Grip on LLMs - HonestCityBench - Incorrect Premise: leaderboard
Grip on LLMs (arXiv 2608.09925v1, Appendix Table 2): HonestCityBench, 530 Dutch prompts that a model cannot or should not answer directly. Share of responses judged to appropriately handle a prompt built on an incorrect premise, percent. Higher is better.
Metric: Appropriate acknowledgement rate (%; LLM judge). Source: arxiv.org. Saturation forecast: Not forecast. 31 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 70 | #357 |
| 2 | Gemma 3 12B (IT) | 68 | #672 |
| 3 | Gemma 3 27B (IT) | 65 | #528 |
| 4 | Mistral Small 3 | 62 | #630 |
| 5 | Qwen 3 8B | 61 | #678 |
| 6 | Mistral Medium 3 | 61 | #507 |
| 7 | Qwen 3 32B | 59 | #449 |
| 8 | Mistral Large 3 | 57 | #410 |
| 9 | aya-expanse-32B | 57 | #687 |
| 10 | OLMo-2-0325-32B-Instruct | 57 | #847 |
| 11 | GPT-4o Mini | 51 | #598 |
| 12 | Llama 3.1 8B Instruct | 49 | #1017 |
| 13 | EuroLLM-9B-Instruct | 49 | #1047 |
| 14 | DeepSeek R1 Distill Qwen 32B | 43 | #662 |
| 15 | Phi-4 Mini Instruct | 43 | #1232 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=grip-on-llms-honestcitybench-incorrect-premise · How It Works · Data refreshed daily, snapshot 2026-11-05.