AdaGround adds adaptive zoom and a verifier for GUI grounding
Original title: AdaGround: Token-Normalized Adaptive Zooming with Verifier-in-the-Loop Refinement for GUI Grounding
On Qwen3-VL, the training-free method raises accuracy 7.4 to 20.6 points on OSWorld-G and ScreenSpot-Pro.
AdaGround is a training-free GUI grounding framework aimed at small targets on high-resolution screens. Token-Normalized Adaptive Zooming sets magnification, field of view and depth from the model’s own target-size estimate, and a marked-click Yes/No verifier can reject zooms, choose chains, re-ground or abstain. A third component composes spreadsheet cells from headers. The authors say it beats other training-free zoom methods and many trained GUI models on Qwen3-VL, with gains that carry to held-out UI-Vision.
Unconfirmed: Accuracy gains are the authors’ reported benchmark results.
- Zoom is set per sample from the model’s target-size estimate.
- A verifier can reject a zoom or abstain via an existence query.
- Spreadsheet cells are composed from their headers.
- Code is at github.com/bcwang-sjtu/AdaGround.
From the original article
1 min read at arxiv.orgGUI grounding on high-resolution screens fails mostly on small targets, and zooming in is a common training-free remedy. However, most existing zooming pipelines do not match the zoom to the target's size, accept the zoomed answer unconditionally and never revisit a wrong answer.
By Bingchao Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.