acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Competitive programming has become a key test of language model reasoning, with IOI and ICPC among its hardest settings. We present an end-to-end specialization pipeline combining problem curation, synthetic reasoning traces, supervised fine-tuning (SFT), reinforcement learning (RL), and test-time compute. Using 22,000 problems, we train Nano-CC (30B-A3B) with SFT and RL and Ultra-CC (550B-A55B) with SFT alone. Across Nano-CC, SFT drives most single-sample gains and RL adds modest improvements, while feedback-driven GenCorrect iteratively generates and refines diverse solutions and provides the largest system-level gain. 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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, post-training, competitive programming, code generation, supervised fine-tuning, reinforcement learning with verifiable rewards, test-time compute, synthetic reasoning data

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nWF9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how post-training and test-time refinement can improve language models on challenging competitive programming problems. The authors train Nano-CC and Ultra-CC using synthetic reasoning data, with SFT applied to both models and additional RL for Nano-CC. They also introduce GenCorrect, which selects diverse candidate solutions and uses execution feedback to guide further attempts. Experiments on IOI, ICPC, and LiveCodeBench Pro show improvements from the overall pipeline, with additional ablations examining the contributions of training and inference strategies.

Critical strengths
  1. Practical end-to-end system. The paper brings together data curation, post-training, and inference in a working system. Its live evaluation during IOI 2026, though unofficial, provides a useful test under real contest constraints.

  2. Substantial engineering effort. The large training corpus and extensive compute investment represent a considerable amount of work. Experiments at this scale offer useful reference points for practitioners developing large models.

  3. Broad experimental coverage. Evaluation across IOI, ICPC, and LiveCodeBench Pro, together with training and inference ablations, provides reasonably strong support for the main findings. Repeated runs and uncertainty estimates further strengthen the evaluation.

Critical weaknesses
  1. The data pipeline is not described in enough detail for others to follow, particularly the deduplication procedure, difficulty assignment, and filtering of teacher-generated traces. Since the training data cannot be released, these missing details make the work difficult to reproduce and reduce its value as a reference for future data construction.

  2. The contest score is impressive, but it does not establish that the underlying models outperform existing leading models. General Ultra-CC still trails its smaller teacher, DeepSeek-V4-Flash, on IOI and ICPC, though it does better on LCB Pro. Given the substantial training investment, this raises concerns about how efficiently the recipe uses its training data.

  3. I appreciate the engineering effort and the practical value of the results. As a research contribution, however, the paper feels more like a large-scale application of existing techniques than a methodological advance. The ablations are useful, but I find the new insights limited relative to the scale of the work.

Questions for authors
  1. Could you clarify how the 22,000 problems are turned into the final SFT datasets? In particular, how is difficulty determined, how many traces are sampled per problem at each difficulty level, and which traces are retained after execution-based checks, if any? I would also like to understand how the self-improvement examples are constructed—what feedback the teacher receives and whether a revision must improve the score to be retained. Since the data cannot be released, these details would make the training recipe much easier to build on.

  2. Could you offer a deeper analysis of the modest RL gains? The explanations in §4.4 seem plausible, but it is unclear which factors actually limit performance beyond the available compute budget. I would also be interested in whether a GenCorrect-style, feedback-conditioned rollout process could improve on independent sampling during RL. This seems worth exploring both as a methodological direction and as a way to better align training with how the models are used at inference time.

  3. Could you evaluate other strong models, particularly the teachers, with the same GenCorrect pipeline under comparable inference budgets and contest constraints? It would also be helpful to include the Ultra initialization checkpoint without the proposed domain-specific SFT. The live result is impressive; these comparisons would help establish how much the specialized training adds beyond applying the inference pipeline to an already strong model.

3 comments

Thank you for recognizing the system’s practical value, prospective evaluation, and experimental coverage.

W1 / Q1: Data construction and reproducibility

Difficulty labels were initialized from source metadata and reviewed by language-model agents using statements, constraints, and required algorithms. Each generation pass allocated 5, 20, and 50 traces per easy, medium, and hard entry. We retained complete outputs with reasoning and syntactically valid C++, removing exact prompt–response duplicates and duplicate normalized code within each prompt. Execution checks validate problem environments, not SFT traces. Self-improvement examples prompt revisions without requiring execution-verified correctness or measured improvement. We will document these procedures, prompts, teacher-specific generation and retention counts, and training–evaluation deduplication in Sections 3.1–3.2 and Appendices A–B.

W2 / Q3 Contribution of specialized training

We additionally evaluated Ultra’s actual initialization checkpoint with GenCorrect on IOI 2025.

