We introduce CUA-Universe, a scalable pipeline that turns real desktop software into hybrid GUI+CLI environments. App-Forge packages applications into reproducible virtual machines, automatically discovers, wraps, or generates command‑line surfaces, and currently supports 16 programs. Task‑Weave builds hybrid tasks of controllable difficulty from reusable operations over seed files, covering visual inspection, file manipulation, and system configuration scenarios. Path‑Steer guides rollout trajectories along efficient GUI+CLI paths during data collection and verifies the harvested trajectories for subsequent training. Training on this data enables a model to move from inefficient GUI‑only interaction or brittle CLI scripting toward effective coordination of both modalities. A 9B‑parameter model gains +39.3 points on CUA‑Verse while cutting steps by 37% and token usage by 60%; on OSWorld success rate improves by +16.8 points with a 57% step reduction and 44% token reduction; on OSWorld‑MCP score rises by +7.84 points with 27% fewer steps and 30% fewer tokens. These results demonstrate that learning hybrid paths markedly boosts agent success and efficiency. CUA-Universe offers a scalable route toward more capable and efficient computer‑use agents.
Review