Sign InOpen Brain
arXivPaperNeeds Review

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

A coding-specialized model paired post-training with an iterative generate-evaluate-refine loop to exceed the top IOI 2026 human score. The reusable idea is feedback-driven test-time search.

arXiv · Sep 2, 2026
Open Source Open MarkdownOpen JSON
Source Summary

The pipeline curated **22,000 problems** for SFT and RL, then used **GenCorrect** to generate, evaluate, and refine diverse solutions. Nano-CC rose from 130 to 468 points on IOI 2025; Ultra-CC reached 502.

Practical Implication

For difficult coding tasks, allocate inference budget to multiple candidates, automated evaluation, and repair rather than relying on one completion. Specialized training data also mattered substantially before test-time search.

Agent-Ready Context
The pipeline curated **22,000 problems** for SFT and RL, then used **GenCorrect** to generate, evaluate, and refine diverse solutions. Nano-CC rose from 130 to 468 points on IOI 2025; Ultra-CC reached 502.

For difficult coding tasks, allocate inference budget to multiple candidates, automated evaluation, and repair rather than relying on one completion. Specialized training data also mattered substantially before test-time search.

The prospective IOI 2026 system scored **535.4/600**, above the top human’s 498.27 under the same stated constraints. Competition performance does not establish comparable reliability on repository-scale engineering.
Connected Context · Feed7 Judgment

This provides strong competition-specific evidence that specialized post-training and generate-evaluate-repair search can outperform a single completion. It shifts coding-model assessment toward the combined training-and-inference system, while explicitly limiting the result: exceptional algorithmic contest scores do not establish repository-scale engineering reliability.

Context Map
modelcoding#reasoning#coding-agents#model-selection
Uncertainty
The prospective IOI 2026 system scored **535.4/600**, above the top human’s 498.27 under the same stated constraints. Competition performance does not establish comparable reliability on repository-scale engineering.