@inproceedings{yang-etal-2026-heal,
title = "{HEAL}: Latent Head Alignment for Post-Training Quantization of Large Language Models",
author = "Yang, Mingzhe and
Lin, Sihao and
Zhang, Ruibo and
Zhu, Xinyuan and
Chen, Lei and
Chen, Weidong and
Chang, Xiaojun",
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.326/",
pages = "7526--7540",
ISBN = "979-8-89176-459-0",
abstract = "Post-training quantization (PTQ) has become a key approach for large language model (LLM) compression. Existing methods commonly adopt block-level reconstruction as a surrogate to calibrate each quantized Transformer block. However, our analysis shows that this loose objective can admit $\textit{cancellation patterns}$, where the aggregated block-level reconstruction error is small even though the per-head errors remain large. We therefore revisit the PTQ formulation from a variational perspective. By treating per-head representations as latent variables and the block output as the observed variable, we find that conventional block-level reconstruction corresponds to optimizing only the observation-level term, while omitting a latent alignment term that captures head-level quantization error. Motivated by this finding, we propose $\textbf{HEAL}$, a Latent ${\textbf {HE}}$ad ${\textbf {AL}}$ignment framework that turns block-level reconstruction calibration into joint observation-latent calibration with stop-gradient head alignment. Experiments on LLaMA-2/3 and Qwen2.5 families under low-bit weight-activation quantization show that HEAL generally improves perplexity and zero-shot accuracy over block-level reconstruction baselines."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yang-etal-2026-heal">
<titleInfo>
<title>HEAL: Latent Head Alignment for Post-Training Quantization of Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Mingzhe</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sihao</namePart>
<namePart type="family">Lin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruibo</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinyuan</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lei</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Weidong</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaojun</namePart>
<namePart type="family">Chang</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>Post-training quantization (PTQ) has become a key approach for large language model (LLM) compression. Existing methods commonly adopt block-level reconstruction as a surrogate to calibrate each quantized Transformer block. However, our analysis shows that this loose objective can admit cancellation patterns, where the aggregated block-level reconstruction error is small even though the per-head errors remain large. We therefore revisit the PTQ formulation from a variational perspective. By treating per-head representations as latent variables and the block output as the observed variable, we find that conventional block-level reconstruction corresponds to optimizing only the observation-level term, while omitting a latent alignment term that captures head-level quantization error. Motivated by this finding, we propose HEAL, a Latent HEad ALignment framework that turns block-level reconstruction calibration into joint observation-latent calibration with stop-gradient head alignment. Experiments on LLaMA-2/3 and Qwen2.5 families under low-bit weight-activation quantization show that HEAL generally improves perplexity and zero-shot accuracy over block-level reconstruction baselines.</abstract>
<identifier type="citekey">yang-etal-2026-heal</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.326/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>7526</start>
<end>7540</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T HEAL: Latent Head Alignment for Post-Training Quantization of Large Language Models
%A Yang, Mingzhe
%A Lin, Sihao
%A Zhang, Ruibo
%A Zhu, Xinyuan
%A Chen, Lei
%A Chen, Weidong
%A Chang, Xiaojun
%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 yang-etal-2026-heal
%X Post-training quantization (PTQ) has become a key approach for large language model (LLM) compression. Existing methods commonly adopt block-level reconstruction as a surrogate to calibrate each quantized Transformer block. However, our analysis shows that this loose objective can admit cancellation patterns, where the aggregated block-level reconstruction error is small even though the per-head errors remain large. We therefore revisit the PTQ formulation from a variational perspective. By treating per-head representations as latent variables and the block output as the observed variable, we find that conventional block-level reconstruction corresponds to optimizing only the observation-level term, while omitting a latent alignment term that captures head-level quantization error. Motivated by this finding, we propose HEAL, a Latent HEad ALignment framework that turns block-level reconstruction calibration into joint observation-latent calibration with stop-gradient head alignment. Experiments on LLaMA-2/3 and Qwen2.5 families under low-bit weight-activation quantization show that HEAL generally improves perplexity and zero-shot accuracy over block-level reconstruction baselines.
%U https://aclanthology.org/2026.emnlp-main.326/
%P 7526-7540
Markdown (Informal)
[HEAL: Latent Head Alignment for Post-Training Quantization of Large Language Models](https://aclanthology.org/2026.emnlp-main.326/) (Yang et al., EMNLP 2026)
ACL