@inproceedings{liang-etal-2026-cd4lm,
title = "{CD}$^4${LM}: Consistency Distillation and a{D}aptive Decoding for Diffusion Language Models",
author = "Liang, Yihao and
Wang, Ze and
Sun, Ximeng and
Liu, Jiang and
Wu, Jialian and
Chen, Hao and
Yu, Xiaodong and
Barsoum, Emad and
Liu, Zicheng and
Jha, Niraj",
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.496/",
pages = "10378--10410",
ISBN = "979-8-89176-462-0",
abstract = "Diffusion language models (DLMs) accelerate generation by decoding multiple tokens in parallel, but pushing them to fast decoding often sacrifices accuracy. We attribute this to a static-to-dynamic mismatch: training optimizes local denoising under fixed schedules, while fast inference skips steps and encounters irregular mask patterns. Existing accelerators largely address only one side of this problem. Training-free inference methods speed up decoding without changing the model, leaving the mismatch unresolved; training-side distillation methods adapt the model, but often tie the student to a specific decoding trajectory, limiting composability and still suffering accuracy loss. We propose CD$^4$LM, a composable framework for accurate DLM acceleration. Its core training component, Discrete-Space Consistency Distillation (DSCD) trains a trajectory-invariant student on diverse masked states, reducing the mismatch without fixing a decoding policy. We pair DSCD with a lightweight Confidence-Adaptive Decoder, while keeping the trained student compatible with external inference-side accelerators. On GSM8K, CD$^4$LM preserves accuracy while achieving a 5.18$\times$ speedup; stacking CD$^4$LM with per-step cache accelerators further pushes speedup to 9.73$\times$ at teacher-level accuracy, suggesting a scalable path for combining training-side robustness with future decoding advances."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liang-etal-2026-cd4lm">
<titleInfo>
<title>CD⁴LM: Consistency Distillation and aDaptive Decoding for Diffusion Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yihao</namePart>
<namePart type="family">Liang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ze</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ximeng</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiang</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jialian</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hao</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaodong</namePart>
<namePart type="family">Yu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Emad</namePart>
<namePart type="family">Barsoum</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zicheng</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Niraj</namePart>
<namePart type="family">Jha</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>Diffusion language models (DLMs) accelerate generation by decoding multiple tokens in parallel, but pushing them to fast decoding often sacrifices accuracy. We attribute this to a static-to-dynamic mismatch: training optimizes local denoising under fixed schedules, while fast inference skips steps and encounters irregular mask patterns. Existing accelerators largely address only one side of this problem. Training-free inference methods speed up decoding without changing the model, leaving the mismatch unresolved; training-side distillation methods adapt the model, but often tie the student to a specific decoding trajectory, limiting composability and still suffering accuracy loss. We propose CD⁴LM, a composable framework for accurate DLM acceleration. Its core training component, Discrete-Space Consistency Distillation (DSCD) trains a trajectory-invariant student on diverse masked states, reducing the mismatch without fixing a decoding policy. We pair DSCD with a lightweight Confidence-Adaptive Decoder, while keeping the trained student compatible with external inference-side accelerators. On GSM8K, CD⁴LM preserves accuracy while achieving a 5.18\times speedup; stacking CD⁴LM with per-step cache accelerators further pushes speedup to 9.73\times at teacher-level accuracy, suggesting a scalable path for combining training-side robustness with future decoding advances.</abstract>
<identifier type="citekey">liang-etal-2026-cd4lm</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.496/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>10378</start>
<end>10410</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T CD⁴LM: Consistency Distillation and aDaptive Decoding for Diffusion Language Models
%A Liang, Yihao
%A Wang, Ze
%A Sun, Ximeng
%A Liu, Jiang
%A Wu, Jialian
%A Chen, Hao
%A Yu, Xiaodong
%A Barsoum, Emad
%A Liu, Zicheng
%A Jha, Niraj
%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 liang-etal-2026-cd4lm
%X Diffusion language models (DLMs) accelerate generation by decoding multiple tokens in parallel, but pushing them to fast decoding often sacrifices accuracy. We attribute this to a static-to-dynamic mismatch: training optimizes local denoising under fixed schedules, while fast inference skips steps and encounters irregular mask patterns. Existing accelerators largely address only one side of this problem. Training-free inference methods speed up decoding without changing the model, leaving the mismatch unresolved; training-side distillation methods adapt the model, but often tie the student to a specific decoding trajectory, limiting composability and still suffering accuracy loss. We propose CD⁴LM, a composable framework for accurate DLM acceleration. Its core training component, Discrete-Space Consistency Distillation (DSCD) trains a trajectory-invariant student on diverse masked states, reducing the mismatch without fixing a decoding policy. We pair DSCD with a lightweight Confidence-Adaptive Decoder, while keeping the trained student compatible with external inference-side accelerators. On GSM8K, CD⁴LM preserves accuracy while achieving a 5.18\times speedup; stacking CD⁴LM with per-step cache accelerators further pushes speedup to 9.73\times at teacher-level accuracy, suggesting a scalable path for combining training-side robustness with future decoding advances.
%U https://aclanthology.org/2026.findings-emnlp.496/
%P 10378-10410
Markdown (Informal)
[CD4LM: Consistency Distillation and aDaptive Decoding for Diffusion Language Models](https://aclanthology.org/2026.findings-emnlp.496/) (Liang et al., Findings 2026)
ACL
- Yihao Liang, Ze Wang, Ximeng Sun, Jiang Liu, Jialian Wu, Hao Chen, Xiaodong Yu, Emad Barsoum, Zicheng Liu, and Niraj Jha. 2026. CD4LM: Consistency Distillation and aDaptive Decoding for Diffusion Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 10378–10410, Budapest, Hungary. Association for Computational Linguistics.