@inproceedings{xie-lin-2026-beyond,
title = "Beyond: Better-than-Full-Precision {KV} Caches via Learnable Non-Uniform Quantization as Implicit Regularization",
author = "Xie, Yuhao and
Lin, Mingjie",
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.526/",
pages = "12135--12158",
ISBN = "979-8-89176-459-0",
abstract = "Key{--}value (KV) caching accelerates autoregressive decoding but makes long-context serving memory-intensive. Most KV quantizers minimize local reconstruction distortion, although this proxy ignores how perturbations propagate through downstream computation. We introduce Beyond, a loss-aware non-uniform quantizer that learns representation levels and decision thresholds from next-token negative log-likelihood while freezing the backbone. Training and inference use the same hard quantizer, with straight-through and boundary-localized pseudo-gradients updating only its parameters. With 4-bit codes and group size 32, Beyond uses 5.0 effective bits per KV scalar, providing 3.2{\texttimes} compression relative to a 16-bit cache. Across three instruction-tuned LLMs, it lowers held-out Pile perplexity below the full-precision baseline by 2.2{--}7.0{\%} and attains the highest LongBench Avg21 among full-precision Native, KV-FP8, and TurboQuant for each model. On an NVIDIA B300, its fused packed-cache kernel achieves a peak 1.50{\texttimes} speedup over FlashAttention-4, while the resulting vLLM path improves end-to-end throughput by up to 1.17{\texttimes}. Our code is available at https://github.com/wsxhjnb1/Beyond-KV-Cache."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xie-lin-2026-beyond">
<titleInfo>
<title>Beyond: Better-than-Full-Precision KV Caches via Learnable Non-Uniform Quantization as Implicit Regularization</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yuhao</namePart>
<namePart type="family">Xie</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mingjie</namePart>
<namePart type="family">Lin</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>Key–value (KV) caching accelerates autoregressive decoding but makes long-context serving memory-intensive. Most KV quantizers minimize local reconstruction distortion, although this proxy ignores how perturbations propagate through downstream computation. We introduce Beyond, a loss-aware non-uniform quantizer that learns representation levels and decision thresholds from next-token negative log-likelihood while freezing the backbone. Training and inference use the same hard quantizer, with straight-through and boundary-localized pseudo-gradients updating only its parameters. With 4-bit codes and group size 32, Beyond uses 5.0 effective bits per KV scalar, providing 3.2× compression relative to a 16-bit cache. Across three instruction-tuned LLMs, it lowers held-out Pile perplexity below the full-precision baseline by 2.2–7.0% and attains the highest LongBench Avg21 among full-precision Native, KV-FP8, and TurboQuant for each model. On an NVIDIA B300, its fused packed-cache kernel achieves a peak 1.50× speedup over FlashAttention-4, while the resulting vLLM path improves end-to-end throughput by up to 1.17×. Our code is available at https://github.com/wsxhjnb1/Beyond-KV-Cache.</abstract>
<identifier type="citekey">xie-lin-2026-beyond</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.526/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>12135</start>
<end>12158</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Beyond: Better-than-Full-Precision KV Caches via Learnable Non-Uniform Quantization as Implicit Regularization
%A Xie, Yuhao
%A Lin, Mingjie
%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-lin-2026-beyond
%X Key–value (KV) caching accelerates autoregressive decoding but makes long-context serving memory-intensive. Most KV quantizers minimize local reconstruction distortion, although this proxy ignores how perturbations propagate through downstream computation. We introduce Beyond, a loss-aware non-uniform quantizer that learns representation levels and decision thresholds from next-token negative log-likelihood while freezing the backbone. Training and inference use the same hard quantizer, with straight-through and boundary-localized pseudo-gradients updating only its parameters. With 4-bit codes and group size 32, Beyond uses 5.0 effective bits per KV scalar, providing 3.2× compression relative to a 16-bit cache. Across three instruction-tuned LLMs, it lowers held-out Pile perplexity below the full-precision baseline by 2.2–7.0% and attains the highest LongBench Avg21 among full-precision Native, KV-FP8, and TurboQuant for each model. On an NVIDIA B300, its fused packed-cache kernel achieves a peak 1.50× speedup over FlashAttention-4, while the resulting vLLM path improves end-to-end throughput by up to 1.17×. Our code is available at https://github.com/wsxhjnb1/Beyond-KV-Cache.
%U https://aclanthology.org/2026.emnlp-main.526/
%P 12135-12158
Markdown (Informal)
[Beyond: Better-than-Full-Precision KV Caches via Learnable Non-Uniform Quantization as Implicit Regularization](https://aclanthology.org/2026.emnlp-main.526/) (Xie & Lin, EMNLP 2026)
ACL