CoAdapt-GUI Improves Mobile Agent Adaptation to New Apps

Mobile GUI agents often struggle when deployed to applications they haven't encountered during training, exhibiting brittleness in novel environments. This research addresses the challenge of generalizing to unseen applications with a minimal budget for target interaction and without relying on target demonstrations. The proposed solution is CoAdapt-GUI, a test-time adaptation (TTA) framework that simultaneously adjusts both the structured workflow context and the agent's policy. CoAdapt-GUI achieves this by learning from the agent's own rollouts and rewards within the target application. The workflow context is designed to retain transferable procedures, common failure modes, and verification rules, while explicitly excluding app-specific details from the source training. This separation allows the reusable workflow knowledge to guide the adaptation process without transferring irrelevant interface states. For policy adaptation, a task-context-matched group-relative optimization updates a LoRA adapter on a frozen vision-language model. Evaluations on two unseen-app benchmarks demonstrated significant performance gains, highlighting the effectiveness of combining transfer-constrained workflow context with joint policy adaptation.