ADeptS-Bench - Disambiguation (Desktop): leaderboard
ADeptS-Bench (2026) tests computer-use agents on single decision points in mobile and desktop screenshots. In the disambiguation stream the desktop instruction is ambiguous at the decision point and the agent should ask; the score is the F1 of its clarification questions against human ground-truth items, judged by Gemini 3.1 Pro, in percent; higher is better.
Metric: Disambiguation F1 (%; harmonic mean of the precision and recall of the clarification questions against human ground-truth items; 363 desktop tasks; ambiguous GUI decision points reached through a trajectory of screenshots; generated clarification questions matched to ground-truth items by a Gemini 3.1 Pro judge; with severity scoring prompt). Source: arxiv.org. Saturation forecast: Around 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 47.6 | #79 |
| 2 | Qwen 3 VL 235B A22B | 46.3 | #479 |
| 3 | Gemini 3.1 Pro (Preview) | 39.8 | #94 |
| 4 | GPT-5.4 | 30.3 | #123 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=adepts-bench-disambiguation-desktop · How It Works · Data refreshed daily, snapshot 2026-11-13.