@inproceedings{xu-etal-2026-transition,
title = "Transition Overlap: Construction-Time Conditioning for Position-Independent {KV} Cache Reuse",
author = "Xu, Liang and
Wang, Zhichun and
Tian, Xuetao",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1520/",
doi = "10.18653/2026.findings-emnlp.1520",
pages = "32366--32386",
ISBN = "979-8-89176-462-0",
abstract = "In long context applications like Retrieval Augmented Generation (RAG), Large Language Models (LLMs) often reuse independently precomputed Key Value (KV) caches to reduce latency. However, directly assembling these cached segments introduces errors that degrade generation quality. We study this problem in the setting of position-independent caching and decompose the reuse error into two components: a geometric error ($G$) caused by position mismatch, which can be largely removed by position alignment, and a dynamic error ($D$) caused by context-dependent representation changes. To characterize $D$, we introduce the smoothness gap, a local diagnostic that measures the discrepancy between local KV-state transitions in isolated caches and those in the full-context baseline. This gap serves as a necessary local consistency diagnostic for reconstruction errors across adjacent positions and reveals that recomputing non-contiguous selected tokens can introduce additional transition discontinuities, thereby amplifying the dynamic error. To overcome this limitation, we propose \textbf{Transition Overlap}, a training-free cache-construction protocol that prepends a short temporary overlap to approximate full-context boundary transitions and stores only the KV states of the original segment tokens. We instantiate it under two deployment regimes: Self-Segment Overlap (SSO) for query-independent offline caches and Cross-Segment Overlap (CSO) for cases in which the predecessor segment is known during construction. In both evaluated deployment settings, Transition Overlap improves the measured serving latency{--}quality trade-off over training-free PIC baselines[Code is available at {\ensuremath{<}}https://github.com/boyoi/Transition-Overlap{\ensuremath{>}}]."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xu-etal-2026-transition">
<titleInfo>
<title>Transition Overlap: Construction-Time Conditioning for Position-Independent KV Cache Reuse</title>
</titleInfo>
<name type="personal">
<namePart type="given">Liang</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhichun</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuetao</namePart>
<namePart type="family">Tian</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>In long context applications like Retrieval Augmented Generation (RAG), Large Language Models (LLMs) often reuse independently precomputed Key Value (KV) caches to reduce latency. However, directly assembling these cached segments introduces errors that degrade generation quality. We study this problem in the setting of position-independent caching and decompose the reuse error into two components: a geometric error (G) caused by position mismatch, which can be largely removed by position alignment, and a dynamic error (D) caused by context-dependent representation changes. To characterize D, we introduce the smoothness gap, a local diagnostic that measures the discrepancy between local KV-state transitions in isolated caches and those in the full-context baseline. This gap serves as a necessary local consistency diagnostic for reconstruction errors across adjacent positions and reveals that recomputing non-contiguous selected tokens can introduce additional transition discontinuities, thereby amplifying the dynamic error. To overcome this limitation, we propose Transition Overlap, a training-free cache-construction protocol that prepends a short temporary overlap to approximate full-context boundary transitions and stores only the KV states of the original segment tokens. We instantiate it under two deployment regimes: Self-Segment Overlap (SSO) for query-independent offline caches and Cross-Segment Overlap (CSO) for cases in which the predecessor segment is known during construction. In both evaluated deployment settings, Transition Overlap improves the measured serving latency–quality trade-off over training-free PIC baselines[Code is available at \ensuremath<https://github.com/boyoi/Transition-Overlap\ensuremath>].</abstract>
<identifier type="citekey">xu-etal-2026-transition</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1520</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1520/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>32366</start>
<end>32386</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Transition Overlap: Construction-Time Conditioning for Position-Independent KV Cache Reuse
%A Xu, Liang
%A Wang, Zhichun
%A Tian, Xuetao
%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 xu-etal-2026-transition
%X In long context applications like Retrieval Augmented Generation (RAG), Large Language Models (LLMs) often reuse independently precomputed Key Value (KV) caches to reduce latency. However, directly assembling these cached segments introduces errors that degrade generation quality. We study this problem in the setting of position-independent caching and decompose the reuse error into two components: a geometric error (G) caused by position mismatch, which can be largely removed by position alignment, and a dynamic error (D) caused by context-dependent representation changes. To characterize D, we introduce the smoothness gap, a local diagnostic that measures the discrepancy between local KV-state transitions in isolated caches and those in the full-context baseline. This gap serves as a necessary local consistency diagnostic for reconstruction errors across adjacent positions and reveals that recomputing non-contiguous selected tokens can introduce additional transition discontinuities, thereby amplifying the dynamic error. To overcome this limitation, we propose Transition Overlap, a training-free cache-construction protocol that prepends a short temporary overlap to approximate full-context boundary transitions and stores only the KV states of the original segment tokens. We instantiate it under two deployment regimes: Self-Segment Overlap (SSO) for query-independent offline caches and Cross-Segment Overlap (CSO) for cases in which the predecessor segment is known during construction. In both evaluated deployment settings, Transition Overlap improves the measured serving latency–quality trade-off over training-free PIC baselines[Code is available at \ensuremath<https://github.com/boyoi/Transition-Overlap\ensuremath>].
%R 10.18653/2026.findings-emnlp.1520
%U https://aclanthology.org/2026.findings-emnlp.1520/
%U https://doi.org/10.18653/2026.findings-emnlp.1520
%P 32366-32386
Markdown (Informal)
[Transition Overlap: Construction-Time Conditioning for Position-Independent KV Cache Reuse](https://aclanthology.org/2026.findings-emnlp.1520/) (Xu et al., Findings 2026)
ACL