Papers.Data.Code
Paper

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

Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar, Boris Ginsburg

A specialization pipeline for coding contests that adds SFT, executable-reward RL, and feedback-driven iterative correction to general LLMs.

arXiv ↗Sep 2, 2026Code ↗

  • llm
  • nlp
  • code generation
  • post-training
  • reinforcement learning
Comparative benchmark bar chart showing model vs baselines across multiple tasks
Comparative benchmark bar chart showing model vs baselines across multiple tasks

What it is

An end-to-end post-training pipeline plus iterative test-time correction specializes LLMs for competitive programming and reaches gold-medal to superhuman IOI results. Instead of relying on base-model scaling alone, it combines curated contest problems, synthetic reasoning traces, SFT, optional executable-reward RL, and GenCorrect: iterative generate-evaluate-refine under submission limits.

Key results

  • IOI 2026: 535.4/600, above top human 498.27 and gold 361.12
  • IOI 2025 Nano-CC: 130 -> 291 Score@1 post-training, 468 with GenCorrect
  • Ultra-CC on IOI 2025: 502 after 5 GenCorrect rounds; Nano-CC reaches 468

For competition coding, most gains come from domain-specific SFT, while iterative evaluator-guided test-time refinement is what pushes models to medal-level performance.

One useful ML email a week.

More from Sep 07 – Sep 13

The whole week →

Picked Sep 11

Picked Sep 10

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.