ReVision: Scaling Computer-Use Agents via Temporal Visual Redundancy Reduction

Amirhossein Abaskohi, Yuhang He, Peter West, Giuseppe Carenini, Pranit Chawla, Vibhav Vineet, [0.5em] University of British Columbia, Microsoft Research · 2026

Read ReVision: Scaling Computer-Use Agents via Temporal Visual Redundancy Reduction by Amirhossein Abaskohi, Yuhang He, Peter West, Giuseppe Carenini, Pranit Chawla, Vibhav Vineet, [0.5em] University of British Columbia, Microsoft Research (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-2605-11212.txt.

Description

Amirhossein Abaskohi and five co-authors present ReVision, a method for scaling computer-use agents by cutting temporal visual redundancy. These agents read screen interfaces through screenshots, but each screenshot becomes many visual tokens, and token cost grows fast as sessions lengthen.

It trains multimodal language models on trajectories where extra visual patches are removed by a learned patch selector. It compares patch vectors across back-to-back screenshots while keeping spatial structure the model needs. Across OSWorld, WebTailBench, and AgentNetBench with five history screenshots on Qwen2.5-VL-7B, it cuts token use by about 46 percent while lifting success rate by 3 percent over the no-drop baseline. Agent builders get longer sessions at lower token cost, and results keep rising as more past context is added.

Catalog record

CREATORAmirhossein Abaskohi, Yuhang He, Peter West, Giuseppe Carenini, Pranit Chawla, Vibhav Vineet, [0.5em] University of British Columbia, Microsoft Research
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE61342
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.11212

Browse the full library · About the Tower of Babel