@inproceedings{xie-etal-2026-aura,
title = "{AURA}-{KV}: Adaptive Subspace Rank Approximation for Efficient Long-Context {LLM} Inference",
author = "Xie, Jincheng and
Shi, Hao and
Jiang, Zhengyi and
Huang, Zhongyi and
Song, Linqi and
Hou, Hanxu",
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.2250/",
pages = "51179--51203",
ISBN = "979-8-89176-459-0",
abstract = "KV caching is essential for efficient autoregressive decoding, but its memory footprint grows linearly with context length and quickly becomes a bottleneck for long-context large language models (LLMs). We study whether feature-dimension compression can serve as an orthogonal second stage for KV-cache reduction, and introduce AURA-KV, an adaptive-rank subspace approximation method for KV caches. AURA-KV is training-free: it decomposes post-RoPE key and value caches into compact bases and coefficients, supports incremental decoding, and allocates rank adaptively across layers under a global memory budget. Because it compresses the feature dimension rather than selecting tokens or lowering precision, it can be combined with token eviction and quantization methods. Across LongBench, RULER, and InfiniteBench, AURA-KV preserves long-context performance on Llama-3-8B, Mistral-7B, and LongChat-7B under moderate compression and degrades gracefully under more aggressive budgets. In latency measurements, AURA-KV improves per-layer decoding speed by up to $2.75 \times$, and by up to $3.23 \times$ when combined with token-level cache reduction, with a small one-time prefill overhead."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xie-etal-2026-aura">
<titleInfo>
<title>AURA-KV: Adaptive Subspace Rank Approximation for Efficient Long-Context LLM Inference</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jincheng</namePart>
<namePart type="family">Xie</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hao</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhengyi</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhongyi</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Linqi</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hanxu</namePart>
<namePart type="family">Hou</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>KV caching is essential for efficient autoregressive decoding, but its memory footprint grows linearly with context length and quickly becomes a bottleneck for long-context large language models (LLMs). We study whether feature-dimension compression can serve as an orthogonal second stage for KV-cache reduction, and introduce AURA-KV, an adaptive-rank subspace approximation method for KV caches. AURA-KV is training-free: it decomposes post-RoPE key and value caches into compact bases and coefficients, supports incremental decoding, and allocates rank adaptively across layers under a global memory budget. Because it compresses the feature dimension rather than selecting tokens or lowering precision, it can be combined with token eviction and quantization methods. Across LongBench, RULER, and InfiniteBench, AURA-KV preserves long-context performance on Llama-3-8B, Mistral-7B, and LongChat-7B under moderate compression and degrades gracefully under more aggressive budgets. In latency measurements, AURA-KV improves per-layer decoding speed by up to 2.75 \times, and by up to 3.23 \times when combined with token-level cache reduction, with a small one-time prefill overhead.</abstract>
<identifier type="citekey">xie-etal-2026-aura</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2250/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>51179</start>
<end>51203</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T AURA-KV: Adaptive Subspace Rank Approximation for Efficient Long-Context LLM Inference
%A Xie, Jincheng
%A Shi, Hao
%A Jiang, Zhengyi
%A Huang, Zhongyi
%A Song, Linqi
%A Hou, Hanxu
%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 xie-etal-2026-aura
%X KV caching is essential for efficient autoregressive decoding, but its memory footprint grows linearly with context length and quickly becomes a bottleneck for long-context large language models (LLMs). We study whether feature-dimension compression can serve as an orthogonal second stage for KV-cache reduction, and introduce AURA-KV, an adaptive-rank subspace approximation method for KV caches. AURA-KV is training-free: it decomposes post-RoPE key and value caches into compact bases and coefficients, supports incremental decoding, and allocates rank adaptively across layers under a global memory budget. Because it compresses the feature dimension rather than selecting tokens or lowering precision, it can be combined with token eviction and quantization methods. Across LongBench, RULER, and InfiniteBench, AURA-KV preserves long-context performance on Llama-3-8B, Mistral-7B, and LongChat-7B under moderate compression and degrades gracefully under more aggressive budgets. In latency measurements, AURA-KV improves per-layer decoding speed by up to 2.75 \times, and by up to 3.23 \times when combined with token-level cache reduction, with a small one-time prefill overhead.
%U https://aclanthology.org/2026.emnlp-main.2250/
%P 51179-51203
Markdown (Informal)
[AURA-KV: Adaptive Subspace Rank Approximation for Efficient Long-Context LLM Inference](https://aclanthology.org/2026.emnlp-main.2250/) (Xie et al., EMNLP 2026)
ACL