@inproceedings{kumar-mishra-2026-quantizing,
title = "Quantizing Masked Diffusion Language Models: Storage-Quality Trade-offs and Runtime Diagnostics",
author = "Kumar, Pawan and
Mishra, Neel",
editor = "Atanasova, Pepa and
Haffari, Gholamreza and
Wang, Minghan",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-industry.241/",
pages = "4067--4089",
ISBN = "979-8-89176-460-6",
abstract = "Masked diffusion language models reuse one denoiser across many steps, so post-training quantization (PTQ) must balance predictive quality, serialized storage, and runtime memory. We evaluate LLaDA-8B-Instruct and run transfer checks on Dream-7B, MDLM-OWT, encoder-only masked LMs, LLaDA-MoE, and LLaDA2.0-mini. The 24-text calibration set is distinct from the main $n{=}500$-example-per-task evaluation. Three findings emerge. First, AWQ-4 is the closest tested 4-bit quality point at the evaluated 1.1{--}2.3 percentage-point resolution, although ARC-Challenge shows a detectable loss. TurboLDM-$(3{+}1)$ instead provides the smallest serialized checkpoint (4.82 GiB, 3.55 effective bits) while losing quality, making it a storage-compression and diagnostic result, not a live-memory result. NF4 alone reduces peak VRAM from roughly 15 GiB to 6 GiB in our tested PyTorch/bitsandbytes runtime; this is a runtime-specific outcome, not an inherent format advantage. Second, individual attention projections are inexpensive to quantize, whereas jointly quantizing the attention block is damaging. Third, more denoising steps partly recover GSM8K loss but not multiple-choice accuracy."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kumar-mishra-2026-quantizing">
<titleInfo>
<title>Quantizing Masked Diffusion Language Models: Storage-Quality Trade-offs and Runtime Diagnostics</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pawan</namePart>
<namePart type="family">Kumar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Neel</namePart>
<namePart type="family">Mishra</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: Industry Track</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pepa</namePart>
<namePart type="family">Atanasova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gholamreza</namePart>
<namePart type="family">Haffari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Minghan</namePart>
<namePart type="family">Wang</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">979-8-89176-460-6</identifier>
</relatedItem>
<abstract>Masked diffusion language models reuse one denoiser across many steps, so post-training quantization (PTQ) must balance predictive quality, serialized storage, and runtime memory. We evaluate LLaDA-8B-Instruct and run transfer checks on Dream-7B, MDLM-OWT, encoder-only masked LMs, LLaDA-MoE, and LLaDA2.0-mini. The 24-text calibration set is distinct from the main n=500-example-per-task evaluation. Three findings emerge. First, AWQ-4 is the closest tested 4-bit quality point at the evaluated 1.1–2.3 percentage-point resolution, although ARC-Challenge shows a detectable loss. TurboLDM-(3+1) instead provides the smallest serialized checkpoint (4.82 GiB, 3.55 effective bits) while losing quality, making it a storage-compression and diagnostic result, not a live-memory result. NF4 alone reduces peak VRAM from roughly 15 GiB to 6 GiB in our tested PyTorch/bitsandbytes runtime; this is a runtime-specific outcome, not an inherent format advantage. Second, individual attention projections are inexpensive to quantize, whereas jointly quantizing the attention block is damaging. Third, more denoising steps partly recover GSM8K loss but not multiple-choice accuracy.</abstract>
<identifier type="citekey">kumar-mishra-2026-quantizing</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-industry.241/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>4067</start>
<end>4089</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Quantizing Masked Diffusion Language Models: Storage-Quality Trade-offs and Runtime Diagnostics
%A Kumar, Pawan
%A Mishra, Neel
%Y Atanasova, Pepa
%Y Haffari, Gholamreza
%Y Wang, Minghan
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-460-6
%F kumar-mishra-2026-quantizing
%X Masked diffusion language models reuse one denoiser across many steps, so post-training quantization (PTQ) must balance predictive quality, serialized storage, and runtime memory. We evaluate LLaDA-8B-Instruct and run transfer checks on Dream-7B, MDLM-OWT, encoder-only masked LMs, LLaDA-MoE, and LLaDA2.0-mini. The 24-text calibration set is distinct from the main n=500-example-per-task evaluation. Three findings emerge. First, AWQ-4 is the closest tested 4-bit quality point at the evaluated 1.1–2.3 percentage-point resolution, although ARC-Challenge shows a detectable loss. TurboLDM-(3+1) instead provides the smallest serialized checkpoint (4.82 GiB, 3.55 effective bits) while losing quality, making it a storage-compression and diagnostic result, not a live-memory result. NF4 alone reduces peak VRAM from roughly 15 GiB to 6 GiB in our tested PyTorch/bitsandbytes runtime; this is a runtime-specific outcome, not an inherent format advantage. Second, individual attention projections are inexpensive to quantize, whereas jointly quantizing the attention block is damaging. Third, more denoising steps partly recover GSM8K loss but not multiple-choice accuracy.
%U https://aclanthology.org/2026.emnlp-industry.241/
%P 4067-4089
Markdown (Informal)
[Quantizing Masked Diffusion Language Models: Storage-Quality Trade-offs and Runtime Diagnostics](https://aclanthology.org/2026.emnlp-industry.241/) (Kumar & Mishra, EMNLP 2026)
ACL