dLLM: Simple Diffusion Language Modeling

Zhanhui Zhou; Lingjie Chen; Hanghang Tong; Dawn Song

cs.CLFeb 26, 2026

dLLM: Simple Diffusion Language Modeling

Authors: Zhanhui Zhou, Lingjie Chen, Hanghang Tong, Dawn Song

Organizations: UIUC

Abstract

Although diffusion language models (DLMs) are evolving quickly, many recent models converge on a set of shared components. These components, however, are distributed across ad-hoc research codebases or lack transparent implementations, making them difficult to reproduce or extend. As the field accelerates, there is a clear need for a unified framework that standardizes these common components while remaining flexible enough to support new methods and architectures. To address this gap, we introduce dLLM, an open-source framework that unifies the core components of diffusion language modeling -- training, inference, and evaluation -- and makes them easy to customize for new designs. With dLLM, users can reproduce, finetune, deploy, and evaluate open-source large DLMs such as LLaDA and Dream through a standardized pipeline. The framework also provides minimal, reproducible recipes for building small DLMs from scratch with accessible compute, including converting any BERT-style encoder or autoregressive LM into a DLM. We also release the checkpoints of these small DLMs to make DLMs more accessible and accelerate future research.

Figures & tables

Panel of Figure 1 — Figure 1: A unified trainer interface supports a variety of purposes via modular trainers and configuration changes. Figure 1(a) shows the MDLM pretraining setup. Figure 1(b) shows the single-line trainer swap from MDLMTrainer to BD3LMTrainer . Figure 1(c) shows the minimal changes to use MDLMTrainer for SFT: NoAttentionMaskWrapper keeps padding EOS visible, and label_pad_token_id=eos_token_id trains the model to generate EOS from extra mask tokens in inputs. Figure 1(d) shows the minimal changes to adapt an autoregressive LM to MDLM: right_shift_logits reuses next-token prediction, and PrependBOSWrapper prepends BOS to provide the predictions for the first mask token.

Panel of Figure 2 — Figure 2: Inference pipeline: sampler swap from vanilla to FastdLLM MDLM sampler.

Panel of Figure 3 — Figure 3: Terminal Visualizer showing transition from masked to decoded tokens.

Panel of Figure 4 — Figure 4: Fast-dLLM evaluation results with max new tokens @ $256$ and $512$ . Model selection follows the original Fast-dLLM evaluation for consistency and fair comparison. Cache uses block-wise approximate KV caching within each decoding block; Parallel uses confidence-based parallel token updates; Cache & Parallel combines both. Note that max new tokens determines the number of pre-allocated padding tokens in the bidirectional context window, therefore affecting compute and measured performance.

Panel of Figure 5 — Figure 5: Sensitivity to decoding hyperparameters. We vary individual sampling hyperparameters at inference time and observe that performance can degrade sharply from the optimal configuration. Baseline denotes the best-performing setting; Suppress does not suppress <eos> from the beginning of generation; CFG sets cfg=0.5 ; Parallel @ $4$ generates four tokens per step; and Temp @ $0$ sets temperature=0.0 .

Model	GSM8K	MATH500	Countdown	Sudoku	HumanEval	MBPP
LLaDA-Instruct	$79.91$	$34.80$	$19.92$	$11.62$	$35.37$	$42.02$
+ SFT (MDLM)	$80.59$	$35.40$	$26.95$	$14.36$	$36.59$	$43.97$
LLaDA-Base	$64.67$	$10.20$	$10.16$	$0.34$	$25.00$	$40.08$
+ SFT (MDLM)	$73.62$	$17.40$	$8.98$	$0.00$	$18.90$	$28.79$
Dream-Instruct	$64.44$	$28.20$	$22.27$	$6.93$	$35.98$	$44.36$
+ SFT (MDLM)	$70.43$	$32.80$	$20.31$	$19.68$	$37.20$	$45.53$

Table 1: MDLM SFT evaluation results . Instruct models show consistent gains, Base models gain on in-distribution math but may regress on out-of-distribution planning and coding.

