PRO-CUA: Process-Reward Optimization for Computer Use Agents

In three linesPRO-CUA introduces a process-reward optimization framework for training computer use agents (CUAs). The method decouples live environment interaction from policy optimization through iterative step-level reinforcement learning, using a process reward model (PRM) to provide dense feedback signals without relying on expert trajectories or golden answers.

Summary generated by Claude — human-verified