@inproceedings{liu-etal-2025-disentangling,
title = "Disentangling Biased Knowledge from Reasoning in Large Language Models via Machine Unlearning",
author = "Liu, Zheyuan and
Maharjan, Suraj and
Wu, Fanyou and
Parikh, Rahil and
Bayar, Belhassen and
Sengamedu, Srinivasan H. and
Jiang, Meng",
editor = "Che, Wanxiang and
Nabende, Joyce and
Shutova, Ekaterina and
Pilehvar, Mohammad Taher",
booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = jul,
year = "2025",
address = "Vienna, Austria",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.acl-long.305/",
doi = "10.18653/v1/2025.acl-long.305",
pages = "6105--6123",
ISBN = "979-8-89176-251-0",
abstract = "The rapid development of Large Language Models (LLMs) has led to their widespread adoption across various domains, leveraging vast pre-training knowledge and impressive generalization capabilities. However, these models often inherit biased knowledge, resulting in unfair decisions in sensitive applications. It is challenging to remove this biased knowledge without compromising reasoning abilities due to the entangled nature of the learned knowledge within LLMs. To solve this problem, existing approaches have attempted to mitigate the bias using techniques such as fine-tuning with unbiased datasets, model merging, and gradient ascent. While these methods have experimentally proven effective, they can still be sub-optimum in fully disentangling biases from reasoning. To address this gap, we propose Selective Disentanglement Unlearning (SDU), a novel unlearning framework that selectively removes biased knowledge while preserving reasoning capabilities. SDU operates in three stages: identifying biased parameters using a shadow LLM, fine-tuning with unbiased data, and performing selective parameter updates based on weight saliency. Experimental results across multiple LLMs show that SDU improves fairness accuracy by 14.7{\%} and enhances reasoning performance by 62.6{\%} compared to existing baselines."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liu-etal-2025-disentangling">
<titleInfo>
<title>Disentangling Biased Knowledge from Reasoning in Large Language Models via Machine Unlearning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zheyuan</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Suraj</namePart>
<namePart type="family">Maharjan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fanyou</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rahil</namePart>
<namePart type="family">Parikh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Belhassen</namePart>
<namePart type="family">Bayar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Srinivasan</namePart>
<namePart type="given">H</namePart>
<namePart type="family">Sengamedu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Meng</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Wanxiang</namePart>
<namePart type="family">Che</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joyce</namePart>
<namePart type="family">Nabende</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ekaterina</namePart>
<namePart type="family">Shutova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohammad</namePart>
<namePart type="given">Taher</namePart>
<namePart type="family">Pilehvar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Vienna, Austria</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-251-0</identifier>
</relatedItem>
<abstract>The rapid development of Large Language Models (LLMs) has led to their widespread adoption across various domains, leveraging vast pre-training knowledge and impressive generalization capabilities. However, these models often inherit biased knowledge, resulting in unfair decisions in sensitive applications. It is challenging to remove this biased knowledge without compromising reasoning abilities due to the entangled nature of the learned knowledge within LLMs. To solve this problem, existing approaches have attempted to mitigate the bias using techniques such as fine-tuning with unbiased datasets, model merging, and gradient ascent. While these methods have experimentally proven effective, they can still be sub-optimum in fully disentangling biases from reasoning. To address this gap, we propose Selective Disentanglement Unlearning (SDU), a novel unlearning framework that selectively removes biased knowledge while preserving reasoning capabilities. SDU operates in three stages: identifying biased parameters using a shadow LLM, fine-tuning with unbiased data, and performing selective parameter updates based on weight saliency. Experimental results across multiple LLMs show that SDU improves fairness accuracy by 14.7% and enhances reasoning performance by 62.6% compared to existing baselines.</abstract>
<identifier type="citekey">liu-etal-2025-disentangling</identifier>
<identifier type="doi">10.18653/v1/2025.acl-long.305</identifier>
<location>
<url>https://aclanthology.org/2025.acl-long.305/</url>
</location>
<part>
<date>2025-07</date>
<extent unit="page">
<start>6105</start>
<end>6123</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Disentangling Biased Knowledge from Reasoning in Large Language Models via Machine Unlearning
%A Liu, Zheyuan
%A Maharjan, Suraj
%A Wu, Fanyou
%A Parikh, Rahil
%A Bayar, Belhassen
%A Sengamedu, Srinivasan H.
%A Jiang, Meng
%Y Che, Wanxiang
%Y Nabende, Joyce
%Y Shutova, Ekaterina
%Y Pilehvar, Mohammad Taher
%S Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2025
%8 July
%I Association for Computational Linguistics
%C Vienna, Austria
%@ 979-8-89176-251-0
%F liu-etal-2025-disentangling
%X The rapid development of Large Language Models (LLMs) has led to their widespread adoption across various domains, leveraging vast pre-training knowledge and impressive generalization capabilities. However, these models often inherit biased knowledge, resulting in unfair decisions in sensitive applications. It is challenging to remove this biased knowledge without compromising reasoning abilities due to the entangled nature of the learned knowledge within LLMs. To solve this problem, existing approaches have attempted to mitigate the bias using techniques such as fine-tuning with unbiased datasets, model merging, and gradient ascent. While these methods have experimentally proven effective, they can still be sub-optimum in fully disentangling biases from reasoning. To address this gap, we propose Selective Disentanglement Unlearning (SDU), a novel unlearning framework that selectively removes biased knowledge while preserving reasoning capabilities. SDU operates in three stages: identifying biased parameters using a shadow LLM, fine-tuning with unbiased data, and performing selective parameter updates based on weight saliency. Experimental results across multiple LLMs show that SDU improves fairness accuracy by 14.7% and enhances reasoning performance by 62.6% compared to existing baselines.
%R 10.18653/v1/2025.acl-long.305
%U https://aclanthology.org/2025.acl-long.305/
%U https://doi.org/10.18653/v1/2025.acl-long.305
%P 6105-6123
Markdown (Informal)
[Disentangling Biased Knowledge from Reasoning in Large Language Models via Machine Unlearning](https://aclanthology.org/2025.acl-long.305/) (Liu et al., ACL 2025)
ACL