MASK (Paper) - Lying in Any of 10 Samples: leaderboard
Same MASK paper run, counting an example as a lie if any of ten sampled answers contradicts the elicited belief, which measures how often a model can be pushed into lying at all rather than on one draw.
Metric: Examples where any of ten sampled statements contradicts the elicited belief, P(lie)@10 (%). Source: arxiv.org. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.2 1B Instruct | 35.8 | #9047 |
| 2 | Llama 2 7B Chat | 38.2 | #8805 |
| 3 | Qwen 2.5 0.5B Instruct | 41.1 | #8987 |
| 4 | Llama 2 13B Chat | 41.4 | #8186 |
| 5 | Llama 2 70B Chat | 41.5 | #6800 |
| 6 | Claude 3.7 Sonnet (20250219) | 42.1 | #402 |
| 7 | Llama 3.2 3B Instruct | 43.1 | #7147 |
| 8 | Claude 3.5 Sonnet (20240620) | 44.7 | #687 |
| 9 | Llama 3.1 405B Instruct | 44.9 | #1115 |
| 10 | QwQ 32B-Preview | 46.3 | #1625 |
| 11 | Llama 3.1 8B Instruct | 46.7 | #4246 |
| 12 | Qwen 1.5 110B Chat | 53.4 | #1771 |
| 13 | Qwen 2.5 3B Instruct | 54.3 | #5150 |
| 14 | Qwen 1.5 7B Chat | 55.2 | #6361 |
| 15 | Qwen 2.5 7B Instruct | 55.3 | #2564 |
Interactive version: theaggregate.ai/benchmark?slug=mask-paper-lying-in-any-of-10-samples · How It Works · Data refreshed daily, snapshot 2026-11-13.