@inproceedings{an-etal-2026-mitigating,
title = "Mitigating Model Merging Conflicts via Knowledge-Aware Task Vector Purification",
author = "An, Bang and
Yang, Yibo and
Torr, Philip and
Ghanem, Bernard",
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.200/",
pages = "4572--4590",
ISBN = "979-8-89176-459-0",
abstract = "Model merging aims to integrate task-specific abilities from individually fine-tuned models into a single model without extra training. In model merging methods, the task vector has become a fundamental building block, as it can encapsulate the information from fine-tuning. However, the merged model often suffers from performance degradation due to conflicts caused by task-irrelevant redundancy in task vectors. To address these challenges, we propose $\textbf{P}$urified T$\textbf{A}$sk $\textbf{VE}$ctors (PAVE) in knowledge-aware subspace. Concretely, we sample training examples from each task, and feed them into their fine-tuned models to acquire covariance matrices before linear layers. We then perform a context-oriented singular value decomposition, which accentuates the weight components most relevant to the target knowledge. As a result, we split fine-tuned model weights into task-relevant and redundant components in the knowledge-aware subspace, and purify the task vector by pruning redundant components. To induce fair pruning efforts across models, we further introduce a spectral rank allocation strategy by optimizing a normalized activated pruning error. The task vector purification by our method as a plug-and-play scheme is applicable across various task vector-based merging methods to improve their performance. In experiments, we demonstrate the effectiveness of PAVE across diverse merging methods, tasks, and model architectures."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="an-etal-2026-mitigating">
<titleInfo>
<title>Mitigating Model Merging Conflicts via Knowledge-Aware Task Vector Purification</title>
</titleInfo>
<name type="personal">
<namePart type="given">Bang</namePart>
<namePart type="family">An</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yibo</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Philip</namePart>
<namePart type="family">Torr</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bernard</namePart>
<namePart type="family">Ghanem</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>Model merging aims to integrate task-specific abilities from individually fine-tuned models into a single model without extra training. In model merging methods, the task vector has become a fundamental building block, as it can encapsulate the information from fine-tuning. However, the merged model often suffers from performance degradation due to conflicts caused by task-irrelevant redundancy in task vectors. To address these challenges, we propose Purified TAsk VEctors (PAVE) in knowledge-aware subspace. Concretely, we sample training examples from each task, and feed them into their fine-tuned models to acquire covariance matrices before linear layers. We then perform a context-oriented singular value decomposition, which accentuates the weight components most relevant to the target knowledge. As a result, we split fine-tuned model weights into task-relevant and redundant components in the knowledge-aware subspace, and purify the task vector by pruning redundant components. To induce fair pruning efforts across models, we further introduce a spectral rank allocation strategy by optimizing a normalized activated pruning error. The task vector purification by our method as a plug-and-play scheme is applicable across various task vector-based merging methods to improve their performance. In experiments, we demonstrate the effectiveness of PAVE across diverse merging methods, tasks, and model architectures.</abstract>
<identifier type="citekey">an-etal-2026-mitigating</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.200/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>4572</start>
<end>4590</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Mitigating Model Merging Conflicts via Knowledge-Aware Task Vector Purification
%A An, Bang
%A Yang, Yibo
%A Torr, Philip
%A Ghanem, Bernard
%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 an-etal-2026-mitigating
%X Model merging aims to integrate task-specific abilities from individually fine-tuned models into a single model without extra training. In model merging methods, the task vector has become a fundamental building block, as it can encapsulate the information from fine-tuning. However, the merged model often suffers from performance degradation due to conflicts caused by task-irrelevant redundancy in task vectors. To address these challenges, we propose Purified TAsk VEctors (PAVE) in knowledge-aware subspace. Concretely, we sample training examples from each task, and feed them into their fine-tuned models to acquire covariance matrices before linear layers. We then perform a context-oriented singular value decomposition, which accentuates the weight components most relevant to the target knowledge. As a result, we split fine-tuned model weights into task-relevant and redundant components in the knowledge-aware subspace, and purify the task vector by pruning redundant components. To induce fair pruning efforts across models, we further introduce a spectral rank allocation strategy by optimizing a normalized activated pruning error. The task vector purification by our method as a plug-and-play scheme is applicable across various task vector-based merging methods to improve their performance. In experiments, we demonstrate the effectiveness of PAVE across diverse merging methods, tasks, and model architectures.
%U https://aclanthology.org/2026.emnlp-main.200/
%P 4572-4590
Markdown (Informal)
[Mitigating Model Merging Conflicts via Knowledge-Aware Task Vector Purification](https://aclanthology.org/2026.emnlp-main.200/) (An et al., EMNLP 2026)
ACL