<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="saeed-etal-2026-surfacing">
    <titleInfo>
        <title>Surfacing Subtle Stereotypes: A Multilingual, Debate-Oriented Evaluation of Modern LLMs</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Muhammed</namePart>
        <namePart type="given">Yahia</namePart>
        <namePart type="given">Gaffar</namePart>
        <namePart type="family">Saeed</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Muhammad</namePart>
        <namePart type="family">Abdul-Mageed</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Shady</namePart>
        <namePart type="family">Shehata</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-05</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the Fifteenth Language Resources and Evaluation Conference</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Stelios</namePart>
            <namePart type="family">Piperidis</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Núria</namePart>
            <namePart type="family">Bel</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Henk</namePart>
            <namePart type="family">van den Heuvel</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Nancy</namePart>
            <namePart type="family">Ide</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Simon</namePart>
            <namePart type="family">Krek</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Antonio</namePart>
            <namePart type="family">Toral</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>ELRA Language Resource Association</publisher>
            <place>
                <placeTerm type="text">Palma de Mallorca, Spain</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
    </relatedItem>
    <abstract>Large language models (LLMs) are widely deployed for open-ended communication, yet most bias evaluations still rely on English, classification-style tasks. We introduce , a new multilingual, debate-style benchmark designed to reveal how narrative bias appears in realistic generative settings. Our dataset includes 8,400 structured debate prompts spanning four sensitive domains – Women’s Rights, Backwardness, Terrorism, and Religion – across seven languages ranging from high-resource (English, Chinese) to low-resource (Swahili, Nigerian Pidgin). Using four flagship models (GPT-4o, Claude 3.5 Haiku, DeepSeek-Chat, and LLaMA-3-70B), we generate over 100,000 debate responses and automatically classify which demographic groups are assigned stereotyped versus modern roles. Results show that all models reproduce entrenched stereotypes despite safety alignment: Arabs are overwhelmingly linked to Terrorism and Religion (\geq89%), Africans to socioeconomic “backwardness” (up to 77%), and Western groups are consistently framed as modern or progressive. Biases grow sharply in lower-resource languages, revealing that alignment trained primarily in English does not generalize globally. Our findings highlight a persistent divide in multilingual fairness: current alignment methods reduce explicit toxicity but fail to prevent biased outputs in open-ended contexts. We release our benchmark and analysis framework to support the next generation of multilingual bias evaluation and safer, culturally inclusive model alignment</abstract>
    <identifier type="citekey">saeed-etal-2026-surfacing</identifier>
    <identifier type="doi">10.63317/4u9x3z4g8jfk</identifier>
    <location>
        <url>https://aclanthology.org/2026.lrec-1.643/</url>
    </location>
    <part>
        <date>2026-05</date>
        <extent unit="page">
            <start>8106</start>
            <end>8121</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Surfacing Subtle Stereotypes: A Multilingual, Debate-Oriented Evaluation of Modern LLMs
%A Saeed, Muhammed Yahia Gaffar
%A Abdul-Mageed, Muhammad
%A Shehata, Shady
%Y Piperidis, Stelios
%Y Bel, Núria
%Y van den Heuvel, Henk
%Y Ide, Nancy
%Y Krek, Simon
%Y Toral, Antonio
%S Proceedings of the Fifteenth Language Resources and Evaluation Conference
%D 2026
%8 May
%I ELRA Language Resource Association
%C Palma de Mallorca, Spain
%F saeed-etal-2026-surfacing
%X Large language models (LLMs) are widely deployed for open-ended communication, yet most bias evaluations still rely on English, classification-style tasks. We introduce , a new multilingual, debate-style benchmark designed to reveal how narrative bias appears in realistic generative settings. Our dataset includes 8,400 structured debate prompts spanning four sensitive domains – Women’s Rights, Backwardness, Terrorism, and Religion – across seven languages ranging from high-resource (English, Chinese) to low-resource (Swahili, Nigerian Pidgin). Using four flagship models (GPT-4o, Claude 3.5 Haiku, DeepSeek-Chat, and LLaMA-3-70B), we generate over 100,000 debate responses and automatically classify which demographic groups are assigned stereotyped versus modern roles. Results show that all models reproduce entrenched stereotypes despite safety alignment: Arabs are overwhelmingly linked to Terrorism and Religion (\geq89%), Africans to socioeconomic “backwardness” (up to 77%), and Western groups are consistently framed as modern or progressive. Biases grow sharply in lower-resource languages, revealing that alignment trained primarily in English does not generalize globally. Our findings highlight a persistent divide in multilingual fairness: current alignment methods reduce explicit toxicity but fail to prevent biased outputs in open-ended contexts. We release our benchmark and analysis framework to support the next generation of multilingual bias evaluation and safer, culturally inclusive model alignment
%R 10.63317/4u9x3z4g8jfk
%U https://aclanthology.org/2026.lrec-1.643/
%U https://doi.org/10.63317/4u9x3z4g8jfk
%P 8106-8121

Markdown (Informal)

[Surfacing Subtle Stereotypes: A Multilingual, Debate-Oriented Evaluation of Modern LLMs](https://aclanthology.org/2026.lrec-1.643/) (Saeed et al., LREC 2026)

• Surfacing Subtle Stereotypes: A Multilingual, Debate-Oriented Evaluation of Modern LLMs (Saeed et al., LREC 2026)

ACL

• Muhammed Yahia Gaffar Saeed, Muhammad Abdul-Mageed, and Shady Shehata. 2026. Surfacing Subtle Stereotypes: A Multilingual, Debate-Oriented Evaluation of Modern LLMs. In Proceedings of the Fifteenth Language Resources and Evaluation Conference, pages 8106–8121, Palma de Mallorca, Spain. ELRA Language Resource Association.