Quality-Aware Self-Distillation Improves GUI Grounding in VLMs

This research introduces a novel quality-aware self-distillation technique designed to significantly improve the performance of vision-language models (VLMs) in graphical user interface (GUI) grounding tasks. GUI grounding requires VLMs to precisely identify small target elements within high-resolution screenshots and predict their exact screen coordinates. The challenge with existing on-policy self-distillation (OPSD) methods for this task is that the quality of teacher signals can degrade when the student model's generated prefix deviates from the correct target coordinate, leading to unreliable supervision. To address this, the proposed method employs two complementary mechanisms: soft correctness-aware gating and teacher-probability scaling. The correctness-aware gate checks if the teacher's current coordinate-token prediction can still lead to the ground-truth box given the student's prefix, down-weighting unreliable signals. Teacher-probability scaling then uses the teacher's confidence to further calibrate the strength of the remaining signals. Experiments across six GUI grounding benchmarks consistently showed performance improvements, indicating that these two mechanisms work together effectively to suppress unreliable supervision and calibrate signal strength.