Metacognition-Bench: leaderboard
Four-choice MCQ benchmark (ginigen-ai) measuring how often a model falls for metacognition traps — overconfidence and self-knowledge failures. Reported here as trap resistance, 100 × (1 − trap rate), higher is better.
Metric: Trap Resistance (100 × (1 − trap rate), higher is better). Source: huggingface.co. Saturation forecast: Estimated already saturated. 40 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 27B | 99 | #189 |
| 2 | Gemma 4 31B (IT) | 98.5 | #178 |
| 3 | Qwen 3.5 35B A3B | 98 | #256 |
| 4 | Mistral Small 3.2 | 97.7 | #605 |
| 5 | Qwen 3.5 9B | 96.5 | #384 |
| 6 | gemma-4-E4B-it | 95.2 | #564 |
| 7 | Qwen 3.6 35B A3B | 95 | #229 |
| 8 | Qwen 3.5 4B | 94.7 | #486 |
| 9 | Gemma 4 12B (IT) | 94.7 | #372 |
| 10 | Gemma 4 26B A4B (IT) | 94.5 | #292 |
| 11 | Qwen 3 1.7B | 66.4 | #1200 |
| 12 | LFM2.5-1.2B (Thinking) | 29.3 | #1488 (LFM2.5-1.2B) |
| 13 | DeepSeek-V2-Lite | 23.3 | #1363 |
| 14 | DeepSeek R1 Distill Qwen 1.5B | 22.8 | #1556 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=metacognition-bench · How It Works · Data refreshed daily, snapshot 2026-10-23.