SYCON-Bench - Unethical Queries: leaderboard
The user asks questions that embed a stereotype and keeps pressing over five turns; the score is the mean turn at which the model stops challenging the premise. Includes base checkpoints as separate rows, unlike the debate board.
Metric: Turn of Flip on stereotype-embedding queries (0-5; higher = keeps challenging the premise longer). Source: arxiv.org. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 2.73 | #458 |
| 2 | DeepSeek R1 | 2.72 | #473 |
| 3 | Gemma 2 9B (IT) | 2.36 | #2262 |
| 4 | O3 Mini | 2.31 | #533 |
| 5 | DeepSeek V3 | 1.99 | #716 |
| 6 | Llama 3.1 70B | 1.99 | #1490 |
| 7 | Qwen 2.5 72B | 1.77 | #1190 |
| 8 | Llama 3.1 8B | 1.53 | #4861 |
| 9 | Qwen 2.5 14B | 1.41 | #1785 |
| 10 | Llama 3.3 70B Instruct | 1.39 | #1357 |
| 11 | Qwen 2.5 7B | 1.33 | #2703 |
| 12 | Qwen 2.5 72B Instruct | 1.32 | #1092 |
| 13 | GPT-4o | 1.23 | #741 |
| 14 | Gemma 2 9B | 1.17 | #2802 |
| 15 | Qwen 2.5 14B Instruct | 1.14 | #1732 |
Interactive version: theaggregate.ai/benchmark?slug=sycon-bench-unethical-queries · How It Works · Data refreshed daily, snapshot 2026-11-13.