@inproceedings{wang-etal-2026-ethical,
title = "Ethical Imprints in {LLM}s: From Corpus Ethics to {ICL}, {SFT}, and Their Interaction on Deception Dilemmas",
author = "Wang, Jiafeng and
Chen, Jiale and
Luo, Zheng and
Yan, Xinyi and
Kong, Lingpeng",
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.1808/",
pages = "38630--38647",
ISBN = "979-8-89176-462-0",
abstract = "Large language models make consistent moral judgments, but whether these tendencies correspond to the moral makeup of public web text in the same language is not well understood. We test this correspondence for deception dilemmas across Chinese and English in a four-layer empirical study. We use DeepSeek-R1 for ICL probing, Llama-3.1-8B and Qwen2.5-7B for fine-tuning, and GPT-4o for cross-model replication. In both languages, roughly 60{\%} of web text favors utilitarian reasoning (59.8{\%} vs. 62.3{\%}), inconsistent with a purely sociolinguistic explanation. In-context learning shifts model decisions sharply in low-resistance scenarios (Hostage Crisis, V = 0.411, p = 0.001). In a high-resistance scenario (Organ Allocation), one DeepSeek-R1 deployment showed near-zero ICL effects (V = 0.067) at small sample size, but a larger replication on another deployment showed a large effect (V = 0.55 to 0.69), indicating deployment-specific rather than stable resistance. Three findings stand out for multilingual alignment. First, fine-tuning on deontological data reverses direction in English (Llama-EN Deon P(A) rises from ${\sim}$0{\%} to 88{\%}), showing that SFT value transfer is not language-neutral. Second, corpus proportions do not carry over to forced choices. The English Llama-3.1-8B base model gives P(A) near zero even though 62.3{\%} of the web text is utilitarian. Third, apparent moral damping in high-resistance scenarios can reflect sample size and deployment variability rather than a stable property of the model or scenario. These results show that corpus moral distributions correspond to model behavior in many conditions, but language, post-training, model architecture, and deployment variability each introduce points where this correspondence breaks."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-ethical">
<titleInfo>
<title>Ethical Imprints in LLMs: From Corpus Ethics to ICL, SFT, and Their Interaction on Deception Dilemmas</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jiafeng</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiale</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zheng</namePart>
<namePart type="family">Luo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinyi</namePart>
<namePart type="family">Yan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lingpeng</namePart>
<namePart type="family">Kong</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>Large language models make consistent moral judgments, but whether these tendencies correspond to the moral makeup of public web text in the same language is not well understood. We test this correspondence for deception dilemmas across Chinese and English in a four-layer empirical study. We use DeepSeek-R1 for ICL probing, Llama-3.1-8B and Qwen2.5-7B for fine-tuning, and GPT-4o for cross-model replication. In both languages, roughly 60% of web text favors utilitarian reasoning (59.8% vs. 62.3%), inconsistent with a purely sociolinguistic explanation. In-context learning shifts model decisions sharply in low-resistance scenarios (Hostage Crisis, V = 0.411, p = 0.001). In a high-resistance scenario (Organ Allocation), one DeepSeek-R1 deployment showed near-zero ICL effects (V = 0.067) at small sample size, but a larger replication on another deployment showed a large effect (V = 0.55 to 0.69), indicating deployment-specific rather than stable resistance. Three findings stand out for multilingual alignment. First, fine-tuning on deontological data reverses direction in English (Llama-EN Deon P(A) rises from \sim0% to 88%), showing that SFT value transfer is not language-neutral. Second, corpus proportions do not carry over to forced choices. The English Llama-3.1-8B base model gives P(A) near zero even though 62.3% of the web text is utilitarian. Third, apparent moral damping in high-resistance scenarios can reflect sample size and deployment variability rather than a stable property of the model or scenario. These results show that corpus moral distributions correspond to model behavior in many conditions, but language, post-training, model architecture, and deployment variability each introduce points where this correspondence breaks.</abstract>
<identifier type="citekey">wang-etal-2026-ethical</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1808/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>38630</start>
<end>38647</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Ethical Imprints in LLMs: From Corpus Ethics to ICL, SFT, and Their Interaction on Deception Dilemmas
%A Wang, Jiafeng
%A Chen, Jiale
%A Luo, Zheng
%A Yan, Xinyi
%A Kong, Lingpeng
%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 wang-etal-2026-ethical
%X Large language models make consistent moral judgments, but whether these tendencies correspond to the moral makeup of public web text in the same language is not well understood. We test this correspondence for deception dilemmas across Chinese and English in a four-layer empirical study. We use DeepSeek-R1 for ICL probing, Llama-3.1-8B and Qwen2.5-7B for fine-tuning, and GPT-4o for cross-model replication. In both languages, roughly 60% of web text favors utilitarian reasoning (59.8% vs. 62.3%), inconsistent with a purely sociolinguistic explanation. In-context learning shifts model decisions sharply in low-resistance scenarios (Hostage Crisis, V = 0.411, p = 0.001). In a high-resistance scenario (Organ Allocation), one DeepSeek-R1 deployment showed near-zero ICL effects (V = 0.067) at small sample size, but a larger replication on another deployment showed a large effect (V = 0.55 to 0.69), indicating deployment-specific rather than stable resistance. Three findings stand out for multilingual alignment. First, fine-tuning on deontological data reverses direction in English (Llama-EN Deon P(A) rises from \sim0% to 88%), showing that SFT value transfer is not language-neutral. Second, corpus proportions do not carry over to forced choices. The English Llama-3.1-8B base model gives P(A) near zero even though 62.3% of the web text is utilitarian. Third, apparent moral damping in high-resistance scenarios can reflect sample size and deployment variability rather than a stable property of the model or scenario. These results show that corpus moral distributions correspond to model behavior in many conditions, but language, post-training, model architecture, and deployment variability each introduce points where this correspondence breaks.
%U https://aclanthology.org/2026.findings-emnlp.1808/
%P 38630-38647
Markdown (Informal)
[Ethical Imprints in LLMs: From Corpus Ethics to ICL, SFT, and Their Interaction on Deception Dilemmas](https://aclanthology.org/2026.findings-emnlp.1808/) (Wang et al., Findings 2026)
ACL