<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ebing-glavas-2026-watch">
    <titleInfo>
        <title>Watch Your Norms: On Test-Time Expert Reduction in Multilingual MoEs</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Benedikt</namePart>
        <namePart type="family">Ebing</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Goran</namePart>
        <namePart type="family">Glavaš</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>Test-time expert reduction (top-k reduction) is a popular way to cut inference cost in Mixture-of-Experts (MoE) language models, avoiding the expensive model retraining. Existing work, however, analyzes the effects of top-k reduction almost exclusively on English benchmarks. In this work, we provide the first systematic multilingual study of test-time top-k reduction across three contemporary MoE models, 100 languages, and 5 downstream tasks. We show that down to 75% of retained experts, the models degrade gracefully, but overall retention collapses sharply at low k and widens a structural multilingual gap: low-resource languages degrade much faster than high-resource ones. We trace the gap to a renormalization-induced norm overshoot in the MoE block, where the renormalization over the surviving expert weights inflates the MoE contribution to the residual stream. This inflation harms low-resource languages disproportionately due to their smaller residual norms. As a remedy, we propose two training-free and latency-free interventions: (i) a grid-searched global scalar to rescale the representations output by the MoE layer (global \ensuremathα) and (ii) a parameter-free per-token correction (token \ensuremathα) of expert weights derived from the routing. The proposed interventions substantially improve overall retention and reduce the multilingual gap.</abstract>
    <identifier type="citekey">ebing-glavas-2026-watch</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.2389/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>54296</start>
            <end>54341</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Watch Your Norms: On Test-Time Expert Reduction in Multilingual MoEs
%A Ebing, Benedikt
%A Glavaš, Goran
%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 ebing-glavas-2026-watch
%X Test-time expert reduction (top-k reduction) is a popular way to cut inference cost in Mixture-of-Experts (MoE) language models, avoiding the expensive model retraining. Existing work, however, analyzes the effects of top-k reduction almost exclusively on English benchmarks. In this work, we provide the first systematic multilingual study of test-time top-k reduction across three contemporary MoE models, 100 languages, and 5 downstream tasks. We show that down to 75% of retained experts, the models degrade gracefully, but overall retention collapses sharply at low k and widens a structural multilingual gap: low-resource languages degrade much faster than high-resource ones. We trace the gap to a renormalization-induced norm overshoot in the MoE block, where the renormalization over the surviving expert weights inflates the MoE contribution to the residual stream. This inflation harms low-resource languages disproportionately due to their smaller residual norms. As a remedy, we propose two training-free and latency-free interventions: (i) a grid-searched global scalar to rescale the representations output by the MoE layer (global \ensuremathα) and (ii) a parameter-free per-token correction (token \ensuremathα) of expert weights derived from the routing. The proposed interventions substantially improve overall retention and reduce the multilingual gap.
%U https://aclanthology.org/2026.emnlp-main.2389/
%P 54296-54341

Markdown (Informal)

[Watch Your Norms: On Test-Time Expert Reduction in Multilingual MoEs](https://aclanthology.org/2026.emnlp-main.2389/) (Ebing & Glavaš, EMNLP 2026)

• Watch Your Norms: On Test-Time Expert Reduction in Multilingual MoEs (Ebing & Glavaš, EMNLP 2026)

ACL

• Benedikt Ebing and Goran Glavaš. 2026. Watch Your Norms: On Test-Time Expert Reduction in Multilingual MoEs. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 54296–54341, Budapest, Hungary. Association for Computational Linguistics.