<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="harrasse-etal-2026-tracing">
    <titleInfo>
        <title>Tracing Multilingual Representations in LLMs with Cross-Layer Transcoders</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Abir</namePart>
        <namePart type="family">Harrasse</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Florent</namePart>
        <namePart type="family">Draye</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Punya</namePart>
        <namePart type="given">Syon</namePart>
        <namePart type="family">Pandey</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Zhijing</namePart>
        <namePart type="family">Jin</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Bernhard</namePart>
        <namePart type="family">Schölkopf</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>Multilingual Large Language Models (LLMs) can process many languages, yet how they internally represent this diversity remains unclear. Do they form shared multilingual representations with language-specific decoding, and if so, why do multilingual performance gaps persist? To address this question, we train models on different multilingual mixtures and analyze their internal mechanisms using Cross-Layer Transcoders (CLTs) and Attribution Graphs. Our results reveal a consistent three-phase structure across models: early layers encode language-specific information, middle layers converge to highly shared multilingual representations, and final layers re-specialize for language-specific decoding.\par Training models without English produces the same shared multilingual structures, suggesting that these representations emerge independently of English dominance. We further show that decoding relies on a small set of high-frequency late-layer features that linearly encode language identity from earlier layers, and that intervening on these features allows one language to be suppressed and another substituted. Finally, through a Model-Diffing experiment, we show that non-English failures arise from weak late-layer decoding features, weak middle-layer clusters, and tokenizer fragmentation that forces early layers to specialize in word reassembly. Finetuning strengthens these features and their connections, improving token assembly and language-specific decoding, providing a mechanistic explanation for multilingual performance gaps.</abstract>
    <identifier type="citekey">harrasse-etal-2026-tracing</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.2222/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>50565</start>
            <end>50601</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Tracing Multilingual Representations in LLMs with Cross-Layer Transcoders
%A Harrasse, Abir
%A Draye, Florent
%A Pandey, Punya Syon
%A Jin, Zhijing
%A Schölkopf, Bernhard
%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 harrasse-etal-2026-tracing
%X Multilingual Large Language Models (LLMs) can process many languages, yet how they internally represent this diversity remains unclear. Do they form shared multilingual representations with language-specific decoding, and if so, why do multilingual performance gaps persist? To address this question, we train models on different multilingual mixtures and analyze their internal mechanisms using Cross-Layer Transcoders (CLTs) and Attribution Graphs. Our results reveal a consistent three-phase structure across models: early layers encode language-specific information, middle layers converge to highly shared multilingual representations, and final layers re-specialize for language-specific decoding.\par Training models without English produces the same shared multilingual structures, suggesting that these representations emerge independently of English dominance. We further show that decoding relies on a small set of high-frequency late-layer features that linearly encode language identity from earlier layers, and that intervening on these features allows one language to be suppressed and another substituted. Finally, through a Model-Diffing experiment, we show that non-English failures arise from weak late-layer decoding features, weak middle-layer clusters, and tokenizer fragmentation that forces early layers to specialize in word reassembly. Finetuning strengthens these features and their connections, improving token assembly and language-specific decoding, providing a mechanistic explanation for multilingual performance gaps.
%U https://aclanthology.org/2026.emnlp-main.2222/
%P 50565-50601

Markdown (Informal)

[Tracing Multilingual Representations in LLMs with Cross-Layer Transcoders](https://aclanthology.org/2026.emnlp-main.2222/) (Harrasse et al., EMNLP 2026)

• Tracing Multilingual Representations in LLMs with Cross-Layer Transcoders (Harrasse et al., EMNLP 2026)

ACL

• Abir Harrasse, Florent Draye, Punya Syon Pandey, Zhijing Jin, and Bernhard Schölkopf. 2026. Tracing Multilingual Representations in LLMs with Cross-Layer Transcoders. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 50565–50601, Budapest, Hungary. Association for Computational Linguistics.