SYCON-Bench - False Presupposition: leaderboard
Questions carry a hidden false premise and the user insists on it over five turns; the score is the mean turn at which the model stops correcting it. It mixes knowing the premise is false with resisting pressure, unlike the debate board.
Metric: Turn of Flip on false-presupposition questions (0-5; higher = keeps correcting the premise longer). Source: arxiv.org. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek R1 | 3.21 | #473 |
| 2 | O3 Mini | 2.98 | #533 |
| 3 | GPT-4o | 2.92 | #741 |
| 4 | Claude 3.7 Sonnet | 2.92 | #458 |
| 5 | DeepSeek V3 | 2.88 | #716 |
| 6 | Llama 3.1 70B | 2.4 | #1490 |
| 7 | Qwen 2.5 72B Instruct | 2.35 | #1092 |
| 8 | Qwen 2.5 14B Instruct | 2.31 | #1732 |
| 9 | Qwen 2.5 72B | 2.15 | #1190 |
| 10 | Qwen 2.5 14B | 2.12 | #1785 |
| 11 | Qwen 2.5 7B Instruct | 1.93 | #2564 |
| 12 | Llama 3.3 70B Instruct | 1.9 | #1357 |
| 13 | Gemma 2 9B (IT) | 1.86 | #2262 |
| 14 | Qwen 2.5 7B | 1.65 | #2703 |
| 15 | Llama 3.1 8B | 1.59 | #4861 |
Interactive version: theaggregate.ai/benchmark?slug=sycon-bench-false-presupposition · How It Works · Data refreshed daily, snapshot 2026-11-13.