<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="thaler-etal-2026-far">
    <titleInfo>
        <title>How Far Can Bias Go? Tracing Bias from Pre-Training Data to Alignment</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Marion</namePart>
        <namePart type="family">Thaler</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Abdullatif</namePart>
        <namePart type="family">Köksal</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Alina</namePart>
        <namePart type="family">Leidinger</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Anna</namePart>
        <namePart type="family">Anna Korhonen</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Hinrich</namePart>
        <namePart type="family">Schütze</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>As LLMs are increasingly integrated into user-facing applications, addressing biases that perpetuate societal inequalities is crucial. While much work has gone into measuring and mitigating biases, fewer studies have investigated their origins. Therefore, this study examines the propagation of representational gender-occupation bias from pre-training data to LLM generations. Using zero-shot prompting and token co-occurrence analyses, we explore how biases in the pre-training data influence model generations. Our findings reveal that representational biases present in the pre-training data are amplified in the model generations, regardless of hyperparameters and prompting type. By comparing gender representation in the pre-training data with real-world distributions, our research highlights discrepancies between the data and the model, underscoring the importance of further work in mitigating bias at the data level.</abstract>
    <identifier type="citekey">thaler-etal-2026-far</identifier>
    <identifier type="doi">10.63317/4zeoky6waeng</identifier>
    <location>
        <url>https://aclanthology.org/2026.lrec-1.315/</url>
    </location>
    <part>
        <date>2026-05</date>
        <extent unit="page">
            <start>3975</start>
            <end>3995</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T How Far Can Bias Go? Tracing Bias from Pre-Training Data to Alignment
%A Thaler, Marion
%A Köksal, Abdullatif
%A Leidinger, Alina
%A Anna Korhonen, Anna
%A Schütze, Hinrich
%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 thaler-etal-2026-far
%X As LLMs are increasingly integrated into user-facing applications, addressing biases that perpetuate societal inequalities is crucial. While much work has gone into measuring and mitigating biases, fewer studies have investigated their origins. Therefore, this study examines the propagation of representational gender-occupation bias from pre-training data to LLM generations. Using zero-shot prompting and token co-occurrence analyses, we explore how biases in the pre-training data influence model generations. Our findings reveal that representational biases present in the pre-training data are amplified in the model generations, regardless of hyperparameters and prompting type. By comparing gender representation in the pre-training data with real-world distributions, our research highlights discrepancies between the data and the model, underscoring the importance of further work in mitigating bias at the data level.
%R 10.63317/4zeoky6waeng
%U https://aclanthology.org/2026.lrec-1.315/
%U https://doi.org/10.63317/4zeoky6waeng
%P 3975-3995

Markdown (Informal)

[How Far Can Bias Go? Tracing Bias from Pre-Training Data to Alignment](https://aclanthology.org/2026.lrec-1.315/) (Thaler et al., LREC 2026)

• How Far Can Bias Go? Tracing Bias from Pre-Training Data to Alignment (Thaler et al., LREC 2026)

ACL

• Marion Thaler, Abdullatif Köksal, Alina Leidinger, Anna Anna Korhonen, and Hinrich Schütze. 2026. How Far Can Bias Go? Tracing Bias from Pre-Training Data to Alignment. In Proceedings of the Fifteenth Language Resources and Evaluation Conference, pages 3975–3995, Palma de Mallorca, Spain. ELRA Language Resource Association.