<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kim-etal-2026-surrogatekv">
    <titleInfo>
        <title>SurrogateKV: Representation-Preserving KV Cache Compression for Long-Context LLMs</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Junwon</namePart>
        <namePart type="family">Kim</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Junghyun</namePart>
        <namePart type="family">Ryu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Farid</namePart>
        <namePart type="family">Talibli</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Jungmin</namePart>
        <namePart type="family">So</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Youngjae</namePart>
        <namePart type="family">Kim</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>KV cache compression is essential for long-context LLM inference. Most token-level KV compressors reduce cache length by retaining selected token KV entries and evicting the rest. Once evicted, those spans no longer receive attention, even though some may still carry information useful enough to preserve in compact form. We propose SurrogateKV, a training-free, runtime-only method that reframes KV compression as region-level representation allocation. Under the same KV-slot budget, SurrogateKV assigns historical regions to raw KV, surrogate KV, or drop. It does not train a compressor, modify model weights, or introduce a new attention operator. After prefill, SurrogateKV constructs lightweight surrogate entries as ordinary KV pairs, keeping their information available during standard attention decoding. This representation decision is orthogonal to layer- and head-wise budget allocation. We evaluate SurrogateKV-Snap, SurrogateKV-Dynamic, and SurrogateKV-Ada, obtained by applying SurrogateKV to SnapKV, DynamicKV, and Ada-KV, respectively. As representative results, SurrogateKV-Snap raises Needle-in-a-Haystack (NIAH) retrieval accuracy from 87.51 to 98.84 at 128 slots per layer. SurrogateKV-Ada raises the LongBench average score from 40.77 to 41.26 at 512 slots per layer while preserving 98.4% of full-cache performance with only 10.5% of the prefill KV cache. Code is available at https://github.com/kimjunewon/SurrogateKV.</abstract>
    <identifier type="citekey">kim-etal-2026-surrogatekv</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-emnlp.1580/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>33643</start>
            <end>33660</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T SurrogateKV: Representation-Preserving KV Cache Compression for Long-Context LLMs
%A Kim, Junwon
%A Ryu, Junghyun
%A Talibli, Farid
%A So, Jungmin
%A Kim, Youngjae
%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 kim-etal-2026-surrogatekv
%X KV cache compression is essential for long-context LLM inference. Most token-level KV compressors reduce cache length by retaining selected token KV entries and evicting the rest. Once evicted, those spans no longer receive attention, even though some may still carry information useful enough to preserve in compact form. We propose SurrogateKV, a training-free, runtime-only method that reframes KV compression as region-level representation allocation. Under the same KV-slot budget, SurrogateKV assigns historical regions to raw KV, surrogate KV, or drop. It does not train a compressor, modify model weights, or introduce a new attention operator. After prefill, SurrogateKV constructs lightweight surrogate entries as ordinary KV pairs, keeping their information available during standard attention decoding. This representation decision is orthogonal to layer- and head-wise budget allocation. We evaluate SurrogateKV-Snap, SurrogateKV-Dynamic, and SurrogateKV-Ada, obtained by applying SurrogateKV to SnapKV, DynamicKV, and Ada-KV, respectively. As representative results, SurrogateKV-Snap raises Needle-in-a-Haystack (NIAH) retrieval accuracy from 87.51 to 98.84 at 128 slots per layer. SurrogateKV-Ada raises the LongBench average score from 40.77 to 41.26 at 512 slots per layer while preserving 98.4% of full-cache performance with only 10.5% of the prefill KV cache. Code is available at https://github.com/kimjunewon/SurrogateKV.
%U https://aclanthology.org/2026.findings-emnlp.1580/
%P 33643-33660

Markdown (Informal)

[SurrogateKV: Representation-Preserving KV Cache Compression for Long-Context LLMs](https://aclanthology.org/2026.findings-emnlp.1580/) (Kim et al., Findings 2026)

• SurrogateKV: Representation-Preserving KV Cache Compression for Long-Context LLMs (Kim et al., Findings 2026)

ACL

• Junwon Kim, Junghyun Ryu, Farid Talibli, Jungmin So, and Youngjae Kim. 2026. SurrogateKV: Representation-Preserving KV Cache Compression for Long-Context LLMs. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 33643–33660, Budapest, Hungary. Association for Computational Linguistics.