Why 1 + 1 < 1 in Visual Token Pruning: Beyond Naïve Integration via Multi-Objective Balanced Covering

Yangfu Li ♣ {}^{\,\clubsuit} , Hongjian Zhan ♣ , ♡ , {}^{\,\clubsuit,\,\heartsuit,\,} , Tianyi Chen ♠ , Qi Liu ♣ {}^{\,\clubsuit} , Yu-jie Xiong ♢ , Yue Lu ♣ {}^{\,\clubsuit} ♣ East China Normal University, ♠ Shanghai Jiao Tong University ♡ Chongqing Institute of East China Normal University ♢ Shang · 2025

Read Why 1 + 1 < 1 in Visual Token Pruning: Beyond Naïve Integration via Multi-Objective Balanced Covering by Yangfu Li ♣ {}^{\,\clubsuit} , Hongjian Zhan ♣ , ♡ , {}^{\,\clubsuit,\,\heartsuit,\,} , Tianyi Chen ♠ , Qi Liu ♣ {}^{\,\clubsuit} , Yu-jie Xiong ♢ , Yue Lu ♣ {}^{\,\clubsuit} ♣ East China Normal University, ♠ Shanghai Jiao Tong University ♡ Chongqing Institute of East China Normal University ♢ Shang (2025) 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-2505-10118.txt.

Description

Yangfu Li and co-authors rethink visual token pruning for image language models in a 2025 paper, based on the head byline. Pruning cuts the image tokens a model must process, but fixed rules balance two goals badly, staying true to the prompt and keeping the picture intact. Their error bound, the first of its kind, ties both goals into one formula based on covering distance.

Covering theory then exposes a built in trade off and its best reachable levels under a token budget. Their method turns pruning into a two goal covering problem solved by trading radii greedily. It kept nearly all model quality on a tenth of the tokens and sped a larger model clearly.

Catalog record

CREATORYangfu Li ♣ {}^{\,\clubsuit} , Hongjian Zhan ♣ , ♡ , {}^{\,\clubsuit,\,\heartsuit,\,} , Tianyi Chen ♠ , Qi Liu ♣ {}^{\,\clubsuit} , Yu-jie Xiong ♢ , Yue Lu ♣ {}^{\,\clubsuit} ♣ East China Normal University, ♠ Shanghai Jiao Tong University ♡ Chongqing Institute of East China Normal University ♢ Shang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE109 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-10118

Browse the full library · About the Tower of Babel