CheckpointMean IOI 2025 GenCorrect score
Ultra RLVR initialization414.06
General Ultra-CC after domain-specific SFT502.00

Both results average five runs using the same GenCorrect procedure and maximum budget of 1,000 candidates and 50 submissions per problem. Domain-specific SFT improves the mean by 87.94 points, demonstrating benefits beyond applying GenCorrect to the initialization checkpoint. Single-sample performance also improves from 45.1% to 50.7% on IOI 2025, 50.1% to 57.4% on ICPC 2025, and 70.2% to 74.5% on LiveCodeBench Pro.

Our IOI 2026 claim concerns a historical, prospective achievement where this is the first AI system to outscore the highest-scoring human contestant on an IOI problem set. We do not claim superiority over all existing AI systems and will make this limitation explicit. Our study addresses a practical question of how we can turn an existing model into a strong competitive-programming system and evaluate post-training and test-time scaling approaches on IOI 2025. We then apply this recipe in a live evaluation at IOI 2026, using a system developed before the problems were revealed and following the same time and submission constraints as contestants.

W3: Research contribution

Our contribution is an empirical study that answers: how can post-training and test-time scaling turn an existing model into a strong competitive-programming system? Beyond demonstrating a combined system, we analyze the contributions of data generation, SFT, RL, and test-time scaling across model settings and training initializations. Matched-budget ablations further show when diversity selection and feedback-conditioned generation help. Together with the new initialization comparison, these findings inform how to adapt existing models and allocate training and inference effort. We will foreground these empirical insights and distinguish them from established methods in the introduction and Related Work.

Q2: Modest RL gains and feedback-conditioned training

Figure 6b shows that initialization strongly affects RL outcomes. After 30 steps, base-initialized Nano reaches 24.9% on IOI 2025, compared with 43.0%, 47.1%, and 48.7% after initialization from successive SFT epochs. Under this budget, RL neither closes these gaps nor consistently improves every SFT checkpoint. The modest gains therefore should not be interpreted as evidence that RL generally offers little value. Our reward structure suggests two further diagnostics. Groups with uniformly unsuccessful or successful rollouts provide no relative task-reward signal, while terminal rewards on trajectories of up to 255K tokens offer limited credit assignment.

Feedback-conditioned RL could better align training with GenCorrect. Toshniwal et al. (2026), “Learning Generative Selection for Best-of-N,” shows that targeted RL improves inference-time candidate selection. Extending this principle to generating revisions conditioned on previous attempts and execution feedback is a promising, untested direction that we will discuss.

We hope these results and clarifications address your concerns. If so, we would be grateful if you would consider revisiting your score.

Thank you for the detailed response. The new results and clarifications address some of my concerns, especially the comparison between Ultra-CC and its initialization checkpoint under the same GenCorrect setup. The additional data details are helpful, but they also raise questions about data quality, since the SFT traces are not execution-verified and self-improvement examples are not required to show actual improvement. I strongly encourage the authors to describe the full data pipeline in the paper and include more analysis of the resulting data. Such analysis may be more valuable to future research than performance gains alone. I would also encourage the authors to report teacher-model performance under the same inference setup and budget. This would make the comparisons more informative. The paper should also carefully scope its claims to avoid the system-level result being interpreted as broader superiority in model capability. Overall, I consider this a solid and practically valuable work. I will raise my score, while keeping my reservations about the experimental rigor, analysis, and framing expected from a research paper.

Thank you for the constructive follow-up and updated assessment! We will expand the pipeline description and data analysis, clarify the limits of trace filtering, and carefully scope our system-level claims. We are hoping to complete teacher evaluations under the same GenCorrect setup for the camera-ready version.

Reviewer_H3phOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores post-training and test-time scaling for large language models in competitive programming scenarios. They trained two models: Nano-CC (SFT and RL based on Nemotron-3-Nano-30B-A3B) and Ultra-CC(SFT based on Nemotron-3- Ultra-550B-A55B). They have curated over 22k problems from previous competitions and programming platforms, and used DeepSeek VisuFlash to generate 1.2 million reasoning traces for training. They apply a test-time scaling method called GenCorrect, which iteratively generates diverse solutions, incorporates evaluative feedback, and refines solutions under a constrained submission budget.

The paper demonstrates big performance gains across both the SFT and test-time scaling stages, while RL stage improves relatively small (within 3%). Through this approach, they raised the score of Nemotron-3-Nano on the IOI 2025 benchmark from 27.2 to 46.1 (pass@200). Furthermore, applying this methodology to IOI 2026 yielded results surpassing those of gold medalists.

