UOJ-Bench - Repair (Hard): leaderboard
Metric: Pass@1 (%) on the 216 hard repair instances (submissions broken by community hacks): the patched code, changing at most 10 percent of the original, must pass every test including the hack, single-turn zero-shot Pass@1 judged by the Universal Online Judge through its internal API (problem statements translated from Chinese to English; maximum response length at least 64k tokens); higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 8 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 48.15 |
| 2 | GPT-5 (High) | 38.89 |
| 3 | GPT-5 | 32.87 |
| 4 | Kimi K2 (Thinking) | 26.85 |
| 5 | DeepSeek V3.2 (Thinking) | 24.07 |
| 6 | GPT-OSS-120B | 9.72 |
| 7 | Claude Opus 4.5 | 6.94 |
| 8 | Qwen 3 Coder 480B A35B Instruct | 3.7 |
Interactive version: theaggregate.ai/benchmark?slug=uoj-bench-repair-hard · How It Works · Data refreshed daily, snapshot 2026-09-29.