PARROT Sycophancy: leaderboard
Each of 1,302 MMLU-style questions is asked once neutrally and once with a user claiming expert authority for a wrong option; the score is how often the model then picks that wrong option. Not the PARROT SQL dialect benchmark; unlike multi-turn tests, the pressure is a single authority claim.
Metric: Follow rate (%): share of 1,302 questions where the model adopts the wrong answer an authoritative user asserts; lower is better. Source: arxiv.org. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 3.6 | #167 |
| 2 | GPT-5 Mini (2025-08-07) | 6.3 | #343 |
| 3 | Grok 4 Fast (Reasoning) | 7.6 | #441 (Grok 4 Fast) |
| 4 | GPT-4.1 | 10.2 | #495 |
| 5 | Claude Sonnet 4.5 | 10.8 | #250 |
| 6 | GPT-4o | 16.1 | #741 |
| 7 | Gemini 2.5 Flash | 17.2 | #453 |
| 8 | GPT-4o Mini | 18.4 | #1623 |
| 9 | Gemini 2.0 Flash | 20.9 | #704 |
| 10 | Gemini 2.0 Flash Lite | 23.1 | #1032 |
| 11 | Grok 4 Fast (Non-reasoning) | 32.9 | #441 (Grok 4 Fast) |
| 12 | Qwen 2.5 14B Instruct | 35.9 | #1732 |
| 13 | Gemma 3 27B (IT) | 40.2 | #1289 |
| 14 | Gemini 2.5 Flash Lite | 50.7 | #958 |
| 15 | Gemma 3 12B (IT) | 51.5 | #1837 |
Interactive version: theaggregate.ai/benchmark?slug=parrot-sycophancy · How It Works · Data refreshed daily, snapshot 2026-11-13.