UOJ-Bench - Hacking (Hard): leaderboard
Code hacking on covert errors in UOJ-Bench (2026): given a problem and an accepted submission that the community later broke, write a generator for an input that makes it fail. UOJ-Bench (2026) draws on real Universal Online Judge submissions and community hacks and judges every answer on the judge itself; Pass@1 in percent, higher is better.
Metric: Pass@1 (%) on the 1,046 hard hacking instances (accepted submissions later broken by community hacks, covert errors): the generated input-synthesis script must make the target submission fail on UOJ, 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 February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 44.17 | #113 |
| 2 | GPT-5 | 38.15 | #167 |
| 3 | GPT-5 (High) | 36.9 | #167 (GPT-5) |
| 4 | Kimi K2 (Thinking) | 28.2 | #439 (Kimi K2) |
| 5 | DeepSeek V3.2 (Thinking) | 25.43 | #406 (DeepSeek V3.2) |
| 6 | GPT-OSS-120B | 20.17 | #700 |
| 7 | Claude Opus 4.5 | 8.99 | #156 |
| 8 | Qwen 3 Coder 480B A35B Instruct | 5.64 | #643 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=uoj-bench-hacking-hard · How It Works · Data refreshed daily, snapshot 2026-11-13.