AndroidIntent - Complete Instructions: leaderboard
GUI-execution setting of AndroidIntent (PersonalAlign, Lyu et al., Harbin Institute of Technology, Shenzhen, ACL 2026), built from two months of real Android users' phone records across 130 apps: the agent is scored offline, step by step against the user's recorded actions, on 775 interactions given complete instructions. The score is the paper's cumulative success rate in percent, which weights step success toward earlier steps; higher is better.
Metric: Cumulative success rate (%; offline step-wise execution of 775 recorded user interactions with complete instructions, step success weighted toward earlier steps). Source: arxiv.org. Saturation forecast: Around November 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-4.5V | 54.5 | #739 |
| 2 | Qwen-VL-Max | 53.3 | #1226 |
| 3 | GPT-5.1 | 52.3 | #225 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=androidintent-complete-instructions · How It Works · Data refreshed daily, snapshot 2026-11-12.