HybridCUA: Learning to Orchestrate GUI and CLI for Computer-Use Agents
The paper trains computer-use agents to decide when shell commands beat clicking through an interface.
HybridCUA builds trajectories for GUI-only, CLI-only, and mixed GUI/CLI task execution. The authors use 5K hybrid trajectories plus 3K verified RLVR tasks, then train with supervised fine-tuning and CLI-aware rewards. Their 9B model reports 53.6% accuracy on OSWorld, 14.8 points over its base model, and a 4.0-point gain on WindowsAgentArena. HF Daily Papers' note
HybridCUA builds trajectories for GUI-only, CLI-only, and mixed GUI/CLI task execution. The authors use 5K hybrid trajectories plus 3K verified RLVR tasks, then train with supervised fine-tuning and CLI-aware rewards. Their 9B model reports 53.6% accuracy on OSWorld, 14.8 points over its base model, and a 4.0-point gain on WindowsAgentArena. HF Daily Papers' note
score 5