@inproceedings{shi-etal-2026-moralsteer,
title = "{M}oral{S}teer: Safety-Preserving Fidelity Gains in Morally Complex Role-Playing via Theory-Guided On-Policy Self-Distillation",
author = "Shi, Shaojie and
Hao, Teqi and
Lv, Bohao and
Xu, Rui and
Shi, Zhengyu and
Yinghui, Xu",
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.1554/",
pages = "35450--35463",
ISBN = "979-8-89176-459-0",
abstract = "Aligned LLMs systematically fail at morally complex role-playing: villains collapse into prosocial defaults due to a structural safety and fidelity tension. We present MoralSteer, a theory-guided on-policy self-distillation framework that improves morally complex role-play fidelity while preserving the safety behavior of aligned base models. MoralSteer treats Moral Foundations Theory (MFT) reasoning chains as privileged information for a frozen teacher, introduces a moral-complexity-adaptive KL interpolation $\alpha(c){=}\sigma(M(c){-}\tau)$ to prevent prosocial collapse, and couples this with $M(c)$-adaptive trajectory mixing to mitigate rollout degeneration. Across four base models, MoralSteer improves Moral RolePlay L3+L4 scores by $+5.8\sim+9.9\%$; across five safety benchmarks, the observed rate differences remain within $\pm$1pp of the aligned bases. We find that villain failure concentrates in Care/Fairness foundations (67{\%} of the L4 gap) and that aligned models perform trait substitution rather than deletion. Our code is available at https://github.com/Sii-yuning/MoralSteer."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shi-etal-2026-moralsteer">
<titleInfo>
<title>MoralSteer: Safety-Preserving Fidelity Gains in Morally Complex Role-Playing via Theory-Guided On-Policy Self-Distillation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shaojie</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Teqi</namePart>
<namePart type="family">Hao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bohao</namePart>
<namePart type="family">Lv</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rui</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhengyu</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xu</namePart>
<namePart type="family">Yinghui</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>Aligned LLMs systematically fail at morally complex role-playing: villains collapse into prosocial defaults due to a structural safety and fidelity tension. We present MoralSteer, a theory-guided on-policy self-distillation framework that improves morally complex role-play fidelity while preserving the safety behavior of aligned base models. MoralSteer treats Moral Foundations Theory (MFT) reasoning chains as privileged information for a frozen teacher, introduces a moral-complexity-adaptive KL interpolation α(c)=σ(M(c)-τ) to prevent prosocial collapse, and couples this with M(c)-adaptive trajectory mixing to mitigate rollout degeneration. Across four base models, MoralSteer improves Moral RolePlay L3+L4 scores by +5.8\sim+9.9%; across five safety benchmarks, the observed rate differences remain within \pm1pp of the aligned bases. We find that villain failure concentrates in Care/Fairness foundations (67% of the L4 gap) and that aligned models perform trait substitution rather than deletion. Our code is available at https://github.com/Sii-yuning/MoralSteer.</abstract>
<identifier type="citekey">shi-etal-2026-moralsteer</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1554/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>35450</start>
<end>35463</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T MoralSteer: Safety-Preserving Fidelity Gains in Morally Complex Role-Playing via Theory-Guided On-Policy Self-Distillation
%A Shi, Shaojie
%A Hao, Teqi
%A Lv, Bohao
%A Xu, Rui
%A Shi, Zhengyu
%A Yinghui, Xu
%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 shi-etal-2026-moralsteer
%X Aligned LLMs systematically fail at morally complex role-playing: villains collapse into prosocial defaults due to a structural safety and fidelity tension. We present MoralSteer, a theory-guided on-policy self-distillation framework that improves morally complex role-play fidelity while preserving the safety behavior of aligned base models. MoralSteer treats Moral Foundations Theory (MFT) reasoning chains as privileged information for a frozen teacher, introduces a moral-complexity-adaptive KL interpolation α(c)=σ(M(c)-τ) to prevent prosocial collapse, and couples this with M(c)-adaptive trajectory mixing to mitigate rollout degeneration. Across four base models, MoralSteer improves Moral RolePlay L3+L4 scores by +5.8\sim+9.9%; across five safety benchmarks, the observed rate differences remain within \pm1pp of the aligned bases. We find that villain failure concentrates in Care/Fairness foundations (67% of the L4 gap) and that aligned models perform trait substitution rather than deletion. Our code is available at https://github.com/Sii-yuning/MoralSteer.
%U https://aclanthology.org/2026.emnlp-main.1554/
%P 35450-35463
Markdown (Informal)
[MoralSteer: Safety-Preserving Fidelity Gains in Morally Complex Role-Playing via Theory-Guided On-Policy Self-Distillation](https://aclanthology.org/2026.emnlp-main.1554/) (Shi et al., EMNLP 2026)
ACL