AdaExploreBench - Factor Recognition: leaderboard
AdaExploreBench's auxiliary check of what the agent understood, asked outside the interaction so it leaks no evidence: the F1 between the factors the agent says the hidden rule uses and the true factors, averaged over the assessments before and after the unannounced rule change (main setting, 256 episodes).
Metric: Factor Rec. (%) (0-100). Source: hkust-knowcomp.github.io. Saturation forecast: Around February 2027. 7 models tracked. Updated 19 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Gemini 3.8 Flash | High | 78.3 | #43 |
| 2 | DeepSeek V4.1 Flash | High | 78.07 | #69 |
| 3 | GPT-6 Sol | High | 73.79 | #33 |
| 4 | Qwen 3.8 27B | Medium | 65.87 | #220 |
| 5 | GPT-6 Luna | High | 56.66 | #106 |
| 6 | Gemma 4 12B (IT) | High | 53.21 | #731 |
| 7 | Qwen 3.5 9B | 39.4 | #829 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=adaexplorebench-factor-recognition · How It Works · Data refreshed daily, snapshot 2026-11-20.