CUA-Gym: Scaling Verifiable Training Environments and Tasks for Computer-Use Agents

Researchers introduced CUA-Gym, a scalable synthetic pipeline for training computer-use agents with verifiable rewards.

The project produced 32,112 verified training tuples across 110 environments, including task instructions, environment states, and reward functions.

Models trained on CUA-Gym improved performance on OSWorld-Verified and also transferred better to WebArena.

The team plans to release the pipeline, environments, dataset, and models through CUA-Gym-Hub.