Model Merging for Knowledge Editing

Zichuan Fu; Xian Wu; Guojing Li; Yingying Zhang; Yefeng Zheng; Tianshi Ming; Yejing Wang; Wanyu Wang; Xiangyu Zhao

doi:10.18653/v1/2025.acl-industry.30

Model Merging for Knowledge Editing

Zichuan Fu
, Xian Wu^*
, Guojing Li
, Yingying Zhang
, Yefeng Zheng
, Tianshi Ming
, Yejing Wang
, Wanyu Wang
, Xiangyu Zhao^*

^*Corresponding author for this work

Department of Data Science

Research output: Chapters, Conference Papers, Creative and Literary Works › RGC 32 - Refereed conference paper (with host publication) › peer-review

5 Downloads (CityUHK Scholars)

Abstract

Large Language Models (LLMs) require continuous updates to maintain accurate and current knowledge as the world evolves. While existing knowledge editing approaches offer various solutions for knowledge updating, they often struggle with sequential editing scenarios and harm the general capabilities of the model, thereby significantly hampering their practical applicability. This paper proposes a two-stage framework combining robust supervised fine-tuning (R-SFT) with model merging for knowledge editing. Our method first fine-tunes the LLM to internalize new knowledge fully, then merges the fine-tuned model with the original foundation model to preserve newly acquired knowledge and general capabilities. Experimental results demonstrate that our approach significantly outperforms existing methods in sequential editing while better preserving the original performance of the model, all without requiring any architectural changes. Code is available at Applied-Machine-Learning-Lab/MM4KE.
©2025 Association for Computational Linguistics

Original language	English
Title of host publication	The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)
Subtitle of host publication	Proceedings of the Conference– Volume 6: Industry Track
Editors	Georg Rehm, Yunyao Li
Publisher	Association for Computational Linguistics
Pages	433–443
Number of pages	11
ISBN (Electronic)	979-8-89176-288-6
DOIs	https://doi.org/10.18653/v1/2025.acl-industry.30
Publication status	Published - Jul 2025
Event	63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025) - Austria Center Vienna, Vienna, Austria Duration: 27 Jul 2025 → 1 Aug 2025 https://2025.aclweb.org/ https://aclanthology.org/2025.acl-long/ https://aclanthology.org/volumes/2025.findings-acl/

Publication series

Name	Proceedings of the Annual Meeting of the Association for Computational Linguistics
Publisher	ACL Anthology
ISSN (Print)	0736-587X

Conference

Conference	63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)
Place	Austria
City	Vienna
Period	27/07/25 → 1/08/25
Internet address	https://2025.aclweb.org/ https://aclanthology.org/2025.acl-long/ https://aclanthology.org/volumes/2025.findings-acl/

Bibliographical note

Research Unit(s) information for this publication is provided by the author(s) concerned.

Funding

This research was partially supported by Research Impact Fund (No.R1015-23), Collaborative Research Fund (No.C1043-24GF) and Tencent (CCF-Tencent Open Fund, Tencent Rhino-Bird Focused Research Program).

Research Keywords

model merging
knowledge edit

Publisher's Copyright Statement

This full text is made available under CC-BY 4.0. https://creativecommons.org/licenses/by/4.0/

RGC Funding Information

RGC-funded

Access Output

10.18653/v1/2025.acl-industry.30

295603264.pdfFinal Published version, 281 KBLicence: CC BY

Other Access Options

Check CityUHK Library

Permanent Link

Right click to copy link

RIF: Integrating ChatGPT with Search Engine, Recommender System and Online Advertising to Enhance User Experience on Online Service Platforms
ZHAO, X. (Principal Investigator / Project Coordinator), KING, I. K. C. (Co-Investigator), LI, Y. D. (Co-Investigator), Li, Q. (Co-Investigator), QIN, S. Z. J. (Co-Investigator) & Xu, J. (Co-Investigator)
1/06/24 → …
Project: Research

Cite this

Fu, Z., Wu, X., Li, G., Zhang, Y., Zheng, Y., Ming, T., Wang, Y., Wang, W., & Zhao, X. (2025). Model Merging for Knowledge Editing. In G. Rehm, & Y. Li (Eds.), The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025): Proceedings of the Conference– Volume 6: Industry Track (pp. 433–443). (Proceedings of the Annual Meeting of the Association for Computational Linguistics). Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.acl-industry.30

Fu, Zichuan ; Wu, Xian ; Li, Guojing et al. / Model Merging for Knowledge Editing. The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025): Proceedings of the Conference– Volume 6: Industry Track. editor / Georg Rehm ; Yunyao Li. Association for Computational Linguistics, 2025. pp. 433–443 (Proceedings of the Annual Meeting of the Association for Computational Linguistics).

