@inproceedings{wan-etal-2026-dspp,
title = "{DSPP}: Dataset-Specific Progressive Pruning for Efficient and Generalizable {LLM} Deployment",
author = "Wan, Lichuan and
Jiang, Huiqiang and
Liu, Wenxuan and
Yang, Hua",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.2391/",
pages = "51383--51396",
ISBN = "979-8-89176-462-0",
abstract = "Large language models (LLMs) have demonstrated exceptional performance across various domains. However, their increasing sizes pose significant deployment challenges, particularly in resource-constrained scenarios. While network pruning provides an effective compression strategy, existing approaches either depend on calibration datasets{---}leading to poor generalization{---}or require training auxiliary predictors, incurring significant computational overhead. In this paper, we propose DSPP ($\textbf{D}$ataset-$\textbf{S}$pecific $\textbf{P}$rogressive $\textbf{P}$runing), a novel training-free structured pruning method that dynamically updates the importance scores of neurons during inference to identify and retain dataset-specific neurons. By employing a progressive pruning strategy, DSPP can further reduce computational overhead. Experimental results show that DSPP demonstrates remarkable generalization capabilities across diverse datasets and models, while achieving a $1.36\times$ speedup at 50{\%} sparsity level."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wan-etal-2026-dspp">
<titleInfo>
<title>DSPP: Dataset-Specific Progressive Pruning for Efficient and Generalizable LLM Deployment</title>
</titleInfo>
<name type="personal">
<namePart type="given">Lichuan</namePart>
<namePart type="family">Wan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Huiqiang</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenxuan</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hua</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Large language models (LLMs) have demonstrated exceptional performance across various domains. However, their increasing sizes pose significant deployment challenges, particularly in resource-constrained scenarios. While network pruning provides an effective compression strategy, existing approaches either depend on calibration datasets—leading to poor generalization—or require training auxiliary predictors, incurring significant computational overhead. In this paper, we propose DSPP (Dataset-Specific Progressive Pruning), a novel training-free structured pruning method that dynamically updates the importance scores of neurons during inference to identify and retain dataset-specific neurons. By employing a progressive pruning strategy, DSPP can further reduce computational overhead. Experimental results show that DSPP demonstrates remarkable generalization capabilities across diverse datasets and models, while achieving a 1.36\times speedup at 50% sparsity level.</abstract>
<identifier type="citekey">wan-etal-2026-dspp</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2391/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>51383</start>
<end>51396</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T DSPP: Dataset-Specific Progressive Pruning for Efficient and Generalizable LLM Deployment
%A Wan, Lichuan
%A Jiang, Huiqiang
%A Liu, Wenxuan
%A Yang, Hua
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F wan-etal-2026-dspp
%X Large language models (LLMs) have demonstrated exceptional performance across various domains. However, their increasing sizes pose significant deployment challenges, particularly in resource-constrained scenarios. While network pruning provides an effective compression strategy, existing approaches either depend on calibration datasets—leading to poor generalization—or require training auxiliary predictors, incurring significant computational overhead. In this paper, we propose DSPP (Dataset-Specific Progressive Pruning), a novel training-free structured pruning method that dynamically updates the importance scores of neurons during inference to identify and retain dataset-specific neurons. By employing a progressive pruning strategy, DSPP can further reduce computational overhead. Experimental results show that DSPP demonstrates remarkable generalization capabilities across diverse datasets and models, while achieving a 1.36\times speedup at 50% sparsity level.
%U https://aclanthology.org/2026.findings-emnlp.2391/
%P 51383-51396
Markdown (Informal)
[DSPP: Dataset-Specific Progressive Pruning for Efficient and Generalizable LLM Deployment](https://aclanthology.org/2026.findings-emnlp.2391/) (Wan et al., Findings 2026)
ACL