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

PDF

Dannong Xu , Wei Pang ·

YOU? · · 2025 · Open Access · · DOI: https://doi.org/10.48550/arxiv.2512.01949

The rapid growth of visual tokens in multimodal large language models (MLLMs) leads to excessive memory consumption and inference latency, especially when handling high-resolution images and videos. Token pruning is a technique used to mitigate this issue by removing redundancy, but existing methods often ignore relevance to the user query or suffer from the limitations of attention mechanisms, reducing their adaptability and effectiveness. To address these challenges, we propose Script, a plug-and-play pruning method that requires no retraining and generalizes across diverse MLLMs. Script comprises two modules: a graph-structured pruning module that removes visually redundant tokens, and a query-conditioned semantic pruning module that preserves query-relevant visual information. Together, they enhance performance on multimodal tasks. Experiments on fourteen benchmarks across image and video understanding tasks show that Script consistently achieves higher model efficiency and predictive accuracy compared to existing pruning methods. On LLaVA-NeXT-7B, it achieves up to 6.8x prefill speedup and 10x FLOP reduction, while retaining 96.88% of the original performance.

Related Topics

Proto-Sinaitic Script

Concepts

No concepts available.

Metadata

Type: preprint
Landing Page: http://arxiv.org/abs/2512.01949
PDF: https://arxiv.org/pdf/2512.01949
OA Status: green
OpenAlex ID: https://openalex.org/W4416968892

All OpenAlex metadata

Raw OpenAlex JSON

OpenAlex ID: https://openalex.org/W4416968892

Canonical identifier for this work in OpenAlex
DOI: https://doi.org/10.48550/arxiv.2512.01949

