<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-etal-2026-multilingual">
    <titleInfo>
        <title>Multilingual Safety Signals Are Multi-Layered: Filtering Safety-Degrading Data for Safer LLMs</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Jiakun</namePart>
        <namePart type="family">Li</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Guowei</namePart>
        <namePart type="family">Song</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Sijia</namePart>
        <namePart type="family">Li</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Xingwei</namePart>
        <namePart type="family">He</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Hongzheng</namePart>
        <namePart type="family">Chai</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yuan</namePart>
        <namePart type="family">Yuan</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>Preserving safety alignment during large language models fine-tuning is critical, however, recent studies have demonstrated that even benign fine-tuning data may contain safety-degrading samples that silently undermine such alignment. Existing approaches typically identify safety-degrading samples using representations from a single safety-sensitive layer. However, whether a single safety-sensitive layer is sufficient for identifying such samples in multilingual settings remains unclear, where representation patterns may vary across languages. Our cross-lingual analysis shows that sensitive layers are only partially shared across languages and that safety-relevant signals often span multiple layers. Motivated by these observations, we propose MMSafe, a multi-layer framework for multilingual safety-degrading data identification that jointly leverages shared and language-specific sensitive layers. Extensive experiments across multiple models, languages, and safety benchmarks demonstrate that MMSafe reduces the average harmful-response ratio by 60% compared with random filtering and outperforms the strongest single-layer baseline, demonstrating the effectiveness of multi-layer modeling for robust multilingual safety alignment. Code is available at https://anonymous.4open.science/r/MMSAFE.</abstract>
    <identifier type="citekey">li-etal-2026-multilingual</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.1492/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>34105</start>
            <end>34120</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Multilingual Safety Signals Are Multi-Layered: Filtering Safety-Degrading Data for Safer LLMs
%A Li, Jiakun
%A Song, Guowei
%A Li, Sijia
%A He, Xingwei
%A Chai, Hongzheng
%A Yuan, Yuan
%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 li-etal-2026-multilingual
%X Preserving safety alignment during large language models fine-tuning is critical, however, recent studies have demonstrated that even benign fine-tuning data may contain safety-degrading samples that silently undermine such alignment. Existing approaches typically identify safety-degrading samples using representations from a single safety-sensitive layer. However, whether a single safety-sensitive layer is sufficient for identifying such samples in multilingual settings remains unclear, where representation patterns may vary across languages. Our cross-lingual analysis shows that sensitive layers are only partially shared across languages and that safety-relevant signals often span multiple layers. Motivated by these observations, we propose MMSafe, a multi-layer framework for multilingual safety-degrading data identification that jointly leverages shared and language-specific sensitive layers. Extensive experiments across multiple models, languages, and safety benchmarks demonstrate that MMSafe reduces the average harmful-response ratio by 60% compared with random filtering and outperforms the strongest single-layer baseline, demonstrating the effectiveness of multi-layer modeling for robust multilingual safety alignment. Code is available at https://anonymous.4open.science/r/MMSAFE.
%U https://aclanthology.org/2026.emnlp-main.1492/
%P 34105-34120

Markdown (Informal)

[Multilingual Safety Signals Are Multi-Layered: Filtering Safety-Degrading Data for Safer LLMs](https://aclanthology.org/2026.emnlp-main.1492/) (Li et al., EMNLP 2026)

• Multilingual Safety Signals Are Multi-Layered: Filtering Safety-Degrading Data for Safer LLMs (Li et al., EMNLP 2026)

ACL

• Jiakun Li, Guowei Song, Sijia Li, Xingwei He, Hongzheng Chai, and Yuan Yuan. 2026. Multilingual Safety Signals Are Multi-Layered: Filtering Safety-Degrading Data for Safer LLMs. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 34105–34120, Budapest, Hungary. Association for Computational Linguistics.