@inproceedings{yeom-etal-2026-mob,
title = "{MOB}-{KV}: Mixture-of-Basis for Key-Value Cache Compression",
author = "Yeom, Jinseop and
Kim, Taehyeon and
Park, Eunhyeok",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.2017/",
pages = "45939--45958",
ISBN = "979-8-89176-459-0",
abstract = "In this paper, we present MOB-KV, a Mixture-of-Basis KV cache compression method that makes adaptive low-rank key compression practical without runtime decomposition. We first analyze KV cache compression through attention output error and reveal a clear asymmetry between keys and values: pre-RoPE keys benefit from preserving dominant low-rank structure while quantizing low-energy residuals, whereas values are better compressed by direct quantization. This analysis motivates an asymmetric design that applies low-rank projection with residual quantization to keys and direct low-bit quantization to values. However, obtaining accurate input-adaptive low-rank bases with online SVD is prohibitively expensive for serving. MOB-KV addresses this by replacing online SVD with a reusable dictionary of offline-constructed key bases, transforming adaptive compression into lightweight basis selection and projection. This design preserves the accuracy benefit of local low-rank subspaces while avoiding the latency bottleneck of runtime decomposition. Across reasoning and long-context benchmarks, MOB-KV substantially reduces KV cache memory, preserves near full-precision accuracy, and establishes a stronger memory{--}accuracy{--}latency Pareto frontier than existing quantization-only, low-rank-only, and online SVD-based baselines."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yeom-etal-2026-mob">
<titleInfo>
<title>MOB-KV: Mixture-of-Basis for Key-Value Cache Compression</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jinseop</namePart>
<namePart type="family">Yeom</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Taehyeon</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Eunhyeok</namePart>
<namePart type="family">Park</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>In this paper, we present MOB-KV, a Mixture-of-Basis KV cache compression method that makes adaptive low-rank key compression practical without runtime decomposition. We first analyze KV cache compression through attention output error and reveal a clear asymmetry between keys and values: pre-RoPE keys benefit from preserving dominant low-rank structure while quantizing low-energy residuals, whereas values are better compressed by direct quantization. This analysis motivates an asymmetric design that applies low-rank projection with residual quantization to keys and direct low-bit quantization to values. However, obtaining accurate input-adaptive low-rank bases with online SVD is prohibitively expensive for serving. MOB-KV addresses this by replacing online SVD with a reusable dictionary of offline-constructed key bases, transforming adaptive compression into lightweight basis selection and projection. This design preserves the accuracy benefit of local low-rank subspaces while avoiding the latency bottleneck of runtime decomposition. Across reasoning and long-context benchmarks, MOB-KV substantially reduces KV cache memory, preserves near full-precision accuracy, and establishes a stronger memory–accuracy–latency Pareto frontier than existing quantization-only, low-rank-only, and online SVD-based baselines.</abstract>
<identifier type="citekey">yeom-etal-2026-mob</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2017/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>45939</start>
<end>45958</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T MOB-KV: Mixture-of-Basis for Key-Value Cache Compression
%A Yeom, Jinseop
%A Kim, Taehyeon
%A Park, Eunhyeok
%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 yeom-etal-2026-mob
%X In this paper, we present MOB-KV, a Mixture-of-Basis KV cache compression method that makes adaptive low-rank key compression practical without runtime decomposition. We first analyze KV cache compression through attention output error and reveal a clear asymmetry between keys and values: pre-RoPE keys benefit from preserving dominant low-rank structure while quantizing low-energy residuals, whereas values are better compressed by direct quantization. This analysis motivates an asymmetric design that applies low-rank projection with residual quantization to keys and direct low-bit quantization to values. However, obtaining accurate input-adaptive low-rank bases with online SVD is prohibitively expensive for serving. MOB-KV addresses this by replacing online SVD with a reusable dictionary of offline-constructed key bases, transforming adaptive compression into lightweight basis selection and projection. This design preserves the accuracy benefit of local low-rank subspaces while avoiding the latency bottleneck of runtime decomposition. Across reasoning and long-context benchmarks, MOB-KV substantially reduces KV cache memory, preserves near full-precision accuracy, and establishes a stronger memory–accuracy–latency Pareto frontier than existing quantization-only, low-rank-only, and online SVD-based baselines.
%U https://aclanthology.org/2026.emnlp-main.2017/
%P 45939-45958
Markdown (Informal)
[MOB-KV: Mixture-of-Basis for Key-Value Cache Compression](https://aclanthology.org/2026.emnlp-main.2017/) (Yeom et al., EMNLP 2026)
ACL
- Jinseop Yeom, Taehyeon Kim, and Eunhyeok Park. 2026. MOB-KV: Mixture-of-Basis for Key-Value Cache Compression. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 45939–45958, Budapest, Hungary. Association for Computational Linguistics.