Critical strengths
  1. The paper significantly strengthens two relatively weak Nemotron models with low coding capability through post-training, and on the Pass@1 scores for three benchmarks, IOI 2025, ICPC 2025, and LiveCodeBench Pro, we saw improvements of 10% to 50% points across the board. And they found that the main gains from this series come from SFT rather than RL.
  2. The paper also explores the effect of test-time compute on boosting model capabilities. By using test-time compute, the authors trained a silver-medal-level model to surpass gold-medal performance. Their method accounts for answer diversity during generation, sampling based on diversity, and employs a scorer to select the best and most diverse samples to facilitate model iteration.
Critical weaknesses
  1. Comparison with other models/humans: This method, after post-training on the Nemotron model, indeed improved its scores on IOI 2025, ICPC, and LCB Pro. However regarding IOI 2026, I'm not sure whether it can be fairly compared with other officially released models and humans. The authors' models receive up to 1,000 candidates per problem, five feedback rounds, and per-subtask score accumulation (§3.4, App. C). Conversely Score@200 (per-subtask maximum over 200 samples) is a stronger selection rule than the 50-submission protocol available to humans.

  2. Fair comparison among methods: In Section 3.4 regarding the GenCorrect method, it appears to be a combination of previous methods. However, in the experiments section, I didn't see a performance comparison across different test-time computation methods. I am uncertain about the following aspects when compared to evaluating the previous methods individually: (1) the actual performance gain of your method, (2) whether this combination yields a significant advantage, and (3) any additional innovations you introduce. These points require further explanation and clarification to demonstrate the superiority of the method.

Questions for authors
  1. Figure 2: According to the diagram, 22K competition problems were used for SFT, and 4K accessible problems were used exclusively for RL. My question is whether the latter might be one of the factors limiting the gains in RL capability. Since RL requires the model to produce at least some successful cases to improve learning effectiveness, one possible approach is to also perform SFT annotation on this portion of data to obtain trajectories. This would raise the ceiling for RL scaling.
  2. The introduction mentions that the small model has 1.2M trajectory data points, while the large model only has 477k for SFT. It doesn't seem to explain specifically why the large model couldn't use the same trajectories for a fairer comparison.
3 comments

Thank you for the thoughtful review and for recognizing the improvements from post-training and test-time scaling.

W1: Evaluation protocol and comparability

We will clarify the distinction between our diagnostic metrics and contest results. Score@200 is a diagnostic sampling metric used to reduce the variance when analysing components of our pipeline and was not used to obtain the live IOI 2026 result. During GenCorrect, candidate selection occurs before observing hidden-test scores, once a candidate is submitted it uses one of the 50 available submissions for that problem and unsubmitted candidates do not receive any hidden-test feedback. Accumulating the best score for each subtask across submissions follows official IOI scoring and is equally available to human contestants.

W2: GenCorrect component comparisons

We want to highlight Appendix E.2 where we compare Random 50, Diversity 50, and full GenCorrect using the same maximum budget of 1,000 candidates and 50 submissions. GenCorrect improves over non-adaptive diversity selection by 26.4 points for Nano-CC and 82.8 points for Ultra-CC, supporting the contribution of feedback-conditioned generation beyond additional sampling. We will move this comparison into the main text and will update our Related Works section to better reflect our contribution from the established techniques on which GenCorrect builds.

Q1: SFT and RL data

Thank you for this helpful question. The 4,000 executable problems used to construct the RL corpus are a subset of the 22,000 curated problems. We acknowledge that Figure 2 and its accompanying text do not clearly convey this relationship and will revise this to make it explicit.

Q2: Different SFT dataset sizes

The different trace budgets reflect compute constraints and earlier observed saturation during Ultra development. We therefore used 477,642 traces for Ultra and 1.2 million for Nano. We will include these observations in the paper and clarify that the experiments measure improvement over each model’s initialization.

We hope these clarifications address your concerns.

Thanks for your response and clarifications that have addressed my main concerns. I will maintain my positive assessment of this article :)

On the other hand, I agree with reviewer#nWF9 on that it's valuable to "describe the full data pipeline in the paper and include more analysis of the resulting data. Such analysis may be more valuable to future research than performance gains alone." If part of the dataset, the data processing methods, and environment construction methods can be open-sourced for the academic community later on to strengthen the capabilities of smaller models, that would be a great resource.

Thanks again for the encouraging feedback! We agree and will incorporate the pipeline details provided in our response to reviewer nWF9 into the paper, together with additional data statistics. We also appreciate the suggestion to share reusable data and tools and will look into what else we can release alongside the model checkpoint and inference/evaluation code already planned for release.

