@inproceedings{yu-etal-2026-dsl,
title = "{DSL}-{LL}a{DA}: Scaling Continuous Denoising to 8{B} Masked Diffusion {LM}s",
author = "Yu, Longxuan and
Wu, Yunshu and
Fu, Yu and
Xiong, Siheng and
Brekelmans, Rob and
Liu, Hui and
Dong, Yue and
Ver Steeg, Greg",
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.2101/",
pages = "44727--44752",
ISBN = "979-8-89176-462-0",
abstract = "Masked diffusion language models (DLMs) generate text through iterative parallel decoding, but at low step budgets, committing many tokens independently can produce premature termination or repetitive outputs. Continuous embedding-space denoising can mitigate this failure mode by maintaining soft states for all positions, yet training such models from scratch at scale remains challenging. We show that a pretrained masked DLM can instead be efficiently adapted to support continuous denoising. Starting from LLaDA-8B-Instruct, we perform only 1,000 steps of continued pretraining with Discrete Stochastic Localization (DSL), replacing binary masking with continuous per-token Gaussian noise mapped to soft embedding mixtures. The adapted model supports SDE-based inference that evolves all positions jointly and defers hard token commitment to the final step. At $\leq$16 denoising evaluations, DSL-LLaDA substantially mitigates the premature-termination and repetition tradeoff in open-ended generation and achieves the best ROUGE-1 among the evaluated diffusion baselines on all four zero-shot summarization benchmarks. The same adaptation also enables selective correction of corrupted tokens while preserving more than 98{\%} of clean tokens. Same-compute controls indicate that these behaviors arise from continuous-noise exposure rather than additional training alone."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yu-etal-2026-dsl">
<titleInfo>
<title>DSL-LLaDA: Scaling Continuous Denoising to 8B Masked Diffusion LMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Longxuan</namePart>
<namePart type="family">Yu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yunshu</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yu</namePart>
<namePart type="family">Fu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Siheng</namePart>
<namePart type="family">Xiong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rob</namePart>
<namePart type="family">Brekelmans</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hui</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Dong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Greg</namePart>
<namePart type="family">Ver Steeg</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>Masked diffusion language models (DLMs) generate text through iterative parallel decoding, but at low step budgets, committing many tokens independently can produce premature termination or repetitive outputs. Continuous embedding-space denoising can mitigate this failure mode by maintaining soft states for all positions, yet training such models from scratch at scale remains challenging. We show that a pretrained masked DLM can instead be efficiently adapted to support continuous denoising. Starting from LLaDA-8B-Instruct, we perform only 1,000 steps of continued pretraining with Discrete Stochastic Localization (DSL), replacing binary masking with continuous per-token Gaussian noise mapped to soft embedding mixtures. The adapted model supports SDE-based inference that evolves all positions jointly and defers hard token commitment to the final step. At łeq16 denoising evaluations, DSL-LLaDA substantially mitigates the premature-termination and repetition tradeoff in open-ended generation and achieves the best ROUGE-1 among the evaluated diffusion baselines on all four zero-shot summarization benchmarks. The same adaptation also enables selective correction of corrupted tokens while preserving more than 98% of clean tokens. Same-compute controls indicate that these behaviors arise from continuous-noise exposure rather than additional training alone.</abstract>
<identifier type="citekey">yu-etal-2026-dsl</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2101/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>44727</start>
<end>44752</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T DSL-LLaDA: Scaling Continuous Denoising to 8B Masked Diffusion LMs
%A Yu, Longxuan
%A Wu, Yunshu
%A Fu, Yu
%A Xiong, Siheng
%A Brekelmans, Rob
%A Liu, Hui
%A Dong, Yue
%A Ver Steeg, Greg
%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 yu-etal-2026-dsl
%X Masked diffusion language models (DLMs) generate text through iterative parallel decoding, but at low step budgets, committing many tokens independently can produce premature termination or repetitive outputs. Continuous embedding-space denoising can mitigate this failure mode by maintaining soft states for all positions, yet training such models from scratch at scale remains challenging. We show that a pretrained masked DLM can instead be efficiently adapted to support continuous denoising. Starting from LLaDA-8B-Instruct, we perform only 1,000 steps of continued pretraining with Discrete Stochastic Localization (DSL), replacing binary masking with continuous per-token Gaussian noise mapped to soft embedding mixtures. The adapted model supports SDE-based inference that evolves all positions jointly and defers hard token commitment to the final step. At łeq16 denoising evaluations, DSL-LLaDA substantially mitigates the premature-termination and repetition tradeoff in open-ended generation and achieves the best ROUGE-1 among the evaluated diffusion baselines on all four zero-shot summarization benchmarks. The same adaptation also enables selective correction of corrupted tokens while preserving more than 98% of clean tokens. Same-compute controls indicate that these behaviors arise from continuous-noise exposure rather than additional training alone.
%U https://aclanthology.org/2026.findings-emnlp.2101/
%P 44727-44752
Markdown (Informal)
[DSL-LLaDA: Scaling Continuous Denoising to 8B Masked Diffusion LMs](https://aclanthology.org/2026.findings-emnlp.2101/) (Yu et al., Findings 2026)
ACL
- Longxuan Yu, Yunshu Wu, Yu Fu, Siheng Xiong, Rob Brekelmans, Hui Liu, Yue Dong, and Greg Ver Steeg. 2026. DSL-LLaDA: Scaling Continuous Denoising to 8B Masked Diffusion LMs. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 44727–44752, Budapest, Hungary. Association for Computational Linguistics.