Reproduce “AdaGround: Token-Normalized Adaptive Zooming with Verifier-in-the-Loop Refinement for GUI Grounding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bingchao Wang, Yongcheng Jing, Chengyu Wang, Jun Huang, Jie Yang, Wei Liu, Dacheng Tao
AdaGround refines GUI grounding by estimating target size, setting a zoom, and checking candidate clicks with a frozen model. On Qwen3-VL-32B-Instruct, it reaches 75.9% point-in-box accuracy on ScreenSpot-Pro, versus 55.3% for single-pass grounding. On Qwen3-VL-8B, verifier-based arbitration raises accuracy from 63.5% to 65.6%. Results cover two Qwen3-VL backbones and four benchmarks; the method uses more computation than a fixed-crop baseline.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.