Reviewer_MoehOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents an end-to-end post-training pipeline for competitive programming, combining curated training data, synthetic reasoning traces, supervised fine-tuning (SFT), reinforcement learning (RL), and a test-time scaling framework called GenCorrect. The authors develop two models, Nano-CC (30B-A3B) and Ultra-CC (550B-A55B), and evaluate them on IOI 2025, ICPC 2025, and LiveCodeBench Pro. The results show substantial improvements from SFT, modest additional gains from RL, and significant improvements from test-time scaling. The paper also reports a prospective IOI 2026 evaluation in which a competition-specific Ultra-CC system scores 535.4/600, surpassing the highest-scoring human contestant.

Critical strengths

S1. Strong empirical results. The paper demonstrates impressive improvements across competitive programming benchmarks, particularly the prospective IOI 2026 result. The experiments also provide useful insights into the contributions of SFT, RL, and test-time scaling.

S2. Potential contributions to the open-source community. The release of data and models could significantly benefit the open-source community and support future research on reasoning and coding.

Critical weaknesses

W1. Limited methodological novelty.

The performance improvements largely come from distillation using stronger teacher models such as DeepSeek-V4-Flash. Moreover, GenCorrect combines existing techniques, including parallel sampling, diversity selection, and feedback-driven refinement. Similar test-time scaling strategies have already been explored in prior work, such as Competitive Programming with Large Reasoning Models (El-Kishky et al., 2025). The authors should better differentiate their methodological contributions.

W2. Overclaiming the significance of IOI 2026 results. While the IOI 2026 performance is impressive, the comparison is limited to human contestants. On IOI 2025, GLM-5.2 and DeepSeek-V4 outperform both proposed models. It is therefore plausible that these stronger models could achieve comparable or better results on IOI 2026 with similar test-time compute. Without such comparisons, the significance of the claimed achievement relative to existing AI systems remains unclear.

W3. Insufficient discussion of related work. The paper could provide a more comprehensive comparison with existing work on competitive programming, reasoning distillation, and test-time scaling.

Questions for authors

How would GLM-5.2 and DeepSeek-V4 perform on IOI 2026 under comparable inference budgets?

2 comments

Thank you for the detailed review. We clarify our contributions, relationship to prior work, and the scope of the IOI 2026 claim below.

W1 & W3: Contribution and related work

We agree that SFT, reinforcement learning, parallel sampling, diversity selection, and feedback-based refinement build on established techniques. Our contribution is an empirical study of how these components contribute to competitive-programming performance improvements culminating in a prospective IOI 2026 evaluation.

In particular, our evidence extends beyond the finding that SFT accounts for most single-sample improvement in Nano-CC. Appendix E.2 compares GenCorrect with non-adaptive baselines under the same maximum budget of 1,000 candidates and 50 submissions per problem. GenCorrect exceeds diversity selection by 26.4 points for Nano-CC and 82.8 points for Ultra-CC. Diversity selection alone substantially benefits Nano-CC but leaves Ultra-CC’s aggregate score nearly unchanged. We will foreground these findings in the main text.

We already cite El-Kishky et al. (2025) as OpenAI (2025a), but agree that the relationship deserves fuller discussion. Their work studies coding-focused RL and specialized inference strategies, whereas ours provides a more detailed and open-weight account of the post-training pipeline including synthetic data generation, SFT, RL, and test-time scaling. GenCorrect also uses successive generation conditioned on subtask feedback and complementary previous solutions, rather than their subtask-specific sampling, execution-based clustering, and reranking. We will expand Related Work to clarify these differences and also distinguish our empirical contributions from the established techniques on which the pipeline builds.

W2 / Question: Scope of the IOI 2026 result

Our claim concerns a historical, prospective achievement where this is the first AI system to outscore the highest-scoring human contestant on an IOI problem set. We do not claim that Ultra-CC outperforms all existing AI systems and will make this limitation explicit. Our study addresses a practical question of how we can turn an existing model into a strong competitive-programming system and evaluate post-training and test-time scaling approaches on IOI 2025. We then apply this recipe in a live evaluation at IOI 2026, using a system developed before the problems were revealed and following the same time and submission constraints as contestants.

We hope these results address your concerns. If so, we would be grateful if you would consider revisiting your score.

Thank you again for your thoughtful review. With less than a day remaining in the discussion period, we would appreciate your thoughts on our response and whether it addresses your concerns. If any concerns remain, we would welcome the opportunity to clarify them before the deadline.

Comments on the paper

1 comment

We want to thank the reviewers for their thoughtful feedback and constructive discussion. Based on the comments below we have uploaded a revised manuscript expanding the data-pipeline description and related work, adding the GenCorrect baselines and Ultra initialization comparison to the main text, and clarifying the evaluation protocol and scope of our claims. We will work to document further pipeline details and complete teacher-model evaluations under the same GenCorrect setup for the camera-ready version.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.