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

Automated programming has evolved from rule-based systems to neural network-based models like Codex and AlphaCode, which have demonstrated impressive code synthesis capabilities. Prior works focused on supervised learning with limited datasets, achieving partial success in competitive environments. Recent advances include large-scale pretraining and fine-tuning, but these still fall short in complex algorithmic reasoning required in IOI and ICPC. The challenge lies in enabling models to generate correct, efficient solutions within strict time and submission constraints, often requiring multi-step reasoning, subtask decomposition, and iterative refinement. Existing methods lack scalable strategies to incorporate synthetic data and reinforcement learning effectively, limiting their performance in high-stakes competitions.

The core challenge is how to significantly improve AI performance in competitive programming by leveraging large-scale data, synthetic reasoning, and multi-stage optimization. Traditional models are constrained by limited training data, shallow fine-tuning, and lack of effective test-time solution refinement. Achieving human-level or superhuman performance requires overcoming these bottlenecks, especially under strict submission limits and complex problem constraints. The key question is how to design a training and inference pipeline that combines data augmentation, multi-round feedback, and reinforcement learning to push the model’s reasoning and coding capabilities beyond current limits.

The main innovations include: 1) Curating an extensive, diverse problem set from global competitions, ensuring broad coverage; 2) Generating synthetic reasoning traces to expose models to complex problem-solving behaviors; 3) Employing long-context supervised fine-tuning to enhance understanding of multi-step reasoning; 4) Integrating reinforcement learning with executable rewards to optimize code solutions; 5) Developing GenCorrect, a multi-round, feedback-driven test-time strategy that refines solutions iteratively. These components form a cohesive system that surpasses previous state-of-the-art models, enabling AI to excel in high-level reasoning tasks within limited attempts.

• �� Data curation: Collect 22,000 problems from international competitions and online platforms, automate packaging into evaluation environments, filter for consistency.
• �� Synthetic trace generation: Use DeepSeek-V4-Flash to produce 1.2 million reasoning traces, covering various difficulty levels.
• �� Supervised fine-tuning: Fine-tune Nano (30B) over three epochs and Ultra (550B) over one epoch, using long-context sequences (up to 262K tokens) to improve comprehension.
• �� Reinforcement learning: Apply GRPO algorithm to Nano, training on 3,219 filtered problems with executable code rewards.
• �� Test-time optimization: Implement GenCorrect, generating up to 200 solutions per round, selecting diverse candidates based on similarity, submitting top solutions, and iteratively refining based on feedback.
• �� Evaluation: Test models on IOI 2025, ICPC 2025, and LiveCodeBench Pro, measuring performance via Score@1 and Pass@1, conducting ablation studies to assess contributions of each component.

The evaluation involved multiple benchmarks: IOI 2025, ICPC 2025, and LCB Pro, with data carefully curated to exclude training overlap. Metrics included Score@1, Pass@1, and solution diversity. Baselines such as DeepSeek-V4, GLM-5.2, and previous models were compared. Hyperparameters like generation rounds, subtask filtering, and reward weights were tuned. Ablation studies confirmed the importance of synthetic data, SFT, RL, and multi-round refinement. Results demonstrated that Nano-CC improved from 130 to 468 on IOI 2025, while Ultra-CC reached 502 with only SFT, validating the pipeline's effectiveness. Multi-round GenCorrect further boosted scores, especially at larger scales.

Nano-CC’s performance on IOI 2025 exceeded the gold threshold with a Score@1 of 468, and during IOI 2026, the competition-specific Ultra-CC scored 535.4, surpassing the top human score of 498.27. The ablation experiments showed that SFT contributed the most to initial gains, while RL provided incremental improvements. Multi-round iterative refinement via GenCorrect significantly enhanced solution quality, especially for larger models. These results establish a new benchmark for AI in competitive programming, demonstrating the effectiveness of combining synthetic data, multi-stage training, and test-time optimization.

The developed pipeline can be applied to automated code synthesis, intelligent tutoring systems, and software automation. It enables AI to generate, evaluate, and refine solutions in real-time, reducing human effort and error. The approach is adaptable to various domains requiring complex reasoning under constraints, such as automated testing, bug fixing, and educational tools. Long-term, this technology could revolutionize software development workflows, enabling fully autonomous programming assistants and adaptive learning platforms.

Despite impressive results, the approach relies heavily on synthetic data generation and high computational resources, limiting accessibility. Its performance on unseen, highly complex, or domain-specific problems remains uncertain. Multi-round optimization introduces latency, making real-time deployment challenging. Further research is needed to improve generalization, reduce costs, and extend capabilities to broader problem domains.