WinDOM: Self-Family Distillation for Small-Model GUI Grounding
Chengheng Li-Chen (EPFL), Zhiqian Zhou (UPC), Hao Chen (UPC), Nicolas Chauvin (Logitech).
ICML 2026 Workshop on Agents in the Wild (AIWILD).
A 2B vision language GUI grounding model trained on synthetic Windows 11 data, with bounding
boxes read straight off the DOM and no human annotation. A Self-Family Distillation cold start
plus GRPO, treating cold start depth as a hyperparameter, reaches +5.4 OOD mean over the base.
This interactive project page needs JavaScript.
Read the paper on OpenReview.