@inproceedings{heo-etal-2026-l,
title = "$\lambda$-Diffusion Language Models: Controllable Generalization of Continuous and Discrete Diffusion Dynamics",
author = "Heo, DongNyeong and
Song, Hyunmin and
Kim, Taehwan",
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.1117/",
pages = "23666--23683",
ISBN = "979-8-89176-462-0",
abstract = "Diffusion language models (DLMs) have emerged as a promising alternative to autoregressive language models. While discrete DLMs achieve strong performance across language generation tasks, continuous DLMs remain relatively underexplored despite their smooth latent-space dynamics. In this work, we present a unified embedding-space interpretation of continuous and discrete DLMs. Within this framework, we analyze continuous diffusion dynamics as a retention channel that preserves intermediate denoising information during sampling. Motivated by this perspective, we propose $\lambda$-DLM, which smoothly interpolates between continuous and discrete diffusion dynamics through a controllable parameter $\lambda$. By training with randomly sampled $\lambda$, a single model generalizes across continuous, discrete, and intermediate diffusion dynamics, unifying diverse retention behaviors within one framework. Unlike prior methods that rely on fixed or self-confidence-guided diffusion dynamics, $\lambda$-DLM enables controllable exploitation{--}exploration trade-offs without retraining. Experiments on LM1B and OWT unconditional generation benchmarks show that $\lambda$-DLM consistently outperforms strong baselines. Moreover, varying $\lambda$ enables semantically diverse generation without additional training. Compared with temperature sweeps, varying $\lambda$ shows a more structured relationship between lexical diversity and sentence-embedding-based mode coverage."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="heo-etal-2026-l">
<titleInfo>
<title>Ĺ‚ambda-Diffusion Language Models: Controllable Generalization of Continuous and Discrete Diffusion Dynamics</title>
</titleInfo>
<name type="personal">
<namePart type="given">DongNyeong</namePart>
<namePart type="family">Heo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hyunmin</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Taehwan</namePart>
<namePart type="family">Kim</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) have emerged as a promising alternative to autoregressive language models. While discrete DLMs achieve strong performance across language generation tasks, continuous DLMs remain relatively underexplored despite their smooth latent-space dynamics. In this work, we present a unified embedding-space interpretation of continuous and discrete DLMs. Within this framework, we analyze continuous diffusion dynamics as a retention channel that preserves intermediate denoising information during sampling. Motivated by this perspective, we propose łambda-DLM, which smoothly interpolates between continuous and discrete diffusion dynamics through a controllable parameter łambda. By training with randomly sampled łambda, a single model generalizes across continuous, discrete, and intermediate diffusion dynamics, unifying diverse retention behaviors within one framework. Unlike prior methods that rely on fixed or self-confidence-guided diffusion dynamics, łambda-DLM enables controllable exploitation–exploration trade-offs without retraining. Experiments on LM1B and OWT unconditional generation benchmarks show that łambda-DLM consistently outperforms strong baselines. Moreover, varying łambda enables semantically diverse generation without additional training. Compared with temperature sweeps, varying łambda shows a more structured relationship between lexical diversity and sentence-embedding-based mode coverage.</abstract>
<identifier type="citekey">heo-etal-2026-l</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1117/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>23666</start>
<end>23683</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Ĺ‚ambda-Diffusion Language Models: Controllable Generalization of Continuous and Discrete Diffusion Dynamics
%A Heo, DongNyeong
%A Song, Hyunmin
%A Kim, Taehwan
%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 heo-etal-2026-l
%X Diffusion language models (DLMs) have emerged as a promising alternative to autoregressive language models. While discrete DLMs achieve strong performance across language generation tasks, continuous DLMs remain relatively underexplored despite their smooth latent-space dynamics. In this work, we present a unified embedding-space interpretation of continuous and discrete DLMs. Within this framework, we analyze continuous diffusion dynamics as a retention channel that preserves intermediate denoising information during sampling. Motivated by this perspective, we propose łambda-DLM, which smoothly interpolates between continuous and discrete diffusion dynamics through a controllable parameter łambda. By training with randomly sampled łambda, a single model generalizes across continuous, discrete, and intermediate diffusion dynamics, unifying diverse retention behaviors within one framework. Unlike prior methods that rely on fixed or self-confidence-guided diffusion dynamics, łambda-DLM enables controllable exploitation–exploration trade-offs without retraining. Experiments on LM1B and OWT unconditional generation benchmarks show that łambda-DLM consistently outperforms strong baselines. Moreover, varying łambda enables semantically diverse generation without additional training. Compared with temperature sweeps, varying łambda shows a more structured relationship between lexical diversity and sentence-embedding-based mode coverage.
%U https://aclanthology.org/2026.findings-emnlp.1117/
%P 23666-23683
Markdown (Informal)
[𝜆-Diffusion Language Models: Controllable Generalization of Continuous and Discrete Diffusion Dynamics](https://aclanthology.org/2026.findings-emnlp.1117/) (Heo et al., Findings 2026)
ACL