Model	GSM8K	BBH	MATH	MMLU	HellaSwag	LAMBADA	WinoGrande
ModernBERT-base-chat-v0.1	$3.6$	$21.1$	$3.1$	$26.2$	$34.5$	$49.3$	$48.8$
ModernBERT-large-chat-v0.1	$9.3$	$25.6$	$3.6$	$29.6$	$40.9$	$46.3$	$49.0$
Qwen1.5-0.5B	$22.0$	$18.3$	$3.1$	$39.2$	$48.2$	$48.6$	$55.0$
Qwen1.5-0.5B-Chat	$11.3$	$18.2$	$2.1$	$35.0$	$36.9$	$41.2$	$52.0$
GPT-2	$0.7$	$6.9$	$1.8$	$22.9$	$31.1$	$46.0$	$51.6$
GPT-2-medium	$2.1$	$17.8$	$1.4$	$22.9$	$39.4$	$55.5$	$53.1$

Table 2: ModernBERT-Chat evaluation results. ModernBERT-Chat ( Warner et al., 2025 ) and GPT-2 ( Radford et al., 2019 ) models are evaluated with dLLM ’s pipeline; Qwen1.5 ( Bai et al., 2023 ) numbers are reported by original sources. See Figure 7 for training curves.

Model	GSM8K	BBH	MATH	MMLU	MMLU-Pro	HellaSwag	HumanEval	MBPP
Qwen3-0.6B-mdlm-v0.1	$29.3$	$26.7$	$8.7$	$40.0$	$17.3$	$42.1$	$30.5$	$29.2$
Qwen3-0.6B-bd3lm-v0.1	$46.3$	$26.6$	$12.9$	$39.1$	$13.8$	$39.3$	$46.3$	$38.2$
Qwen2.5-0.5B	$41.6$	$20.3$	$19.5$	$47.5$	$15.7$	$52.1$	$30.5$	$39.3$
Qwen3-0.6B-Base	$59.6$	$41.5$	$32.4$	$52.8$	$24.7$	$47.4$	$32.3$	$36.6$

Table 3: Qwen-A2D evaluation results. MDLM and BD3LM models are evaluated with dLLM ’s pipeline; Autoregressive Qwen2.5/Qwen3 baselines are reported by original sources ( Qwen Team et al., 2024 ; Yang et al., 2025 ) . See Figure 8 for training curves.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Training loss for finetuning open-weight DLMs to reason (Section 4.1 ).

Figure 7: Training loss for finetuning BERT to chat (Section 4.2.1 ).

Figure 8: Training loss for finetuning autoregressive LMs to be DLMs (Section 4.2.2 ).

	MMLU	BBH	ARC-C	Hella.	WinoG.	PIQA	GSM8K	MATH	GPQA	HEval	MBPP
Official	65.9	49.7	45.9	70.5	74.8	73.6	70.3	31.4	25.2	35.4	40.0
dLLM	65.9	47.2	44.1	69.2	70.4	70.7	70.7	32.4	31.9	32.9	38.8

Appendix

Table 4: LLaDA evaluation results. “Official” denotes results from the original paper; “ dLLM ” denotes results from our dLLM reimplementation. “Hella.” stands for HellaSwag, “HEval” stands for HumanEval and “WinoG.” stands for WinoGrande.

	MMLU	BBH	ARC-C	Hella.	WinoG.	PIQA	GSM8K	MATH	GPQA	HEval	MBPP
Official	69.5	57.9	59.9	73.3	74.8	75.8	77.2	39.6	36.6	57.9	56.2
dLLM	70.0	63.7	59.0	73.5	72.5	76.4	77.0	42.4	34.6	56.7	56.0

Appendix

Table 5: Dream evaluation results. “Official” denotes results from the original paper; “ dLLM ” denotes results from our dLLM reimplementation. “Hella.” stands for HellaSwag, “HEval” stands for HumanEval and “WinoG.” stands for WinoGrande.

Benchmark	Source	Baseline		+Cache		+Parallel		+Both
		Acc	Tok/s ( $\times$ )	Acc	( $\times$ )	Acc	( $\times$ )	Acc	( $\times$ )
GSM8K	Official	$79.3$	$6.7\ (1.0\times)$	$79.5$	$3.2\times$	$79.2$	$2.5\times$	$78.5$	$8.1\times$
	dLLM	$78.0$	$8.1\ (1.0\times)$	$78.2$	$3.2\times$	$78.9$	$2.3\times$	$78.0$	$6.5\times$
MATH	Official	$33.5$	$9.1\ (1.0\times)$	$33.3$	$2.6\times$	$33.4$	$2.7\times$	$33.2$	$5.7\times$
	dLLM	$38.3$	$9.7\ (1.0\times)$	$37.6$	$2.7\times$	$38.6$	$2.0\times$	$37.5$	$5.0\times$
HumanEval	Official	$41.5$	$30.5\ (1.0\times)$	$42.7$	$1.3\times$	$43.9$	$3.3\times$	$43.3$	$3.7\times$

Appendix

Table 6: Fast-dLLM LLaDA-Instruct evaluation results with max new tokens @

256

(a) and

512

(b). “Official” denotes results from the Fast-dLLM paper; “ dLLM ” denotes results from our dLLM reimplementation.

