Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents
Qwen reports a GUI agent trained for real-device, cross-platform workflows, with mobile benchmark scores leading its evaluation claims.
The paper presents Qwen-UI-Agent across mobile, computer-use, web, and DeepSearch settings. It combines GUI actions with CLI execution, can batch actions in one model turn, and uses agent-generated tasks and failure diagnosis to drive further training. The authors say online RL supports trajectories beyond 100 turns, scaled across more than 10,000 concurrent environments. Reported results include 82.1% on MobileWorld, 92.2% on MobileWorld-Real, and 97.5% on AndroidDaily, with competitive scores on OSWorld, WebArena, and ScreenSpot-Pro. HF Daily Papers' note
The paper presents Qwen-UI-Agent across mobile, computer-use, web, and DeepSearch settings. It combines GUI actions with CLI execution, can batch actions in one model turn, and uses agent-generated tasks and failure diagnosis to drive further training. The authors say online RL supports trajectories beyond 100 turns, scaled across more than 10,000 concurrent environments. Reported results include 82.1% on MobileWorld, 92.2% on MobileWorld-Real, and 97.5% on AndroidDaily, with competitive scores on OSWorld, WebArena, and ScreenSpot-Pro. HF Daily Papers' note
score 6