AdaExploreBench - Calibrated Accuracy: leaderboard
AdaExploreBench's headline score: the agent discovers a hidden decision rule by designing its own probes and must notice when the rule silently drifts (256 test episodes, main setting). Calibrated accuracy gives near-optimal choices partial credit by the utility they earn and is rescaled so a uniformly random policy scores 0 and the optimal policy 100.
Metric: CA (%) (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 | 67.25 | #43 |
| 2 | GPT-6 Sol | High | 61.87 | #33 |
| 3 | DeepSeek V4.1 Flash | High | 52.87 | #69 |
| 4 | Qwen 3.8 27B | Medium | 37.73 | #220 |
| 5 | GPT-6 Luna | High | 35.1 | #106 |
| 6 | Gemma 4 12B (IT) | High | 20.59 | #731 |
| 7 | Qwen 3.5 9B | 12.31 | #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-calibrated-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-20.