Cua's RL-Trained Decision Model Hits 92.9% on GUI-360, Open-Sourced

Cua released Cua-S1-4B-0.2, which it calls the first multimodal decision model trained with RLOO, a reinforcement-learning method, on live computer-use tasks with task-completion rewards. At each step the model receives the screen state, the task goal and a fixed list of candidate actions, and returns one. Training first teaches that output format with supervised learning, then runs the model in live task environments where only a completed task earns reward.

Cua reports 92.9% versus 60.1% for an untrained djev baseline on a frozen 168-task multimodal split of GUI-360, an existing benchmark of Windows office application tasks, with the same inputs and scoring and no accessibility tree. Its other results, 17/18 versus 16/18 text episodes and 13/18 versus 12/18 multimodal, come from Cua's own held-out tasks and differ by one episode. Cua released Apache 2.0 adapters and training code, so others can check the numbers, which remain Cua's own.