AI 代理
Qwen-UI-Agent Unifies Mobile, Desktop, and Browser Operations, Extending Long-Trajectory Training Beyond 100 Steps
The Qwen team has introduced a unified GUI agent spanning mobile, desktop, web, and DeepSearch environments, with on-screen interactions and CLI commands sharing the same action space. Its technical report claims a score of 79.5% on OSWorld-Verified, but the model weights, training data, and complete trajectories have not yet been released.
The key idea behind Qwen-UI-Agent is not to add yet another mouse-clicking model, but to consolidate previously fragmented mobile operations, desktop software, browser tasks, and DeepSearch into a single foundation agent. Within the same trajectory, the model can interleave GUI actions with CLI commands. It can also generate a batch of actions at once, reducing the latency caused by having to invoke the model again for every click. The system additionally uses a lightweight harness to preserve state across devices and allows the agent to proactively launch services rather than merely respond passively to commands.
On the training side, it uses an AutoResearch-like data flywheel: the agent creates its own tasks and sandboxes, diagnoses failures, and uses those findings to plan the next round of data generation. Online reinforcement learning can handle trajectories exceeding 100 turns, while the team says its rollout infrastructure can run more than 10,000 environments concurrently. This design suggests that the scaling bottleneck for GUI agents is shifting away from single-step visual grounding toward resettable environments, long-horizon credit assignment, and the recycling of failure data.
The report lists scores of 82.1% on MobileWorld, 92.2% on MobileWorld-Real, 97.5% on AndroidDaily, 73.6% on WebArena, and 79.5% on OSWorld-Verified. For the newer OSWorld-v2, however, it reports only a partial-progress score of 40.0%. Engineering teams should pay close attention to evaluation methodology when interpreting these figures: OSWorld execution results can be affected by virtual machine images, account and OAuth configuration, website anti-bot measures, and the agent’s maximum step limit. The official leaderboard also requires implementation submissions or monitored trajectories for verification.
The biggest limitation is that the technical report has arrived before the assets needed for reproducibility. The paper does not provide the model weights, complete training dataset, deployment code, or per-task trajectories, so comparisons with closed frontier models remain based on the team’s internal evaluations. Key issues to watch next include whether the model will be released, how batched actions are interrupted when the on-screen state changes, and whether cross-GUI/CLI permissions include fine-grained isolation and human-confirmation mechanisms.