BeHonest - Preference Sycophancy: leaderboard
The model states a preference, then the user states the opposite one; the number is the share of times it switches. Lower is better.
Metric: Preferences changed after the user states the opposite preference (%). Source: arxiv.org. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-05-13) | 24.11 | #735 |
| 2 | Llama 3 70B Instruct | 33.07 | #1676 |
| 3 | Llama 2 70B Chat | 46.52 | #6800 |
| 4 | GPT-3.5 Turbo (0125) | 48.78 | #3500 |
| 5 | Llama 2 13B Chat | 54.35 | #8186 |
| 6 | Qwen 1.5 14B Chat | 57.39 | #2974 |
| 7 | Llama 2 7B Chat | 61.74 | #8805 |
| 8 | Llama 3 8B Instruct | 78.02 | #4612 |
| 9 | Mistral 7B Instruct (v0.2) | 80.21 | #7007 |
Interactive version: theaggregate.ai/benchmark?slug=behonest-preference-sycophancy · How It Works · Data refreshed daily, snapshot 2026-11-13.