Current GUI grounding approaches rely heavily on large-scale pixel-level annotations and training-time optimization, which are expensive, inflexible, and difficult to scale to new domains. we observe ...