BeHonest - Persona Sycophancy: leaderboard
The model answers the same question with and without a user persona that implies a view; the number is the share of answers that move toward the persona. Lower is better.
Metric: Answers changed to match the persona the user adopts (%). Source: arxiv.org. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 2 7B Chat | 23.4 | #8805 |
| 2 | Llama 3 8B Instruct | 25.74 | #4612 |
| 3 | Llama 2 70B Chat | 26.81 | #6800 |
| 4 | Llama 2 13B Chat | 27.66 | #8186 |
| 5 | Qwen 1.5 14B Chat | 30.64 | #2974 |
| 6 | Llama 3 70B Instruct | 33.62 | #1676 |
| 7 | GPT-3.5 Turbo (0125) | 38.39 | #3500 |
| 8 | GPT-4o (2024-05-13) | 39.44 | #735 |
| 9 | Mistral 7B Instruct (v0.2) | 39.53 | #7007 |
Interactive version: theaggregate.ai/benchmark?slug=behonest-persona-sycophancy · How It Works · Data refreshed daily, snapshot 2026-11-13.