CUA-Universe: A Scalable and Dynamic Environment for Hybrid GUI+CLI Agents

TL;DR - CUA-Universe is a scalable pipeline for building environments and training data for computer-use agents that coordinate GUI perception with efficient CLI execution. A trained 9B model substantially improved task success while reducing interaction steps and token use across multiple benchmarks.

• App-Forge converts real desktop software into reproducible virtual machines with discovered, wrapped, or generated command-line interfaces, covering 16 applications.
• Task-Weave synthesizes hybrid GUI+CLI tasks with controllable difficulty, while Path-Steer generates and verifies efficient trajectories for post-training.
• On CUA-Verse, the 9B model gained 39.3 score points while using 37% fewer steps and 60% fewer tokens.
• It also improved OSWorld success rate by 16.8 points and OSWorld-MCP score by 7.84 points, with sizable efficiency gains on both.