Applied Scientific RefusalBench: leaderboard
Models answer biological-research requests, some benign and some that should be refused; the score is the share of should-refuse prompts refused minus the share of benign prompts refused, so refusing everything and refusing nothing both score near zero. Unlike over-refusal boards such as OR-Bench or XSTest, it rewards discrimination between the two sets rather than willingness alone.
Metric: Refusal calibration, Youden's J (percentage points; should-refuse refusal rate minus benign refusal rate, the source's headline column). Source: huggingface.co. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.20 | 97.01 | #297 |
| 2 | Gemini 3.1 Pro (Preview) | 92.17 | #94 |
| 3 | Nemotron 3 Super | 88.91 | #768 |
| 4 | GLM-5 | 72.91 | #260 |
| 5 | Claude Sonnet 4.6 | 67.66 | #161 |
| 6 | Claude Opus 4.5 | 67.09 | #156 |
| 7 | Claude Opus 4.6 | 66.67 | #102 |
| 8 | MiniMax-M2.7 | 66.04 | #518 |
| 9 | Qwen 3 Next 80B A3B | 45.45 | #757 |
| 10 | Gemini 3.1 Flash Lite | 45.33 | #364 |
| 11 | GPT-5.5 | 42.13 | #49 |
| 12 | Claude Opus 4.7 | 23.4 | #79 |
| 13 | DeepSeek R1 | 14.24 | #473 |
| 14 | Nova Pro | 9.33 | #1555 |
| 15 | Kimi K2.6 | 8.51 | #178 |
Interactive version: theaggregate.ai/benchmark?slug=applied-scientific-refusalbench · How It Works · Data refreshed daily, snapshot 2026-11-13.