@inproceedings{yang-etal-2026-spsr,
title = "{SPSR}: Achieving Superior Large Language Model Layer Pruning Performance by Super-fast Recovery",
author = "Yang, Zhiguo and
Deng, Changjian and
Chen, Qinke and
Zhou, Zijing and
Cheng, Jian",
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.1057/",
pages = "22378--22398",
ISBN = "979-8-89176-462-0",
abstract = "Layer pruning aims to accelerate the inference of Large Language Models on general hardware by removing less important Transformer layers. However, naively deleting layers typically degrades model performance. To reduce this degradation, prior work approximates removed layers with replacement components such as linear layers. These approaches often require large datasets and heavy replacement modules, which increase computational and memory costs. Their time complexity also restricts them to static, consecutive-layer pruning. In this work, we introduce a novel method, Superior layer Pruning performance by Super-fast Recovery (SPSR), which uses Pearson correlation coefficients to quantify a layer{'}s recovery ability and applies a super-fast fine-tuning procedure with a lightweight replacement component. SPSR readily extends to non-consecutive pruning and to dynamic pruning. Experimental results show that SPSR achieves performance competitive with state-of-the-art methods while substantially improving recovery efficiency."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yang-etal-2026-spsr">
<titleInfo>
<title>SPSR: Achieving Superior Large Language Model Layer Pruning Performance by Super-fast Recovery</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zhiguo</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Changjian</namePart>
<namePart type="family">Deng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qinke</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zijing</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jian</namePart>
<namePart type="family">Cheng</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>Layer pruning aims to accelerate the inference of Large Language Models on general hardware by removing less important Transformer layers. However, naively deleting layers typically degrades model performance. To reduce this degradation, prior work approximates removed layers with replacement components such as linear layers. These approaches often require large datasets and heavy replacement modules, which increase computational and memory costs. Their time complexity also restricts them to static, consecutive-layer pruning. In this work, we introduce a novel method, Superior layer Pruning performance by Super-fast Recovery (SPSR), which uses Pearson correlation coefficients to quantify a layer’s recovery ability and applies a super-fast fine-tuning procedure with a lightweight replacement component. SPSR readily extends to non-consecutive pruning and to dynamic pruning. Experimental results show that SPSR achieves performance competitive with state-of-the-art methods while substantially improving recovery efficiency.</abstract>
<identifier type="citekey">yang-etal-2026-spsr</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1057/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>22378</start>
<end>22398</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T SPSR: Achieving Superior Large Language Model Layer Pruning Performance by Super-fast Recovery
%A Yang, Zhiguo
%A Deng, Changjian
%A Chen, Qinke
%A Zhou, Zijing
%A Cheng, Jian
%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 yang-etal-2026-spsr
%X Layer pruning aims to accelerate the inference of Large Language Models on general hardware by removing less important Transformer layers. However, naively deleting layers typically degrades model performance. To reduce this degradation, prior work approximates removed layers with replacement components such as linear layers. These approaches often require large datasets and heavy replacement modules, which increase computational and memory costs. Their time complexity also restricts them to static, consecutive-layer pruning. In this work, we introduce a novel method, Superior layer Pruning performance by Super-fast Recovery (SPSR), which uses Pearson correlation coefficients to quantify a layer’s recovery ability and applies a super-fast fine-tuning procedure with a lightweight replacement component. SPSR readily extends to non-consecutive pruning and to dynamic pruning. Experimental results show that SPSR achieves performance competitive with state-of-the-art methods while substantially improving recovery efficiency.
%U https://aclanthology.org/2026.findings-emnlp.1057/
%P 22378-22398
Markdown (Informal)
[SPSR: Achieving Superior Large Language Model Layer Pruning Performance by Super-fast Recovery](https://aclanthology.org/2026.findings-emnlp.1057/) (Yang et al., Findings 2026)
ACL