AdaGround framework enhances GUI grounding accuracy with adaptive zooming

Researchers have developed AdaGround, a novel framework designed to improve GUI grounding accuracy on high-resolution screens, particularly for small targets. The system employs Token-Normalized Adaptive Zooming to dynamically adjust magnification based on the model's target size estimation. It also incorporates a Verifier-in-the-Loop Refinement mechanism that uses a click-based verifier to accept or reject zooms and trigger re-grounding. AdaGround has demonstrated significant accuracy improvements over existing methods when applied to Qwen3-VL models, outperforming both training-free zooming techniques and some trained GUI models on benchmark datasets. AI

IMPACT Improves accuracy for AI models in understanding and interacting with graphical user interfaces, particularly on high-resolution displays.

RANK_REASON This is a research paper detailing a new framework for GUI grounding. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• AdaGround
• arXiv
• Compositional Structure Grounding
• Hugging Face
• OSWorld-G
• Qwen3 VL
• ScreenSpot-Pro
• Token-Normalized Adaptive Zooming
• UI-Vision
• Verifier-in-the-Loop Refinement

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →