems.The split is made at the parent-problem level to prevent related subtasks from crossing sets.
- Reinforcement learning: Each rollout receives reward 1 for full credit and 0 otherwise after compiling and executing a C++17 solution.Partial scores provide no intermediate rewards, and group-relative advantages use a leave-one-out baseline.
- Candidate selection: GenCorrect filters candidates, clusters them into K=10 centers, and selects locally well-formed representatives to promote diversity.The score-blind heuristic Q(c) uses program features, exact frequency, generated-token length, and deterministic run id without competition scores.
- Feedback and carry-forward: Each round selects a primary solution plus three complementary references based on accumulated subtask feedback and remaining score gaps.The next generation is conditioned on the problem, feedback state, primary solution, and references; official evaluator scoring is unchanged.
C.1. Test-Time Compute Prompts
The test-time prompts instruct the model to choose one incompletely solved subtask, use prior candidates as references, and produce a complete implementation for the official grader.
- Output constraints: The final response must contain one complete C++20 code block, and the official grader supplies input and output through routines rather than standard streams.The prompt also forbids adding a main function or sample grader unless explicitly requested.
- Inputs and references: The prompt supplies the problem statement, prior candidate solutions, subtask scores, best achieved scores, and maximum subtask scores.Candidate solutions are treated as potentially buggy or incomplete peer references rather than authoritative solutions.
- Target selection: The model must choose exactly one eligible subtask whose achieved score is below its maximum and generally target the largest remaining gap.Already fully achieved subtasks should not receive additional effort.
- Solution strategy: The reasoning process identifies unsolved targets, extracts useful candidate ideas, explains a sufficient approach, and then provides an implementation.If references focus on completed subtasks, the model is instructed to design a fresh approach.