@inproceedings{wu-etal-2026-subspace,
title = "Subspace Alignment Score: Locating Domain-Relevant Important Parameters in Large Language Models",
author = "Wu, Le and
Wu, Jiayi and
Xiong, Jixiang and
Mo, Mengfa and
Xin, Jingmin",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.1504/",
pages = "34347--34366",
ISBN = "979-8-89176-459-0",
abstract = "Although large language models (LLMs) possess powerful general capabilities, their insufficient domain expertise limits their application in high-reliability scenarios. The sparse activation patterns of neurons in LLM feed-forward layers suggest that structured fine-tuning strategies for domain adaptation can be constructed by identifying important parameters, thereby enhancing domain-specific performance while preserving general capabilities. However, due to high computational costs or weak scalability, current studies struggle to support large-scale and fine-grained domain-relevant parameter analysis. To address this, we investigate the distribution patterns and hierarchical evolution patterns of domain knowledge within the hidden state space of LLMs, and propose a novel parameter evaluation method: Subspace Alignment Score. Intervention experiments across three models and six domains illustrate that, compared with baseline methods, our method demonstrates superior performance in locating domain-relevant important parameters. Furthermore, structured fine-tuning strategies based on the important parameters achieve performance comparable to full fine-tuning while substantially reducing training costs, which is helpful for understanding efficient and interpretable domain adaptation. Our code is available at https://github.com/lewu0723/subspace-alignment-score."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wu-etal-2026-subspace">
<titleInfo>
<title>Subspace Alignment Score: Locating Domain-Relevant Important Parameters in Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Le</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiayi</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jixiang</namePart>
<namePart type="family">Xiong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mengfa</namePart>
<namePart type="family">Mo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jingmin</namePart>
<namePart type="family">Xin</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Although large language models (LLMs) possess powerful general capabilities, their insufficient domain expertise limits their application in high-reliability scenarios. The sparse activation patterns of neurons in LLM feed-forward layers suggest that structured fine-tuning strategies for domain adaptation can be constructed by identifying important parameters, thereby enhancing domain-specific performance while preserving general capabilities. However, due to high computational costs or weak scalability, current studies struggle to support large-scale and fine-grained domain-relevant parameter analysis. To address this, we investigate the distribution patterns and hierarchical evolution patterns of domain knowledge within the hidden state space of LLMs, and propose a novel parameter evaluation method: Subspace Alignment Score. Intervention experiments across three models and six domains illustrate that, compared with baseline methods, our method demonstrates superior performance in locating domain-relevant important parameters. Furthermore, structured fine-tuning strategies based on the important parameters achieve performance comparable to full fine-tuning while substantially reducing training costs, which is helpful for understanding efficient and interpretable domain adaptation. Our code is available at https://github.com/lewu0723/subspace-alignment-score.</abstract>
<identifier type="citekey">wu-etal-2026-subspace</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1504/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>34347</start>
<end>34366</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Subspace Alignment Score: Locating Domain-Relevant Important Parameters in Large Language Models
%A Wu, Le
%A Wu, Jiayi
%A Xiong, Jixiang
%A Mo, Mengfa
%A Xin, Jingmin
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F wu-etal-2026-subspace
%X Although large language models (LLMs) possess powerful general capabilities, their insufficient domain expertise limits their application in high-reliability scenarios. The sparse activation patterns of neurons in LLM feed-forward layers suggest that structured fine-tuning strategies for domain adaptation can be constructed by identifying important parameters, thereby enhancing domain-specific performance while preserving general capabilities. However, due to high computational costs or weak scalability, current studies struggle to support large-scale and fine-grained domain-relevant parameter analysis. To address this, we investigate the distribution patterns and hierarchical evolution patterns of domain knowledge within the hidden state space of LLMs, and propose a novel parameter evaluation method: Subspace Alignment Score. Intervention experiments across three models and six domains illustrate that, compared with baseline methods, our method demonstrates superior performance in locating domain-relevant important parameters. Furthermore, structured fine-tuning strategies based on the important parameters achieve performance comparable to full fine-tuning while substantially reducing training costs, which is helpful for understanding efficient and interpretable domain adaptation. Our code is available at https://github.com/lewu0723/subspace-alignment-score.
%U https://aclanthology.org/2026.emnlp-main.1504/
%P 34347-34366
Markdown (Informal)
[Subspace Alignment Score: Locating Domain-Relevant Important Parameters in Large Language Models](https://aclanthology.org/2026.emnlp-main.1504/) (Wu et al., EMNLP 2026)
ACL