RubricBench (Self-Generated Rubrics): leaderboard
Metric: Preference accuracy (%) on RubricBench's 1,147 preference pairs (instruction following, STEM, code, safety and chat pairs re-curated from HelpSteer3, PPE and RewardBench2 and filtered so that surface cues such as length, formatting or tone favour the rejected response); the judge first writes a rubric of binary checks from the instruction and then verifies both responses against it (the OpenRubric pipeline of Table 2); higher is better. Source: arxiv.org. Saturation forecast: Around March 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 60.4 | #77 |
| 2 | Qwen 3.5 Plus | 59.3 | #123 |
| 3 | Gemini 3 Flash | 58 | #93 |
| 4 | DeepSeek V3.2 (Non-reasoning) | 57.8 | #198 (DeepSeek V3.2) |
| 5 | GPT-OSS-120B | 56.4 | #330 |
| 6 | GPT-5.1 | 54.6 | #131 |
| 7 | GPT-4o Mini | 46.7 | #588 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=rubricbench-self-generated-rubrics · How It Works · Data refreshed daily, snapshot 2026-10-11.