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

Work title
Type: preprint

OpenAlex work type
Publication year: 2025

Year of publication
Publication date: 2025-12-01

Full publication date if available
Authors: Dannong Xu, Wei Pang

List of authors in order
Landing page: https://arxiv.org/abs/2512.01949

Publisher landing page
PDF URL: https://arxiv.org/pdf/2512.01949

Direct link to full text PDF
Open access: Yes

Whether a free full text is available
OA status: green

Open access status per OpenAlex
OA URL: https://arxiv.org/pdf/2512.01949

Direct OA link when available
Cited by: 0

Total citation count in OpenAlex

Full payload

id	https://openalex.org/W4416968892
doi	https://doi.org/10.48550/arxiv.2512.01949
ids.doi	https://doi.org/10.48550/arxiv.2512.01949
ids.openalex	https://openalex.org/W4416968892
fwci
type	preprint
title	Script: Graph-Structured and Query-Conditioned Semantic Token Pruning for Multimodal Large Language Models
biblio.issue
biblio.volume
biblio.last_page
biblio.first_page
is_xpac	False
apc_list
apc_paid
language
locations[0].id	pmh:oai:arXiv.org:2512.01949
locations[0].is_oa	True
locations[0].source.id	https://openalex.org/S4306400194
locations[0].source.issn
locations[0].source.type	repository
locations[0].source.is_oa	True
locations[0].source.issn_l
locations[0].source.is_core	False
locations[0].source.is_in_doaj	False
locations[0].source.display_name	arXiv (Cornell University)
locations[0].source.host_organization	https://openalex.org/I205783295
locations[0].source.host_organization_name	Cornell University
locations[0].source.host_organization_lineage	https://openalex.org/I205783295
locations[0].license	cc-by
locations[0].pdf_url	https://arxiv.org/pdf/2512.01949
locations[0].version	submittedVersion
locations[0].raw_type	text
locations[0].license_id	https://openalex.org/licenses/cc-by
locations[0].is_accepted	False
locations[0].is_published	False
locations[0].raw_source_name
locations[0].landing_page_url	http://arxiv.org/abs/2512.01949
locations[1].id	doi:10.48550/arxiv.2512.01949
locations[1].is_oa	True
locations[1].source.id	https://openalex.org/S4306400194
locations[1].source.issn
locations[1].source.type	repository
locations[1].source.is_oa	True
locations[1].source.issn_l
locations[1].source.is_core	False
locations[1].source.is_in_doaj	False
locations[1].source.display_name	arXiv (Cornell University)
locations[1].source.host_organization	https://openalex.org/I205783295
locations[1].source.host_organization_name	Cornell University
locations[1].source.host_organization_lineage	https://openalex.org/I205783295
locations[1].license	cc-by
locations[1].pdf_url
locations[1].version
locations[1].raw_type	article
locations[1].license_id	https://openalex.org/licenses/cc-by
locations[1].is_accepted	False
locations[1].is_published
locations[1].raw_source_name
locations[1].landing_page_url	https://doi.org/10.48550/arxiv.2512.01949
indexed_in	arxiv, datacite
authorships[0].author.id	https://openalex.org/A5111005645
authorships[0].author.orcid
authorships[0].author.display_name	Dannong Xu
authorships[0].author_position	last
authorships[0].raw_author_name	Xu, Dannong
authorships[0].is_corresponding	False
authorships[1].author.id	https://openalex.org/A5080597597
authorships[1].author.orcid
authorships[1].author.display_name	Wei Pang
authorships[1].author_position	middle
authorships[1].raw_author_name	Pang, Wei
authorships[1].is_corresponding	False
has_content.pdf	True
has_content.grobid_xml	False
is_paratext	False
open_access.is_oa	True
open_access.oa_url	https://arxiv.org/pdf/2512.01949
open_access.oa_status	green
open_access.any_repository_has_fulltext	False
created_date	2025-12-03T00:00:00
display_name	Script: Graph-Structured and Query-Conditioned Semantic Token Pruning for Multimodal Large Language Models
has_fulltext	False
is_retracted	False
updated_date	2025-12-03T23:12:59.920255
primary_topic
cited_by_count	0
locations_count	2
best_oa_location.id	pmh:oai:arXiv.org:2512.01949
best_oa_location.is_oa	True
best_oa_location.source.id	https://openalex.org/S4306400194
best_oa_location.source.issn
best_oa_location.source.type	repository
best_oa_location.source.is_oa	True
best_oa_location.source.issn_l
best_oa_location.source.is_core	False
best_oa_location.source.is_in_doaj	False
best_oa_location.source.display_name	arXiv (Cornell University)
best_oa_location.source.host_organization	https://openalex.org/I205783295
best_oa_location.source.host_organization_name	Cornell University
best_oa_location.source.host_organization_lineage	https://openalex.org/I205783295
best_oa_location.license	cc-by
best_oa_location.pdf_url	https://arxiv.org/pdf/2512.01949
best_oa_location.version	submittedVersion
best_oa_location.raw_type	text
best_oa_location.license_id	https://openalex.org/licenses/cc-by
best_oa_location.is_accepted	False
best_oa_location.is_published	False
best_oa_location.raw_source_name
best_oa_location.landing_page_url	http://arxiv.org/abs/2512.01949
primary_location.id	pmh:oai:arXiv.org:2512.01949
primary_location.is_oa	True
primary_location.source.id	https://openalex.org/S4306400194
primary_location.source.issn
primary_location.source.type	repository
primary_location.source.is_oa	True
primary_location.source.issn_l
primary_location.source.is_core	False
primary_location.source.is_in_doaj	False
primary_location.source.display_name	arXiv (Cornell University)
