TreeCUA: Efficiently Scaling GUI Automation with Tree-Structured Verifiable Evolution

Effectively scaling GUI automation is essential for computer-use agents (CUAs); however, existing work primarily focuses on scaling GUI grounding rather than the more crucial GUI planning, which requires more sophisticated data collection. In reality, the exploration process of a CUA across apps/desktops/web pages typically follows a tree structure, with earlier functional entry points often being explored more frequently. In this work, we find that organizing large-scale GUI trajectories into tree structures can effectively eliminate redundant exploration costs, while each branch node also provides key reasoning evidence for distinguishing adjacent trajectories. Therefore, we propose TreeCUA to efficiently scale GUI automation with tree-structured verifiable evolution. We propose a multi-agent collaborative framework to explore the environment, verify actions, summarize trajectories, and evaluate quality to generate high-quality and scalable GUI trajectories. To improve efficiency, we devise a novel tree-based topology to store and replay duplicate exploration nodes, and design an adaptive exploration algorithm to balance the depth (i.e., trajectory difficulty) and breadth (i.e., trajectory diversity). Moreover, we develop world knowledge guidance and global memory backtracking to avoid low-quality generation. Finally, we naturally extend and propose the TreeCUA-DPO method from abundant tree node information, improving GUI planning capability by referring to the branch information of adjacent trajectories. Experimental results show that TreeCUA and TreeCUA-DPO offer significant improvements, and out-of-domain (OOD) studies further demonstrate strong generalization. All trajectory node information and code will be open-sourced.

Imagine teaching a robot assistant to navigate your computer — opening apps, clicking menus, and completing tasks just like a human would. Collecting the training data for such agents traditionally requires human annotators to record their every click, which is both expensive and difficult to scale up. Recent efforts have turned to automated data collection, but these approaches have their own problems: they record every task independently from scratch, wasting effort by repeating the same early steps over and over. They also tend to explore only the most obvious, common actions — missing the richer, less-traveled corners of an application. We noticed that computer navigation naturally follows a tree shape — many tasks share the same starting clicks before branching off in different directions. So instead of recording each task from scratch, we store shared steps once and reuse them across thousands of tasks, dramatically cutting collection costs. We also guide the exploration with official software documentation to uncover rare but useful functionalities, and deploy a team of AI agents to verify each action, filter low-quality data, and keep explorations diverse. The result is TreeCUA, a computer-use agent trained on over 150,000 high-quality interaction examples collected at a fraction of the usual cost. It outperforms existing open-source agents on standard benchmarks and generalizes well to software it has never seen before — a meaningful step toward AI assistants that can reliably handle real-world computer tasks.