SYCON-Bench - Debate: leaderboard
The model is told to defend a stance on a controversial topic while a simulated user argues against it for five turns; the score is the mean turn at which it first abandons the stance (5 = never). Unlike single-turn sycophancy tests, it measures how long a model holds out under repeated pressure.
Metric: Turn of Flip in the debate setting (0-5; higher = holds its assigned stance longer). Source: arxiv.org. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 Mini | 4.97 | #533 |
| 2 | Qwen 2.5 72B Instruct | 4.9 | #1092 |
| 3 | Llama 3.3 70B Instruct | 4.85 | #1357 |
| 4 | DeepSeek R1 | 4.85 | #473 |
| 5 | GPT-4o | 4.67 | #741 |
| 6 | Claude 3.7 Sonnet | 4.47 | #458 |
| 7 | DeepSeek V3 | 4.16 | #716 |
| 8 | Qwen 2.5 14B Instruct | 3.65 | #1732 |
| 9 | Gemma 2 9B (IT) | 3.25 | #2262 |
| 10 | Llama 3.1 8B Instruct | 2.44 | #4246 |
| 11 | Qwen 2.5 7B Instruct | 0.83 | #2564 |
Interactive version: theaggregate.ai/benchmark?slug=sycon-bench-debate · How It Works · Data refreshed daily, snapshot 2026-11-13.