The paper introduces CUA-Universe, a pipeline that converts real desktop applications into hybrid GUI+CLI environments for training computer-use agents. It consists of three components: App-Forge adapts applications into reproducible VMs with command-line surfaces, Task-Weave synthesizes diverse hybrid tasks from operations on seed files, and Path-Steer collects efficient hybrid trajectories for post-training. The authors train a 9B model on data generated by this pipeline and test it on several benchmarks.
The trained model shows substantial improvements over baselines: on CUA-Verse its score increases by 39.3 points while reducing steps by 37% and tokens by 60%; on OSWorld success rate improves by 16.8 points with 57% fewer steps and 44% fewer tokens; and on OSWorld-MCP score improves by 7.84 points with 27% fewer steps and 30% fewer tokens. This work is interesting because it addresses the inefficiency of GUI-only agents and provides a scalable method to create hybrid training environments without per-application manual engineering.