@inproceedings{wu-etal-2026-centralkv,
title = "{C}entral{KV}: Compressing Long-Context {LLM}s via Attention-Graph Centrality",
author = "Wu, Yusen and
Yinjun, Huang and
Tan, Jia Yee and
Li, Hao and
Guo, Rongfeng and
Fan, Liang and
Luo, Jiachen and
Dong, Guangyuan and
Xiang, Sike",
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.382/",
pages = "8727--8748",
ISBN = "979-8-89176-459-0",
abstract = "Long-context inference in large language models is limited by the linear memory growth of the key-value (KV) cache. Existing eviction methods typically retain tokens with high accumulated attention or remove semantic redundancy, but may discard low-frequency tokens that connect distant dependencies. We propose CentralKV, a training-free framework that models attention as a dynamic sparse causal graph. Its Stream-Influence algorithm updates token centrality online and preserves structurally important bridge tokens during KV compression. The attention graph is constructed directly from the attention operation without separate similarity computation. Experiments on LongBench show that CentralKV improves over H2O, SnapKV, and GraphKV under strict physical memory constraints, particularly on reasoning-intensive tasks, with minimal latency overhead."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wu-etal-2026-centralkv">
<titleInfo>
<title>CentralKV: Compressing Long-Context LLMs via Attention-Graph Centrality</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yusen</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Huang</namePart>
<namePart type="family">Yinjun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jia</namePart>
<namePart type="given">Yee</namePart>
<namePart type="family">Tan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hao</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rongfeng</namePart>
<namePart type="family">Guo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Liang</namePart>
<namePart type="family">Fan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiachen</namePart>
<namePart type="family">Luo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Guangyuan</namePart>
<namePart type="family">Dong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sike</namePart>
<namePart type="family">Xiang</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>Long-context inference in large language models is limited by the linear memory growth of the key-value (KV) cache. Existing eviction methods typically retain tokens with high accumulated attention or remove semantic redundancy, but may discard low-frequency tokens that connect distant dependencies. We propose CentralKV, a training-free framework that models attention as a dynamic sparse causal graph. Its Stream-Influence algorithm updates token centrality online and preserves structurally important bridge tokens during KV compression. The attention graph is constructed directly from the attention operation without separate similarity computation. Experiments on LongBench show that CentralKV improves over H2O, SnapKV, and GraphKV under strict physical memory constraints, particularly on reasoning-intensive tasks, with minimal latency overhead.</abstract>
<identifier type="citekey">wu-etal-2026-centralkv</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.382/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>8727</start>
<end>8748</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T CentralKV: Compressing Long-Context LLMs via Attention-Graph Centrality
%A Wu, Yusen
%A Yinjun, Huang
%A Tan, Jia Yee
%A Li, Hao
%A Guo, Rongfeng
%A Fan, Liang
%A Luo, Jiachen
%A Dong, Guangyuan
%A Xiang, Sike
%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 wu-etal-2026-centralkv
%X Long-context inference in large language models is limited by the linear memory growth of the key-value (KV) cache. Existing eviction methods typically retain tokens with high accumulated attention or remove semantic redundancy, but may discard low-frequency tokens that connect distant dependencies. We propose CentralKV, a training-free framework that models attention as a dynamic sparse causal graph. Its Stream-Influence algorithm updates token centrality online and preserves structurally important bridge tokens during KV compression. The attention graph is constructed directly from the attention operation without separate similarity computation. Experiments on LongBench show that CentralKV improves over H2O, SnapKV, and GraphKV under strict physical memory constraints, particularly on reasoning-intensive tasks, with minimal latency overhead.
%U https://aclanthology.org/2026.emnlp-main.382/
%P 8727-8748
Markdown (Informal)
[CentralKV: Compressing Long-Context LLMs via Attention-Graph Centrality](https://aclanthology.org/2026.emnlp-main.382/) (Wu et al., EMNLP 2026)
ACL
- Yusen Wu, Huang Yinjun, Jia Yee Tan, Hao Li, Rongfeng Guo, Liang Fan, Jiachen Luo, Guangyuan Dong, and Sike Xiang. 2026. CentralKV: Compressing Long-Context LLMs via Attention-Graph Centrality. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 8727–8748, Budapest, Hungary. Association for Computational Linguistics.