Benchmark	Source	Baseline		+Cache		+Parallel		+Both
		Acc	Tok/s ( $\times$ )	Acc	( $\times$ )	Acc	( $\times$ )	Acc	( $\times$ )
GSM8K	Official	$75.0$	$9.1\ (1.0\times)$	$74.3$	$3.6\times$	$74.2$	$1.6\times$	$74.8$	$5.3\times$
	dLLM	$75.4$	$9.0\ (1.0\times)$	$75.0$	$3.7\times$	$72.6$	$1.4\times$	$74.2$	$4.7\times$
MATH	Official	$38.4$	$11.4\ (1.0\times)$	$36.8$	$3.0\times$	$37.9$	$2.4\times$	$37.6$	$5.9\times$
	dLLM	$31.5$	$25.1\ (1.0\times)$	$33.3$	$1.5\times$	$23.5$	$2.1\times$	$31.1$	$3.1\times$
HumanEval	Official	$49.4$	$23.3\ (1.0\times)$	$53.7$	$1.5\times$	$49.4$	$2.0\times$	$54.3$	$2.8\times$

Appendix

Table 7: Fast-dLLM Dream-Base evaluation results with max new tokens @

256

(a) and

512

(b). “Official” denotes results from the Fast-dLLM paper; “ dLLM ” denotes results from our dLLM reimplementation.

Diffusion Language Model Training

See all 40 papers

DARE: Diffusion Large Language Models Alignment and Reinforcement Executor
Apr 5, 2026Jingyi Yang, Yuxian Jiang, Xuhao Hu +3Diffusion Language Models Diffusion Model Alignment
Shanghai Artificial Intelligence Laboratory · Fudan University · Zhejiang University +1
ES-dLLM: Efficient Inference for Diffusion Large Language Models by Early-Skipping
Mar 10, 2026Zijian Zhu, Fei Ren, Zhanhong Tan +1Diffusion Language Model Inference Diffusion Language Models
IIIS, Tsinghua University · Polar Bear Tech.
DARE: Diffusion Language Model Activation Reuse for Efficient Inference
May 1, 2026Natalia Frumkin, Bokun Wang, Hung-Yueh Chiang +3Self-Attention Diffusion Language Model Inference
Chandra Family Department of Electrical and Computer Engineering The University of Texas at Austin · Cornell University

Apr 5, 2026J/K move · Enter open · S save

DARE: Diffusion Large Language Models Alignment and Reinforcement Executor

Jingyi Yang, Yuxian Jiang, Xuhao Hu, Shuang Cheng+2

Shanghai Artificial Intelligence Laboratory · Fudan University · Zhejiang University+1

Diffusion Language Models Diffusion Model Alignment RL for Diffusion Models

Open paper AskPDF arXiv

Diffusion large language models (dLLMs) are emerging as a compelling alternative to dominant autoregressive models, replacing strictly sequential token generation with iterative denoising and parallel generation dynamics. However, their open-source ecosystem remains fragmented across model families and, in particular, across post-training pipelines, where reinforcement learning objectives, rollout implementations and evaluation scripts are often released as paper-specific codebases. This fragmentation slows research iteration, raises the engineering burden of reproduction, and makes fair comparison across algorithms difficult. We present DARE (dLLMs Alignment and Reinforcement Executor), an open framework for post-training and evaluating dLLMs. Built on top of verl~\cite{sheng2024hybridflow} and OpenCompass~\cite{2023opencompass}, DARE unifies supervised fine-tuning, parameter-efficient fine-tuning, preference optimization, and dLLM-specific reinforcement learning under a shared execution stack for both masked and block diffusion language models. Across representative model families including LLaDA, Dream, SDAR, and LLaDA2.x, DARE provides broad algorithmic coverage, reproducible benchmark evaluation, and practical acceleration. Extensive empirical results position that DARE serves as a reusable research substrate for developing, comparing, and deploying post-training methods for current and emerging dLLMs.

Abstract

Diffusion Language Model Training

Explore similar work

DARE: Diffusion Large Language Models Alignment and Reinforcement Executor

ES-dLLM: Efficient Inference for Diffusion Large Language Models by Early-Skipping

DARE: Diffusion Language Model Activation Reuse for Efficient Inference