AdaExploreBench - Exact Accuracy: leaderboard
An agent makes case-by-case decisions (four ordered actions) in finance, healthcare, cybersecurity or manufacturing under a hidden rule it must discover by designing its own probe cases, and the rule changes without announcement partway through the episode (HKUST KnowComp; 256 test episodes, main setting with masked evaluation history). Exact accuracy is the share of evaluation cases where the agent picks the optimal action.
Metric: EA (%) (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 | 69.47 | #43 |
| 2 | GPT-6 Sol | High | 65.69 | #33 |
| 3 | DeepSeek V4.1 Flash | High | 58.01 | #69 |
| 4 | GPT-6 Luna | High | 46.5 | #106 |
| 5 | Qwen 3.8 27B | Medium | 45.95 | #220 |
| 6 | Gemma 4 12B (IT) | High | 37.98 | #731 |
| 7 | Qwen 3.5 9B | 32.17 | #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-exact-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-20.