Script: Graph-Structured and Query-Conditioned Semantic Token Pruning for Multimodal Large Language Models

Zhongyu Yang [email protected] · 2025

Read Script: Graph-Structured and Query-Conditioned Semantic Token Pruning for Multimodal Large Language Models by Zhongyu Yang [email protected] (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-2512-01949.txt.

Description

Multimodal language models turn images and video into long streams of visual tokens, and processing all of them is slow and memory-hungry. Zhongyu Yang's 2025 paper, called Script, prunes those tokens, discarding the redundant ones, while keeping the ones that matter for the user's actual question.

Earlier pruning methods often ignored the query, cutting tokens by generic importance, or leaned on attention scores with known limits. Script instead builds a graph over the tokens and conditions its choices on the query's meaning. Smarter pruning promises faster, lighter multimodal models that still answer with the visual detail the question needs.

Catalog record

CREATORZhongyu Yang [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE118 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-01949

Browse the full library · About the Tower of Babel