@inproceedings{ye-etal-2026-less,
title = "Less Random, More Incoherent: Greedy {H}adamard Alignment for {LLM} Quantization",
author = "Ye, Yun and
Wu, Zhaoqi and
Wang, Xueshu and
Zhang, Cong and
Yongbo, Zhu and
Sun, Zhonghua and
XU, Beryl",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.70/",
pages = "1526--1540",
ISBN = "979-8-89176-439-2",
abstract = "Hadamard preprocessing is widely used in large language model (LLM) quantization to reduce activation coherence and suppress peaks. Existing methods typically rely on random sign flips before the Hadamard transform, which do not use information from the target data distribution. We propose Greedy Hadamard Alignment (GHA), a simple training-free method to build a signed-permutation preconditioner from calibration data. GHA uses a pooled representative activation, a greedy search over flat targets, and an efficient alignment step that is analytically optimal once the target is fixed. We also study when one shared GHA preconditioner can transfer across LLM activations. GHA works best when activations are directionally aligned and stable across calibration sources. Overall, GHA improves post-quantization model quality over randomized Hadamard baselines, with the strongest gains in KV-cache quantization, more moderate gains for hidden states, and little compute overhead during inference."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ye-etal-2026-less">
<titleInfo>
<title>Less Random, More Incoherent: Greedy Hadamard Alignment for LLM Quantization</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yun</namePart>
<namePart type="family">Ye</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhaoqi</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xueshu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cong</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhu</namePart>
<namePart type="family">Yongbo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhonghua</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Beryl</namePart>
<namePart type="family">XU</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-439-2</identifier>
</relatedItem>
<abstract>Hadamard preprocessing is widely used in large language model (LLM) quantization to reduce activation coherence and suppress peaks. Existing methods typically rely on random sign flips before the Hadamard transform, which do not use information from the target data distribution. We propose Greedy Hadamard Alignment (GHA), a simple training-free method to build a signed-permutation preconditioner from calibration data. GHA uses a pooled representative activation, a greedy search over flat targets, and an efficient alignment step that is analytically optimal once the target is fixed. We also study when one shared GHA preconditioner can transfer across LLM activations. GHA works best when activations are directionally aligned and stable across calibration sources. Overall, GHA improves post-quantization model quality over randomized Hadamard baselines, with the strongest gains in KV-cache quantization, more moderate gains for hidden states, and little compute overhead during inference.</abstract>
<identifier type="citekey">ye-etal-2026-less</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.70/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>1526</start>
<end>1540</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Less Random, More Incoherent: Greedy Hadamard Alignment for LLM Quantization
%A Ye, Yun
%A Wu, Zhaoqi
%A Wang, Xueshu
%A Zhang, Cong
%A Yongbo, Zhu
%A Sun, Zhonghua
%A XU, Beryl
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F ye-etal-2026-less
%X Hadamard preprocessing is widely used in large language model (LLM) quantization to reduce activation coherence and suppress peaks. Existing methods typically rely on random sign flips before the Hadamard transform, which do not use information from the target data distribution. We propose Greedy Hadamard Alignment (GHA), a simple training-free method to build a signed-permutation preconditioner from calibration data. GHA uses a pooled representative activation, a greedy search over flat targets, and an efficient alignment step that is analytically optimal once the target is fixed. We also study when one shared GHA preconditioner can transfer across LLM activations. GHA works best when activations are directionally aligned and stable across calibration sources. Overall, GHA improves post-quantization model quality over randomized Hadamard baselines, with the strongest gains in KV-cache quantization, more moderate gains for hidden states, and little compute overhead during inference.
%U https://aclanthology.org/2026.aacl-long.70/
%P 1526-1540
Markdown (Informal)
[Less Random, More Incoherent: Greedy Hadamard Alignment for LLM Quantization](https://aclanthology.org/2026.aacl-long.70/) (Ye et al., AACL-IJCNLP 2026)
ACL
- Yun Ye, Zhaoqi Wu, Xueshu Wang, Cong Zhang, Zhu Yongbo, Zhonghua Sun, and Beryl XU. 2026. Less Random, More Incoherent: Greedy Hadamard Alignment for LLM Quantization. In Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1526–1540, Hengqin, China. Association for Computational Linguistics.