@inproceedings{zhu-etal-2026-preserving,
title = "Preserving General Capability in Personality-Aligned {LLM}s via Orthogonal Gradient Projection",
author = "Zhu, Yangfu and
Wei, Linyue and
Ning, Nianwen and
Zhou, Pengpeng and
Song, Chenguang and
Wang, Yuandong and
Shao, Zhenzhou",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1631/",
doi = "10.18653/2026.findings-emnlp.1631",
pages = "34698--34714",
ISBN = "979-8-89176-462-0",
abstract = "Personality alignment tailors Large Language Models (LLMs) for personalized interactions. Most existing studies focus on prompting, post- training, or representation steering to elicit trait expression, but they often compromise the general capabilities like reasoning and truthfulness. In this work, we empirically verified the existence of the Personality Alignment Tax (PAT) that distribution shifts and conflicting objectives during personality updates overwrite pre-trained competencies. To mitigate this tax, we propose a Personality-Orthogonal Gradient Projection (POGP) plug-in that preserves the capability in personality alignment. Specifically, POGP estimates the general-capability subspace from capability reference benchmarks and projects personality gradients onto the orthogonal complement before updating. This isolates and neutralizes conflicting gradients while maintaining the pre-trained reasoning capabilities. Experiments on the BIG5-CHAT dataset with two backbone LLMs demonstrate that POGP achieves a stronger personality plasticity and capability stability trade-off than baselines."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhu-etal-2026-preserving">
<titleInfo>
<title>Preserving General Capability in Personality-Aligned LLMs via Orthogonal Gradient Projection</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yangfu</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Linyue</namePart>
<namePart type="family">Wei</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nianwen</namePart>
<namePart type="family">Ning</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pengpeng</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chenguang</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuandong</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhenzhou</namePart>
<namePart type="family">Shao</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>Findings of the Association for Computational Linguistics: EMNLP 2026</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">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Personality alignment tailors Large Language Models (LLMs) for personalized interactions. Most existing studies focus on prompting, post- training, or representation steering to elicit trait expression, but they often compromise the general capabilities like reasoning and truthfulness. In this work, we empirically verified the existence of the Personality Alignment Tax (PAT) that distribution shifts and conflicting objectives during personality updates overwrite pre-trained competencies. To mitigate this tax, we propose a Personality-Orthogonal Gradient Projection (POGP) plug-in that preserves the capability in personality alignment. Specifically, POGP estimates the general-capability subspace from capability reference benchmarks and projects personality gradients onto the orthogonal complement before updating. This isolates and neutralizes conflicting gradients while maintaining the pre-trained reasoning capabilities. Experiments on the BIG5-CHAT dataset with two backbone LLMs demonstrate that POGP achieves a stronger personality plasticity and capability stability trade-off than baselines.</abstract>
<identifier type="citekey">zhu-etal-2026-preserving</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1631</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1631/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>34698</start>
<end>34714</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Preserving General Capability in Personality-Aligned LLMs via Orthogonal Gradient Projection
%A Zhu, Yangfu
%A Wei, Linyue
%A Ning, Nianwen
%A Zhou, Pengpeng
%A Song, Chenguang
%A Wang, Yuandong
%A Shao, Zhenzhou
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F zhu-etal-2026-preserving
%X Personality alignment tailors Large Language Models (LLMs) for personalized interactions. Most existing studies focus on prompting, post- training, or representation steering to elicit trait expression, but they often compromise the general capabilities like reasoning and truthfulness. In this work, we empirically verified the existence of the Personality Alignment Tax (PAT) that distribution shifts and conflicting objectives during personality updates overwrite pre-trained competencies. To mitigate this tax, we propose a Personality-Orthogonal Gradient Projection (POGP) plug-in that preserves the capability in personality alignment. Specifically, POGP estimates the general-capability subspace from capability reference benchmarks and projects personality gradients onto the orthogonal complement before updating. This isolates and neutralizes conflicting gradients while maintaining the pre-trained reasoning capabilities. Experiments on the BIG5-CHAT dataset with two backbone LLMs demonstrate that POGP achieves a stronger personality plasticity and capability stability trade-off than baselines.
%R 10.18653/2026.findings-emnlp.1631
%U https://aclanthology.org/2026.findings-emnlp.1631/
%U https://doi.org/10.18653/2026.findings-emnlp.1631
%P 34698-34714
Markdown (Informal)
[Preserving General Capability in Personality-Aligned LLMs via Orthogonal Gradient Projection](https://aclanthology.org/2026.findings-emnlp.1631/) (Zhu et al., Findings 2026)
ACL