跳转至

文章背景与核心概要

竞技编程(Competitive Programming)是评估大语言模型(LLM)高级推理能力的关键基准,其中国际信息学奥林匹克竞赛(IOI)和国际大学生程序设计竞赛(ICPC)最具挑战性。本文介绍了一种端到端的模型专业化流水线,整合了大规模问题筛选(22,000个精选问题)、合成推理轨迹、监督微调(SFT)和强化学习(RL)。利用该方法,作者训练出了经过SFT和RL训练的 Nemotron-3-Nano-CC(30B-A3B)模型,以及仅通过SFT训练的 Nemotron-3-Ultra-CC(550B-A55B)模型。此外,他们还开发了 GenCorrect——一种反馈驱动的测试时计算技术,能够迭代生成、评估和精炼多样化的解决方案。

在IOI 2025的评估中,Nano-CC 模型在后训练后从初始的130分提升至291分,并在结合 GenCorrect 后达到了 468 分(超过了 438.3 的金牌线),而 Ultra-CC 模型则达到了 502 分。在 IOI 2026 的前瞻性评估中,在与人类参赛者完全相同的约束条件(时间限制、互联网访问限制和提交次数上限)下,特定竞赛版 Ultra-CC 系统在 600 分中获得了 535.4 分,超越了 361.12 的金牌线以及人类最高分 498.27。据作者所知,这是人工智能系统在官方 IOI 题目集上的得分首次超过人类最高分选手。


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

arXiv: 2609.02849 [cs.LG]
Submitted: September 2, 2026
Authors: Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar, Boris Ginsburg
Primary Subject: Machine Learning (cs.LG)
Additional Subjects: Artificial Intelligence (cs.AI), Computation and Language (cs.CL), Multiagent Systems (cs.MA), Software Engineering (cs.SE)


Executive Summary / 执行摘要

竞技编程是评估大语言模型(LLM)高级推理能力的决定性基准,特别是通过国际信息学奥林匹克竞赛(IOI)和国际大学生程序设计竞赛(ICPC)等久负盛名的国际赛事。

Competitive programming serves as a definitive benchmark for evaluating the advanced reasoning capabilities of Large Language Models (LLMs), particularly through prestigious international contests like the International Olympiad in Informatics (IOI) and the International Collegiate Programming Contest (ICPC).

本文介绍了一种端到端的模型专业化流水线,集成了: * 大规模问题筛选(22,000个精选问题) * 合成推理轨迹 * 监督微调(SFT) * 强化学习(RL)

This paper introduces an end-to-end model specialization pipeline that integrates: * Large-scale problem curation (22,000 curated problems) * Synthetic reasoning traces * Supervised Fine-Tuning (SFT) * Reinforcement Learning (RL)

利用这种方法,作者使用 SFT 和 RL 训练了 Nemotron-3-Nano-CC (30B-A3B),并仅使用 SFT 训练了 Nemotron-3-Ultra-CC (550B-A55B)。此外,他们开发了 GenCorrect,这是一种反馈驱动的测试时计算技术,旨在迭代生成、评估和精炼多样化的解决方案。

Using this methodology, the authors trained Nemotron-3-Nano-CC (30B-A3B) with both SFT and RL, and Nemotron-3-Ultra-CC (550B-A55B) using SFT alone. Furthermore, they developed GenCorrect, a feedback-driven test-time compute technique designed to iteratively generate, evaluate, and refine diverse solutions.

关键里程碑与结果

  • IOI 2025 评估:
  • Nano-CC 模型在后训练后从最初的 130 分提升至 291 分,在使用 GenCorrect 后达到 468 分(超过了 438.3 的金牌门槛)。
  • Ultra-CC 模型达到了 502 分
  • IOI 2026 前瞻性评估:
  • 在与人类参赛者相同的约束条件(时间限制、互联网访问限制和提交上限)下,特定竞赛的 Ultra-CC 系统在 600 分中获得了 535.4 分
  • 这一表现超过了 361.12 的金牌门槛以及 498.27 的人类最高分。
  • 意义: 据作者所知,这标志着人工智能系统在官方 IOI 题目集上的得分首次超过人类最高分选手

Key Milestones & Results

  • IOI 2025 Evaluation:
  • The Nano-CC model improved from an initial 130 points to 291 points post-training, and reached 468 points using GenCorrect (surpassing the gold medal threshold of 438.3).
  • The Ultra-CC model reached 502 points.
  • IOI 2026 Prospective Evaluation:
  • Operating under identical constraints to human contestants (time limits, internet access restrictions, and submission caps), a competition-specific Ultra-CC system scored 535.4 out of 600.
  • This performance surpassed both the gold threshold of 361.12 and the top human score of 498.27.
  • Significance: To the best of the authors' knowledge, this marks the first time an AI system has outscored the highest-scoring human contestant on an official IOI problem set.

Abstract / 摘要

竞技编程已成为大语言模型推理能力的关键测试,IOI 和 ICPC 等国际赛事代表了其最具挑战性的场景。我们提出了一种端到端的专业化流水线,结合了大范围问题筛选、合成推理轨迹、监督微调(SFT)和强化学习(RL)。利用 22,000 个精选问题,我们使用 SFT 和 RL 训练了 Nemotron-3-Nano-CC (30B-A3B),仅使用 SFT 训练了 Nemotron-3-Ultra-CC (550B-A55B)。我们进一步引入了 GenCorrect,这是一种反馈驱动的测试时计算策略,可以迭代生成、评估和精炼多样化的解决方案。在 IOI 2025 上,Nano-CC 在后训练后从 130 分提高到 291 分,使用 GenCorrect 后提高到 468 分,超过了 438.3 的金牌门槛,而 Ultra-CC 则达到了 502 分。在这些结果的指导下,我们开发了一个特定竞赛的 Ultra-CC 系统,并在 IOI 2026 期间对其进行了前瞻性评估。在与人类参赛者相同的时间、互联网访问和提交约束条件下,它在 600 分中得到了 535.4 分,超过了 361.12 的金牌门槛和 498.27 的人类最高分。据我们所知,这是第一个在 IOI 题目集上得分超过最高分人类选手的 AI 系统。

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.