Applied Scientific RefusalBench - Benign Over-Refusal: leaderboard
Share of trials in which the model refuses a benign biological-research request, judged by a three-model council. It is the over-refusal half of the RefusalBench headline and, read alone, favours models that refuse nothing.
Metric: Benign over-refusal rate (%; share of trials on 47 benign biological-research prompts x 5 trials whose council modal label is a direct or indirect refusal). Source: huggingface.co. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.3 70B Instruct | 0 | #1357 |
| 2 | DeepSeek V3.2 | 0 | #406 |
| 3 | Gemini 3.1 Flash Lite | 0 | #364 |
| 4 | Nova Pro | 0 | #1555 |
| 5 | DeepSeek R1 | 0.43 | #473 |
| 6 | GLM-5 | 0.43 | #260 |
| 7 | Mistral Large 3 | 0.43 | #817 |
| 8 | Nemotron 3 Super | 0.43 | #768 |
| 9 | Qwen 3 Next 80B A3B | 2.55 | #757 |
| 10 | Grok 4.20 | 2.99 | #297 |
| 11 | Gemini 3.1 Pro (Preview) | 3.83 | #94 |
| 12 | GPT-5.4 Mini | 4.26 | #389 |
| 13 | MiniMax-M2.7 | 5.96 | #518 |
| 14 | Claude Sonnet 4.6 | 32.34 | #161 |
| 15 | Claude Opus 4.5 | 32.91 | #156 |
Interactive version: theaggregate.ai/benchmark?slug=applied-scientific-refusalbench-benign-over-refusal · How It Works · Data refreshed daily, snapshot 2026-11-13.