BeHonest - Admitting Unknowns: leaderboard
The model is asked 7,648 questions nobody can answer (SelfAware, UnknownBench); the number is the share it declines without being told to. Unlike hallucination benchmarks that grade answers, it rewards admitting ignorance.
Metric: Unanswerable questions the model declines to answer without being told to (%). Source: arxiv.org. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral 7B Instruct (v0.2) | 50.03 | #7007 |
| 2 | Llama 3 70B Instruct | 48.81 | #1676 |
| 3 | Llama 3 8B Instruct | 37.8 | #4612 |
| 4 | Qwen 1.5 14B Chat | 37.03 | #2974 |
| 5 | Llama 2 13B Chat | 32.24 | #8186 |
| 6 | GPT-4o (2024-05-13) | 31.37 | #735 |
| 7 | Llama 2 7B Chat | 27.82 | #8805 |
| 8 | Llama 2 70B Chat | 26.4 | #6800 |
| 9 | GPT-3.5 Turbo (0125) | 21.78 | #3500 |
Interactive version: theaggregate.ai/benchmark?slug=behonest-admitting-unknowns · How It Works · Data refreshed daily, snapshot 2026-11-13.