<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="feng-etal-2026-prompted">
    <titleInfo>
        <title>Prompted Values and Deployed Policies: A Truth–Propriety Diagnostic in Language Models</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Ziyuan</namePart>
        <namePart type="family">Feng</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Ajian</namePart>
        <namePart type="family">Liu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Hui</namePart>
        <namePart type="family">Ma</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yanyan</namePart>
        <namePart type="family">Liang</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>Prompting a language model to endorse a value does not by itself establish how that value is deployed when responses require tradeoffs. We study this gap in truth–propriety dilemmas, where direct truth-telling can conflict with tact, face-saving, deference, or role-sensitive restraint. We introduce Truth–Propriety Alignment (TPA), a 600-prompt diagnostic benchmark over 100 scenario groups, and conduct a single-model diagnostic evaluation of Qwen2.5-7B-Instruct under deterministic decoding. The model remains predominantly truth-oriented, but trigger-assigned scenario groups differ sharply in how residual responses are distributed. Because value triggers are fixed within scenario groups, we interpret this pattern as trigger-associated residual composition rather than a causal trigger effect. A small semantic audit shows that compatible residuals are mixed: they include substantive reframings but also tonal and lexical tracking. Finally, a hidden-state probe is nearly matched by an exact-prompt BoW baseline, and deleting the explicit value-trigger sentence does not reduce BoW performance, indicating that behavioral-label predictability is distributed across the prompt surface. We therefore frame TPA as a diagnostic audit of prompt intent, observed behavior, scenario/trigger priors, annotation boundaries, and probe evidence, rather than as evidence for prompt-independent value representations.</abstract>
    <identifier type="citekey">feng-etal-2026-prompted</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-emnlp.1165/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>24673</start>
            <end>24686</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Prompted Values and Deployed Policies: A Truth–Propriety Diagnostic in Language Models
%A Feng, Ziyuan
%A Liu, Ajian
%A Ma, Hui
%A Liang, Yanyan
%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 feng-etal-2026-prompted
%X Prompting a language model to endorse a value does not by itself establish how that value is deployed when responses require tradeoffs. We study this gap in truth–propriety dilemmas, where direct truth-telling can conflict with tact, face-saving, deference, or role-sensitive restraint. We introduce Truth–Propriety Alignment (TPA), a 600-prompt diagnostic benchmark over 100 scenario groups, and conduct a single-model diagnostic evaluation of Qwen2.5-7B-Instruct under deterministic decoding. The model remains predominantly truth-oriented, but trigger-assigned scenario groups differ sharply in how residual responses are distributed. Because value triggers are fixed within scenario groups, we interpret this pattern as trigger-associated residual composition rather than a causal trigger effect. A small semantic audit shows that compatible residuals are mixed: they include substantive reframings but also tonal and lexical tracking. Finally, a hidden-state probe is nearly matched by an exact-prompt BoW baseline, and deleting the explicit value-trigger sentence does not reduce BoW performance, indicating that behavioral-label predictability is distributed across the prompt surface. We therefore frame TPA as a diagnostic audit of prompt intent, observed behavior, scenario/trigger priors, annotation boundaries, and probe evidence, rather than as evidence for prompt-independent value representations.
%U https://aclanthology.org/2026.findings-emnlp.1165/
%P 24673-24686

Markdown (Informal)

[Prompted Values and Deployed Policies: A Truth–Propriety Diagnostic in Language Models](https://aclanthology.org/2026.findings-emnlp.1165/) (Feng et al., Findings 2026)

• Prompted Values and Deployed Policies: A Truth–Propriety Diagnostic in Language Models (Feng et al., Findings 2026)

ACL

• Ziyuan Feng, Ajian Liu, Hui Ma, and Yanyan Liang. 2026. Prompted Values and Deployed Policies: A Truth–Propriety Diagnostic in Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 24673–24686, Budapest, Hungary. Association for Computational Linguistics.