HybridCUA: Learning to Orchestrate GUI and CLI for Computer-Use Agents

This work introduces HybridCUA, a framework that combines graphical user interfaces (GUI) with command-line interfaces (CLI) to enhance computer-use agents. The authors construct HybridCUA-8K containing 5K hybrid trajectories and 3K verified RLVR tasks through a dedicated data pipeline. Using a two-stage training approach with supervised fine-tuning and reinforcement learning via CLI-aware rewards, the method successfully guides agents to leverage CLI efficiency alongside GUI generality.

HybridCUA-9B achieves 53.6% accuracy on OSWorld, improving over the base model by 14.8 percentage points, and improves performance on WindowsAgentArena by 4.0 percentage points.