AQuaUI: Visual Token Reduction for GUI Agents with Adaptive Quadtrees

Yuankai Li, Tinghui Zhu, Ha Min Son, Zhe Zhao, Xin Liu, Muhao Chen · 2026

Read AQuaUI: Visual Token Reduction for GUI Agents with Adaptive Quadtrees by Yuankai Li, Tinghui Zhu, Ha Min Son, Zhe Zhao, Xin Liu, Muhao Chen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-19260.txt.

Description

Yuankai Li and colleagues cut the visual cost of GUI agents, which feed high-resolution screenshots to large multimodal models. Screenshots waste tokens on blank regions while text and icons still need fine detail, and older fixes need retraining or ignore screen layout. Their AQuaUI builds an adaptive quadtree per screenshot, keeping one merged token per leaf and preserving positions, then reuses past quadtrees for steadier work. On a leading GUI model it runs up to 13.22 percent faster with 29.52 percent fewer tokens while keeping 99.06 percent of full-token performance.

Catalog record

CREATORYuankai Li, Tinghui Zhu, Ha Min Son, Zhe Zhao, Xin Liu, Muhao Chen
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46641
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.19260

Browse the full library · About the Tower of Babel