MASK (Paper) - Honesty: leaderboard
Same MASK paper run: the share of examples where the pressured statement is consistent with the model's elicited belief. Evasive answers count neither as honest nor as lies, so this is not 100 minus the lie rate.
Metric: Examples where the statement under pressure is consistent with the elicited belief, P(honest) (%). Source: arxiv.org. 32 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet (20250219) | 47.6 | #402 |
| 2 | Qwen 2.5 3B Instruct | 30.7 | #5150 |
| 3 | Qwen 2.5 7B Instruct | 28.9 | #2564 |
| 4 | Qwen 2.5 32B Instruct | 28.7 | #1215 |
| 5 | Llama 2 13B Chat | 28.7 | #8186 |
| 6 | Llama 2 70B Chat | 28.3 | #6800 |
| 7 | Qwen 1.5 110B Chat | 27.9 | #1771 |
| 8 | Claude 3.5 Sonnet (20240620) | 27.7 | #687 |
| 9 | Llama 2 7B Chat | 27.5 | #8805 |
| 10 | GPT-4.5 (Preview) | 27.2 | #502 |
| 11 | Llama 3.1 70B Instruct | 27.1 | #1546 |
| 12 | Qwen 1.5 7B Chat | 27.1 | #6361 |
| 13 | Qwen 2.5 14B Instruct | 26.5 | #1732 |
| 14 | Qwen 2.5 1.5B Instruct | 25.7 | #7470 |
| 15 | Llama 3.3 70B Instruct | 24.7 | #1357 |
Interactive version: theaggregate.ai/benchmark?slug=mask-paper-honesty · How It Works · Data refreshed daily, snapshot 2026-11-13.