Fully Decoupled Reinforcement Learning with Adaptive Data Curation for GUI Agents

DART is a reinforcement learning framework for GUI agents that splits training into four independent modules (environment cluster, rollout service, data manager, and trainer) and adds multi-level adaptive data curation across tasks, trajectories, steps, and tokens. The authors report 1.6× GPU utilization, 1.9× training throughput, and 5.5× environment utilization over a non-decoupled baseline. Their DART-GUI-7B model reaches a 43.28% success rate on OSWorld, a 15.76-point gain over its base model.