@inproceedings{yao-etal-2026-wnw,
title = "{W}n{W}: Waxing-and-Waning {KV} Cache for Long-Form Speech {LLM}s",
author = "Yao, Yiming and
Lyu, Chenyang and
Ni, Xuanfan and
Wang, Longyue and
Luo, Weihua and
Yang, Yazheng and
Su, Jinsong",
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.1076/",
pages = "24735--24748",
ISBN = "979-8-89176-459-0",
abstract = "Long-form audio makes the Key-Value (KV) cache the dominant bottleneck in speech LLMs, yet existing KV compression methods commit to an irreversible retention set at prefill. However, we reveal a severe attention mismatch in this static paradigm particularly for long-form audio: prefill attention concentrates at the audio{'}s start, whereas decode-time attention distributes broadly. This weak overlap between their top-attended positions leads to performance degradation, with generation collapsing into repetition at extreme compression rates. To bridge this gap, we propose WnW (Waxing-and-Waning KV cache), which makes eviction reversible by partitioning KV-heads into three roles via offline calibration. Anchor heads keep all audio KV on GPU and provide a per-step signal of which audio region the current token needs; tidal heads, guided by that signal, recall evicted KV from CPU on demand; fixed heads discard theirs permanently. Evaluations on LibriSpeech-Long Benchmark across two 3B backbones (Voxtral-Mini and Qwen2.5-Omni) demonstrate that WnW preserves near-full-cache accuracy while maintaining only 20{\%} of audio tokens on GPU{---}a regime where irreversible methods fail to terminate. WnW exhibits strong generalization across languages, tasks, and domains, with CPU-GPU chunk swapping introducing negligible decoding overhead. More broadly, WnW transforms KV compression into a reversible memory hierarchy in which a small set of calibrated observer heads governs the placement of the remaining cache."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yao-etal-2026-wnw">
<titleInfo>
<title>WnW: Waxing-and-Waning KV Cache for Long-Form Speech LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yiming</namePart>
<namePart type="family">Yao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chenyang</namePart>
<namePart type="family">Lyu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuanfan</namePart>
<namePart type="family">Ni</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Longyue</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Weihua</namePart>
<namePart type="family">Luo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yazheng</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jinsong</namePart>
<namePart type="family">Su</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-form audio makes the Key-Value (KV) cache the dominant bottleneck in speech LLMs, yet existing KV compression methods commit to an irreversible retention set at prefill. However, we reveal a severe attention mismatch in this static paradigm particularly for long-form audio: prefill attention concentrates at the audio’s start, whereas decode-time attention distributes broadly. This weak overlap between their top-attended positions leads to performance degradation, with generation collapsing into repetition at extreme compression rates. To bridge this gap, we propose WnW (Waxing-and-Waning KV cache), which makes eviction reversible by partitioning KV-heads into three roles via offline calibration. Anchor heads keep all audio KV on GPU and provide a per-step signal of which audio region the current token needs; tidal heads, guided by that signal, recall evicted KV from CPU on demand; fixed heads discard theirs permanently. Evaluations on LibriSpeech-Long Benchmark across two 3B backbones (Voxtral-Mini and Qwen2.5-Omni) demonstrate that WnW preserves near-full-cache accuracy while maintaining only 20% of audio tokens on GPU—a regime where irreversible methods fail to terminate. WnW exhibits strong generalization across languages, tasks, and domains, with CPU-GPU chunk swapping introducing negligible decoding overhead. More broadly, WnW transforms KV compression into a reversible memory hierarchy in which a small set of calibrated observer heads governs the placement of the remaining cache.</abstract>
<identifier type="citekey">yao-etal-2026-wnw</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1076/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>24735</start>
<end>24748</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T WnW: Waxing-and-Waning KV Cache for Long-Form Speech LLMs
%A Yao, Yiming
%A Lyu, Chenyang
%A Ni, Xuanfan
%A Wang, Longyue
%A Luo, Weihua
%A Yang, Yazheng
%A Su, Jinsong
%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 yao-etal-2026-wnw
%X Long-form audio makes the Key-Value (KV) cache the dominant bottleneck in speech LLMs, yet existing KV compression methods commit to an irreversible retention set at prefill. However, we reveal a severe attention mismatch in this static paradigm particularly for long-form audio: prefill attention concentrates at the audio’s start, whereas decode-time attention distributes broadly. This weak overlap between their top-attended positions leads to performance degradation, with generation collapsing into repetition at extreme compression rates. To bridge this gap, we propose WnW (Waxing-and-Waning KV cache), which makes eviction reversible by partitioning KV-heads into three roles via offline calibration. Anchor heads keep all audio KV on GPU and provide a per-step signal of which audio region the current token needs; tidal heads, guided by that signal, recall evicted KV from CPU on demand; fixed heads discard theirs permanently. Evaluations on LibriSpeech-Long Benchmark across two 3B backbones (Voxtral-Mini and Qwen2.5-Omni) demonstrate that WnW preserves near-full-cache accuracy while maintaining only 20% of audio tokens on GPU—a regime where irreversible methods fail to terminate. WnW exhibits strong generalization across languages, tasks, and domains, with CPU-GPU chunk swapping introducing negligible decoding overhead. More broadly, WnW transforms KV compression into a reversible memory hierarchy in which a small set of calibrated observer heads governs the placement of the remaining cache.
%U https://aclanthology.org/2026.emnlp-main.1076/
%P 24735-24748
Markdown (Informal)
[WnW: Waxing-and-Waning KV Cache for Long-Form Speech LLMs](https://aclanthology.org/2026.emnlp-main.1076/) (Yao et al., EMNLP 2026)
ACL
- Yiming Yao, Chenyang Lyu, Xuanfan Ni, Longyue Wang, Weihua Luo, Yazheng Yang, and Jinsong Su. 2026. WnW: Waxing-and-Waning KV Cache for Long-Form Speech LLMs. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 24735–24748, Budapest, Hungary. Association for Computational Linguistics.