@inproceedings{lee-etal-2026-opera,
title = "{OPERA}-{M}o{E}: Output-aware Routing Perturbation Alignment via Observable Discrepancy for {M}o{E} Quantization",
author = "Lee, Geonho and
Park, Hancheol and
Kim, Tae-Ho and
Choi, Jungwook",
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.2201/",
pages = "47132--47148",
ISBN = "979-8-89176-462-0",
abstract = "Mixture-of-Experts (MoE) architectures have become a dominant paradigm for scaling large language models (LLMs), but their large memory footprint still makes post-training quantization (PTQ) essential for practical deployment. Unlike dense LLMs, quantized MoE models exhibit routing perturbations, where quantization alters expert routing behavior. Existing router alignment methods mitigate this issue by uniformly minimizing routing discrepancies across all tokens. However, we observe that routing perturbations exhibit highly heterogeneous relationships with observable output discrepancy: many tokens preserve nearly identical output distributions despite substantial routing changes. Motivated by this observation, we propose OPERA, an output-aware router alignment strategy for quantized MoE models that selectively prioritizes tokens exhibiting observable output discrepancy after quantization. We further analyze output-preserving routing perturbations and show that they frequently occur on highly predictable tokens, where substantial routing perturbations and hidden-state deviations may not strongly affect projected output distributions. Experiments across multiple MoE LLMs demonstrate that OPERA generally improves over conventional router alignment, with particularly noticeable gains under aggressive low-bit quantization settings."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="lee-etal-2026-opera">
<titleInfo>
<title>OPERA-MoE: Output-aware Routing Perturbation Alignment via Observable Discrepancy for MoE Quantization</title>
</titleInfo>
<name type="personal">
<namePart type="given">Geonho</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hancheol</namePart>
<namePart type="family">Park</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tae-Ho</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jungwook</namePart>
<namePart type="family">Choi</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>Mixture-of-Experts (MoE) architectures have become a dominant paradigm for scaling large language models (LLMs), but their large memory footprint still makes post-training quantization (PTQ) essential for practical deployment. Unlike dense LLMs, quantized MoE models exhibit routing perturbations, where quantization alters expert routing behavior. Existing router alignment methods mitigate this issue by uniformly minimizing routing discrepancies across all tokens. However, we observe that routing perturbations exhibit highly heterogeneous relationships with observable output discrepancy: many tokens preserve nearly identical output distributions despite substantial routing changes. Motivated by this observation, we propose OPERA, an output-aware router alignment strategy for quantized MoE models that selectively prioritizes tokens exhibiting observable output discrepancy after quantization. We further analyze output-preserving routing perturbations and show that they frequently occur on highly predictable tokens, where substantial routing perturbations and hidden-state deviations may not strongly affect projected output distributions. Experiments across multiple MoE LLMs demonstrate that OPERA generally improves over conventional router alignment, with particularly noticeable gains under aggressive low-bit quantization settings.</abstract>
<identifier type="citekey">lee-etal-2026-opera</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2201/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>47132</start>
<end>47148</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T OPERA-MoE: Output-aware Routing Perturbation Alignment via Observable Discrepancy for MoE Quantization
%A Lee, Geonho
%A Park, Hancheol
%A Kim, Tae-Ho
%A Choi, Jungwook
%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 lee-etal-2026-opera
%X Mixture-of-Experts (MoE) architectures have become a dominant paradigm for scaling large language models (LLMs), but their large memory footprint still makes post-training quantization (PTQ) essential for practical deployment. Unlike dense LLMs, quantized MoE models exhibit routing perturbations, where quantization alters expert routing behavior. Existing router alignment methods mitigate this issue by uniformly minimizing routing discrepancies across all tokens. However, we observe that routing perturbations exhibit highly heterogeneous relationships with observable output discrepancy: many tokens preserve nearly identical output distributions despite substantial routing changes. Motivated by this observation, we propose OPERA, an output-aware router alignment strategy for quantized MoE models that selectively prioritizes tokens exhibiting observable output discrepancy after quantization. We further analyze output-preserving routing perturbations and show that they frequently occur on highly predictable tokens, where substantial routing perturbations and hidden-state deviations may not strongly affect projected output distributions. Experiments across multiple MoE LLMs demonstrate that OPERA generally improves over conventional router alignment, with particularly noticeable gains under aggressive low-bit quantization settings.
%U https://aclanthology.org/2026.findings-emnlp.2201/
%P 47132-47148
Markdown (Informal)
[OPERA-MoE: Output-aware Routing Perturbation Alignment via Observable Discrepancy for MoE Quantization](https://aclanthology.org/2026.findings-emnlp.2201/) (Lee et al., Findings 2026)
ACL