@inproceedings{391a1edf2efa4e7ca2187b90a4d65b1a,

title = "Model Merging for Knowledge Editing",

abstract = "Large Language Models (LLMs) require continuous updates to maintain accurate and current knowledge as the world evolves. While existing knowledge editing approaches offer various solutions for knowledge updating, they often struggle with sequential editing scenarios and harm the general capabilities of the model, thereby significantly hampering their practical applicability. This paper proposes a two-stage framework combining robust supervised fine-tuning (R-SFT) with model merging for knowledge editing. Our method first fine-tunes the LLM to internalize new knowledge fully, then merges the fine-tuned model with the original foundation model to preserve newly acquired knowledge and general capabilities. Experimental results demonstrate that our approach significantly outperforms existing methods in sequential editing while better preserving the original performance of the model, all without requiring any architectural changes. Code is available at Applied-Machine-Learning-Lab/MM4KE. {\textcopyright}2025 Association for Computational Linguistics",

keywords = "model merging, knowledge edit",

author = "Zichuan Fu and Xian Wu and Guojing Li and Yingying Zhang and Yefeng Zheng and Tianshi Ming and Yejing Wang and Wanyu Wang and Xiangyu Zhao",

note = "Research Unit(s) information for this publication is provided by the author(s) concerned.; 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025) ; Conference date: 27-07-2025 Through 01-08-2025",

year = "2025",

month = jul,

doi = "10.18653/v1/2025.acl-industry.30",

language = "English",

series = "Proceedings of the Annual Meeting of the Association for Computational Linguistics",

publisher = "Association for Computational Linguistics",

pages = "433–443",

editor = "Georg Rehm and Yunyao Li",

booktitle = "The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)",

url = "https://2025.aclweb.org/, https://aclanthology.org/2025.acl-long/, https://aclanthology.org/volumes/2025.findings-acl/",

}

Fu, Z, Wu, X, Li, G, Zhang, Y, Zheng, Y, Ming, T, Wang, Y , Wang, W & Zhao, X 2025, Model Merging for Knowledge Editing. in G Rehm & Y Li (eds), The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025): Proceedings of the Conference– Volume 6: Industry Track. Proceedings of the Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, pp. 433–443, 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), Vienna, Austria, 27/07/25. https://doi.org/10.18653/v1/2025.acl-industry.30

Model Merging for Knowledge Editing. / Fu, Zichuan; Wu, Xian; Li, Guojing et al.
The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025): Proceedings of the Conference– Volume 6: Industry Track. ed. / Georg Rehm; Yunyao Li. Association for Computational Linguistics, 2025. p. 433–443 (Proceedings of the Annual Meeting of the Association for Computational Linguistics).

Research output: Chapters, Conference Papers, Creative and Literary Works › RGC 32 - Refereed conference paper (with host publication) › peer-review

TY - GEN

T1 - Model Merging for Knowledge Editing

AU - Fu, Zichuan

AU - Wu, Xian

AU - Li, Guojing

AU - Zhang, Yingying

AU - Zheng, Yefeng

AU - Ming, Tianshi

AU - Wang, Yejing

AU - Wang, Wanyu

AU - Zhao, Xiangyu

N1 - Research Unit(s) information for this publication is provided by the author(s) concerned.

PY - 2025/7

Y1 - 2025/7

N2 - Large Language Models (LLMs) require continuous updates to maintain accurate and current knowledge as the world evolves. While existing knowledge editing approaches offer various solutions for knowledge updating, they often struggle with sequential editing scenarios and harm the general capabilities of the model, thereby significantly hampering their practical applicability. This paper proposes a two-stage framework combining robust supervised fine-tuning (R-SFT) with model merging for knowledge editing. Our method first fine-tunes the LLM to internalize new knowledge fully, then merges the fine-tuned model with the original foundation model to preserve newly acquired knowledge and general capabilities. Experimental results demonstrate that our approach significantly outperforms existing methods in sequential editing while better preserving the original performance of the model, all without requiring any architectural changes. Code is available at Applied-Machine-Learning-Lab/MM4KE. ©2025 Association for Computational Linguistics

AB - Large Language Models (LLMs) require continuous updates to maintain accurate and current knowledge as the world evolves. While existing knowledge editing approaches offer various solutions for knowledge updating, they often struggle with sequential editing scenarios and harm the general capabilities of the model, thereby significantly hampering their practical applicability. This paper proposes a two-stage framework combining robust supervised fine-tuning (R-SFT) with model merging for knowledge editing. Our method first fine-tunes the LLM to internalize new knowledge fully, then merges the fine-tuned model with the original foundation model to preserve newly acquired knowledge and general capabilities. Experimental results demonstrate that our approach significantly outperforms existing methods in sequential editing while better preserving the original performance of the model, all without requiring any architectural changes. Code is available at Applied-Machine-Learning-Lab/MM4KE. ©2025 Association for Computational Linguistics

KW - model merging

KW - knowledge edit

U2 - 10.18653/v1/2025.acl-industry.30

DO - 10.18653/v1/2025.acl-industry.30

M3 - RGC 32 - Refereed conference paper (with host publication)

T3 - Proceedings of the Annual Meeting of the Association for Computational Linguistics

SP - 433

EP - 443

BT - The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

A2 - Rehm, Georg

A2 - Li, Yunyao

PB - Association for Computational Linguistics

T2 - 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Y2 - 27 July 2025 through 1 August 2025

ER -

Fu Z, Wu X, Li G, Zhang Y, Zheng Y, Ming T et al. Model Merging for Knowledge Editing. In Rehm G, Li Y, editors, The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025): Proceedings of the Conference– Volume 6: Industry Track. Association for Computational Linguistics. 2025. p. 433–443. (Proceedings of the Annual Meeting of the Association for Computational Linguistics). doi: 10.18653/v1/2025.acl-industry.30

Model Merging for Knowledge Editing

Abstract

Publication series

Conference

Bibliographical note

Funding

Research Keywords

Publisher's Copyright Statement

RGC Funding Information

Access Output

Other Access Options

Permanent Link

Fingerprint

Projects

RIF: Integrating ChatGPT with Search Engine, Recommender System and Online Advertising to Enhance User Experience on Online Service Platforms

Cite this