arXiv: AdaZoom-GUI — Adaptive Zoom-Based Grounding Reduces GUI Agent Localization Failures on Dense UIs

arXiv·low signal

AdaZoom-GUI introduces instruction-guided adaptive zooming for GUI grounding in vision-language models, dynamically focusing on relevant UI regions before element localization rather than operating at fixed resolution. The method directly addresses the failure mode where small or densely packed UI elements (buttons, form fields, dropdowns) fall below the effective resolution of standard GUI agent vision pipelines. Relevant for browser automation and computer-use agent deployments where precise element identification is a primary reliability bottleneck.