Qwen-UI-Agent Hits 97.5% on AndroidDaily and 79.5% on OSWorld-Verified With 100+ Turn Online RL
Alibaba Tongyi Lab's Qwen-UI-Agent technical report (arXiv 2607.28227, July 30) describes a foundation GUI agent spanning mobile, computer-use, web, and DeepSearch, with a unified action space that interleaves GUI operations with CLI execution and emits batched actions in a single model turn. Reported results are 82.1% on MobileWorld, 92.2% on MobileWorld-Real, 97.5% on AndroidDaily, and 79.5% on OSWorld-Verified, competitive with Opus 4.8, Gemini 3.1 Pro, and GPT-5.6 Sol. Training used online RL on trajectories exceeding 100 turns with over 10,000 concurrent environments plus an AutoResearch-style data flywheel where agents build tasks, diagnose failures, and plan the next iteration; it is currently the #1 trending paper on Hugging Face with 242 upvotes.
Source
↳ Follow the thread