The paper proposes a two-stage adaptive training method for multi-step GUI agents that cuts the number of environment interactions needed during reinforcement learning.
Where to learn: The method builds reusable checkpoints from successful expert trajectories, picks uncertain decision points using action disagreement, and runs one-step RL at those pivots.
How far to roll out: Stage II expands training from Local to Suffix to Root according to task progress, so already-mastered prefixes don't have to be replayed.
Results: On MobileGym and CUA-Gym-Hub, the authors report 56-61% less environment interaction than Vanilla GRPO, with better task performance.
Sign in to comment.
TL;DR of the paper, generated automatically.
The paper proposes a two-stage adaptive training method for multi-step GUI agents that cuts the number of environment interactions needed during reinforcement learning.