TF-RefusalBench - Summarization Refusal: leaderboard
Over-refusal on legitimate court work (Wuhrmann et al., 2026): 2,800 summarization prompts over severe criminal-law passages of public Swiss Federal Supreme Court rulings in German, French, Italian and English. Share of responses that decline the task, by majority of three LLM judges, in percent; lower is better.
Metric: Refusal rate (%; share of 2,800 summarization prompts over Swiss Federal Supreme Court criminal-law passages in which the model declines the task for content or ethical reasons; 3 samples per prompt at temperature 0.7, 2-of-3 majority of three LLM judges). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 32B | 0 | #960 |
| 2 | Gemma 4 31B (IT) | 0 | #343 |
| 3 | Llama 3.3 70B Instruct | 7.2 | #1353 |
| 4 | GPT-OSS-20B | 14.7 | #1150 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=tf-refusalbench-summarization-refusal · How It Works · Data refreshed daily, snapshot 2026-11-12.