<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="karmim-etal-2026-multi">
    <titleInfo>
        <title>Multi-Hop Knowledge Composition is Bound by Pretraining Exposure</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Yannis</namePart>
        <namePart type="family">Karmim</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Luis</namePart>
        <namePart type="family">Martí</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Djamé</namePart>
        <namePart type="family">Seddah</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Valentin</namePart>
        <namePart type="family">Barriere</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 fail at implicit multi-hop reasoning: a model answers “When was born?” and “Who is ‘s closest friend?” correctly but fails on “When was ‘s closest friend born?” in a single forward pass, even when both facts are perfectly memorized and individually retrievable. We study this failure in a controlled natural language setting with a strict separation between individuals exposed to compositional contexts during pretraining and those that never appear in any such context. We confirm that compositional failure persists even at 97% 1-hop accuracy, establishing the gap as a pretraining failure rather than a knowledge absence. We propose and test nine data-centric augmentation formats and find that compositional pretraining transfers to unseen questions for exposed individuals, but never to individuals absent from compositional pretraining, suggesting that exposure to compositional contexts during pretraining is a necessary condition for implicit multi-hop reasoning. Our code is \hrefhttps://github.com/ykrmm/composition-boundavailable online.</abstract>
    <identifier type="citekey">karmim-etal-2026-multi</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-emnlp.2144/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>45815</start>
            <end>45825</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Multi-Hop Knowledge Composition is Bound by Pretraining Exposure
%A Karmim, Yannis
%A Martí, Luis
%A Seddah, Djamé
%A Barriere, Valentin
%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 karmim-etal-2026-multi
%X Large Language Models fail at implicit multi-hop reasoning: a model answers “When was born?” and “Who is ‘s closest friend?” correctly but fails on “When was ‘s closest friend born?” in a single forward pass, even when both facts are perfectly memorized and individually retrievable. We study this failure in a controlled natural language setting with a strict separation between individuals exposed to compositional contexts during pretraining and those that never appear in any such context. We confirm that compositional failure persists even at 97% 1-hop accuracy, establishing the gap as a pretraining failure rather than a knowledge absence. We propose and test nine data-centric augmentation formats and find that compositional pretraining transfers to unseen questions for exposed individuals, but never to individuals absent from compositional pretraining, suggesting that exposure to compositional contexts during pretraining is a necessary condition for implicit multi-hop reasoning. Our code is \hrefhttps://github.com/ykrmm/composition-boundavailable online.
%U https://aclanthology.org/2026.findings-emnlp.2144/
%P 45815-45825

Markdown (Informal)

[Multi-Hop Knowledge Composition is Bound by Pretraining Exposure](https://aclanthology.org/2026.findings-emnlp.2144/) (Karmim et al., Findings 2026)

• Multi-Hop Knowledge Composition is Bound by Pretraining Exposure (Karmim et al., Findings 2026)

ACL

• Yannis Karmim, Luis Martí, Djamé Seddah, and Valentin Barriere. 2026. Multi-Hop Knowledge Composition is Bound by Pretraining Exposure. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 45815–45825, Budapest, Hungary. Association for Computational Linguistics.