@inproceedings{vijjini-chaturvedi-2026-fair,
title = "{FAIR}: Balancing Identity-Associated Training Signals to Mitigate Personalization Bias",
author = "Vijjini, Anvesh Rao and
Chaturvedi, Snigdha",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.160/",
pages = "3191--3208",
ISBN = "979-8-89176-444-6",
abstract = "Personalization of LLMs can greatly improve user experience by catering to individual preferences and contexts. However, unnecessary differences in model behavior across users based on their personal information can be undesirable. Such differences introduce \textit{personalization bias}, leading to inconsistent or unsafe behavior across social attributes such as religion, race, and gender. We introduce (Fairness Aware Imbalance Reduction), a framework for mitigating personalization bias by addressing imbalances in stereotype-related content during preference optimization. includes two variants: {\_}D, which applies data-level intervention based on data augmentation with LLMs, and {\_}T, which incorporates fairness-aware reweighting directly into the optimization objective. We evaluate both variants using multiple personalization bias metrics across diverse identities and models. Across all three LLMs, both methods consistently reduce bias relative to baseline, with up to 20 and 30.1 relative improvement by {\_}D and {\_}T, respectively."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="vijjini-chaturvedi-2026-fair">
<titleInfo>
<title>FAIR: Balancing Identity-Associated Training Signals to Mitigate Personalization Bias</title>
</titleInfo>
<name type="personal">
<namePart type="given">Anvesh</namePart>
<namePart type="given">Rao</namePart>
<namePart type="family">Vijjini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Snigdha</namePart>
<namePart type="family">Chaturvedi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Personalization of LLMs can greatly improve user experience by catering to individual preferences and contexts. However, unnecessary differences in model behavior across users based on their personal information can be undesirable. Such differences introduce personalization bias, leading to inconsistent or unsafe behavior across social attributes such as religion, race, and gender. We introduce (Fairness Aware Imbalance Reduction), a framework for mitigating personalization bias by addressing imbalances in stereotype-related content during preference optimization. includes two variants: _D, which applies data-level intervention based on data augmentation with LLMs, and _T, which incorporates fairness-aware reweighting directly into the optimization objective. We evaluate both variants using multiple personalization bias metrics across diverse identities and models. Across all three LLMs, both methods consistently reduce bias relative to baseline, with up to 20 and 30.1 relative improvement by _D and _T, respectively.</abstract>
<identifier type="citekey">vijjini-chaturvedi-2026-fair</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.160/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>3191</start>
<end>3208</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T FAIR: Balancing Identity-Associated Training Signals to Mitigate Personalization Bias
%A Vijjini, Anvesh Rao
%A Chaturvedi, Snigdha
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F vijjini-chaturvedi-2026-fair
%X Personalization of LLMs can greatly improve user experience by catering to individual preferences and contexts. However, unnecessary differences in model behavior across users based on their personal information can be undesirable. Such differences introduce personalization bias, leading to inconsistent or unsafe behavior across social attributes such as religion, race, and gender. We introduce (Fairness Aware Imbalance Reduction), a framework for mitigating personalization bias by addressing imbalances in stereotype-related content during preference optimization. includes two variants: _D, which applies data-level intervention based on data augmentation with LLMs, and _T, which incorporates fairness-aware reweighting directly into the optimization objective. We evaluate both variants using multiple personalization bias metrics across diverse identities and models. Across all three LLMs, both methods consistently reduce bias relative to baseline, with up to 20 and 30.1 relative improvement by _D and _T, respectively.
%U https://aclanthology.org/2026.findings-aacl.160/
%P 3191-3208
Markdown (Informal)
[FAIR: Balancing Identity-Associated Training Signals to Mitigate Personalization Bias](https://aclanthology.org/2026.findings-aacl.160/) (Vijjini & Chaturvedi, Findings 2026)
ACL