primary_location.source.host_organization	https://openalex.org/I205783295
primary_location.source.host_organization_name	Cornell University
primary_location.source.host_organization_lineage	https://openalex.org/I205783295
primary_location.license	cc-by
primary_location.pdf_url	https://arxiv.org/pdf/2512.01949
primary_location.version	submittedVersion
primary_location.raw_type	text
primary_location.license_id	https://openalex.org/licenses/cc-by
primary_location.is_accepted	False
primary_location.is_published	False
primary_location.raw_source_name
primary_location.landing_page_url	http://arxiv.org/abs/2512.01949
publication_date	2025-12-01
publication_year	2025
referenced_works_count	0
abstract_inverted_index.a	30, 70, 87, 97
abstract_inverted_index.On	140
abstract_inverted_index.To	63
abstract_inverted_index.by	37
abstract_inverted_index.in	6
abstract_inverted_index.is	29
abstract_inverted_index.it	142
abstract_inverted_index.no	76
abstract_inverted_index.of	3, 55, 156
abstract_inverted_index.on	111, 115
abstract_inverted_index.or	50
abstract_inverted_index.to	13, 33, 46, 136, 145
abstract_inverted_index.up	144
abstract_inverted_index.we	67
abstract_inverted_index.10x	150
abstract_inverted_index.The	0
abstract_inverted_index.and	17, 25, 61, 78, 96, 120, 132, 149
abstract_inverted_index.but	40
abstract_inverted_index.the	47, 53, 157
abstract_inverted_index.two	85
abstract_inverted_index.6.8x	146
abstract_inverted_index.FLOP	151
abstract_inverted_index.from	52
abstract_inverted_index.show	124
abstract_inverted_index.that	74, 91, 102, 125
abstract_inverted_index.they	108
abstract_inverted_index.this	35
abstract_inverted_index.used	32
abstract_inverted_index.user	48
abstract_inverted_index.when	21
abstract_inverted_index.Token	27
abstract_inverted_index.image	119
abstract_inverted_index.issue	36
abstract_inverted_index.large	8
abstract_inverted_index.leads	12
abstract_inverted_index.model	130
abstract_inverted_index.often	43
abstract_inverted_index.query	49
abstract_inverted_index.rapid	1
abstract_inverted_index.tasks	123
abstract_inverted_index.their	59
abstract_inverted_index.these	65
abstract_inverted_index.video	121
abstract_inverted_index.while	153
abstract_inverted_index.96.88%	155
abstract_inverted_index.MLLMs.	82
abstract_inverted_index.Script	83, 126
abstract_inverted_index.across	80, 118
abstract_inverted_index.growth	2
abstract_inverted_index.higher	129
abstract_inverted_index.ignore	44
abstract_inverted_index.images	24
abstract_inverted_index.memory	15
abstract_inverted_index.method	73
abstract_inverted_index.models	10
abstract_inverted_index.module	90, 101
abstract_inverted_index.suffer	51
abstract_inverted_index.tasks.	113
abstract_inverted_index.tokens	5
abstract_inverted_index.visual	4, 105
abstract_inverted_index.(MLLMs)	11
abstract_inverted_index.Script,	69
abstract_inverted_index.address	64
abstract_inverted_index.diverse	81
abstract_inverted_index.enhance	109
abstract_inverted_index.methods	42
abstract_inverted_index.prefill	147
abstract_inverted_index.propose	68
abstract_inverted_index.pruning	28, 72, 89, 100, 138
abstract_inverted_index.removes	92
abstract_inverted_index.speedup	148
abstract_inverted_index.tokens,	95
abstract_inverted_index.videos.	26
abstract_inverted_index.accuracy	134
abstract_inverted_index.achieves	128, 143
abstract_inverted_index.compared	135
abstract_inverted_index.existing	41, 137
abstract_inverted_index.fourteen	116
abstract_inverted_index.handling	22
abstract_inverted_index.language	9
abstract_inverted_index.latency,	19
abstract_inverted_index.methods.	139
abstract_inverted_index.mitigate	34
abstract_inverted_index.modules:	86
abstract_inverted_index.original	158
abstract_inverted_index.reducing	58
abstract_inverted_index.removing	38
abstract_inverted_index.requires	75
abstract_inverted_index.semantic	99
abstract_inverted_index.visually	93
abstract_inverted_index.Together,	107
abstract_inverted_index.attention	56
abstract_inverted_index.comprises	84
abstract_inverted_index.excessive	14
abstract_inverted_index.inference	18
abstract_inverted_index.preserves	103
abstract_inverted_index.redundant	94
abstract_inverted_index.relevance	45
abstract_inverted_index.retaining	154
abstract_inverted_index.technique	31
abstract_inverted_index.benchmarks	117
abstract_inverted_index.efficiency	131
abstract_inverted_index.especially	20
abstract_inverted_index.multimodal	7, 112
abstract_inverted_index.predictive	133
abstract_inverted_index.reduction,	152
abstract_inverted_index.retraining	77
abstract_inverted_index.Experiments	114
abstract_inverted_index.challenges,	66
abstract_inverted_index.consumption	16
abstract_inverted_index.generalizes	79
abstract_inverted_index.limitations	54
abstract_inverted_index.mechanisms,	57
abstract_inverted_index.performance	110
abstract_inverted_index.redundancy,	39
abstract_inverted_index.adaptability	60
abstract_inverted_index.consistently	127
abstract_inverted_index.information.	106
abstract_inverted_index.performance.	159
abstract_inverted_index.plug-and-play	71
abstract_inverted_index.understanding	122
abstract_inverted_index.LLaVA-NeXT-7B,	141
abstract_inverted_index.effectiveness.	62
abstract_inverted_index.query-relevant	104
abstract_inverted_index.high-resolution	23
abstract_inverted_index.graph-structured	88
abstract_inverted_index.query-conditioned	98
cited_by_percentile_year
countries_distinct_count	0
institutions_distinct_count	2
citation_normalized_percentile