Submitted by Rui Yang 15 GUI-Libra: Training Native GUI Agents to Reason and Act with Action-aware Supervision and Partially Verifiable RL UIUC ScaleML Lab 27 3