<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ding-etal-2026-meki">
    <titleInfo>
        <title>MeKi: Memory-based Expert Knowledge Injection for Efficient LLM Scaling</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Ning</namePart>
        <namePart type="family">Ding</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Fangcheng</namePart>
        <namePart type="family">Liu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Kyungrae</namePart>
        <namePart type="family">Kim</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Linji</namePart>
        <namePart type="family">Hao</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Kyenghun</namePart>
        <namePart type="family">Lee</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Hyeonmok</namePart>
        <namePart type="family">Ko</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yehui</namePart>
        <namePart type="family">Tang</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>Scaling Large Language Models (LLMs) typically relies on increasing the number of parameters or test-time computations to boost performance. However, these strategies are impractical for edge device deployment due to limited RAM and NPU resources. Despite hardware constraints, deploying performant LLM on edge devices such as smartphone remains crucial for user experience. To address this, we propose MeKi (Memory-based Expert Knowledge Injection), a novel system that scales LLM capacity via storage space rather than FLOPs. MiKi equips each Transformer layer with token-level memory experts that injects pre-stored semantic knowledge into the generation process. To bridge the gap between training capacity and inference efficiency, we employ a re-parameterization strategy to fold parameter matrices used during training into a compact static lookup table. By offloading the knowledge to ROM, MiKi decouples model capacity from computational cost, introducing zero inference latency overhead. Extensive experiments demonstrate that MiKi significantly outperforms dense LLM baselines with identical inference speed, validating the effectiveness of memory-based scaling paradigm for on-device LLMs. Code is available at https://github.com/ningding-o/MeKi.</abstract>
    <identifier type="citekey">ding-etal-2026-meki</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.2661/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>60572</start>
            <end>60584</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T MeKi: Memory-based Expert Knowledge Injection for Efficient LLM Scaling
%A Ding, Ning
%A Liu, Fangcheng
%A Kim, Kyungrae
%A Hao, Linji
%A Lee, Kyenghun
%A Ko, Hyeonmok
%A Tang, Yehui
%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 ding-etal-2026-meki
%X Scaling Large Language Models (LLMs) typically relies on increasing the number of parameters or test-time computations to boost performance. However, these strategies are impractical for edge device deployment due to limited RAM and NPU resources. Despite hardware constraints, deploying performant LLM on edge devices such as smartphone remains crucial for user experience. To address this, we propose MeKi (Memory-based Expert Knowledge Injection), a novel system that scales LLM capacity via storage space rather than FLOPs. MiKi equips each Transformer layer with token-level memory experts that injects pre-stored semantic knowledge into the generation process. To bridge the gap between training capacity and inference efficiency, we employ a re-parameterization strategy to fold parameter matrices used during training into a compact static lookup table. By offloading the knowledge to ROM, MiKi decouples model capacity from computational cost, introducing zero inference latency overhead. Extensive experiments demonstrate that MiKi significantly outperforms dense LLM baselines with identical inference speed, validating the effectiveness of memory-based scaling paradigm for on-device LLMs. Code is available at https://github.com/ningding-o/MeKi.
%U https://aclanthology.org/2026.emnlp-main.2661/
%P 60572-60584

Markdown (Informal)

[MeKi: Memory-based Expert Knowledge Injection for Efficient LLM Scaling](https://aclanthology.org/2026.emnlp-main.2661/) (Ding et al., EMNLP 2026)

• MeKi: Memory-based Expert Knowledge Injection for Efficient LLM Scaling (Ding et al., EMNLP 2026)

ACL

• Ning Ding, Fangcheng Liu, Kyungrae Kim, Linji Hao, Kyenghun Lee, Hyeonmok Ko, and Yehui Tang. 2026. MeKi: Memory-based Expert Knowledge Injection for Efficient LLM Scaling. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 60572–60584, Budapest, Hungary. Association for Computational Linguistics.