Saved in:
Bibliographic Details
Main Authors: Liu, Yue, Zhao, Yuzhong, Xie, Zheyong, Ye, Qixiang, Jiao, Jianbin, Hu, Yao, Cao, Shaosheng, Liu, Yunfan
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2602.01362
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918318167818240
author Liu, Yue
Zhao, Yuzhong
Xie, Zheyong
Ye, Qixiang
Jiao, Jianbin
Hu, Yao
Cao, Shaosheng
Liu, Yunfan
author_facet Liu, Yue
Zhao, Yuzhong
Xie, Zheyong
Ye, Qixiang
Jiao, Jianbin
Hu, Yao
Cao, Shaosheng
Liu, Yunfan
contents In discrete generative modeling, two dominant paradigms demonstrate divergent capabilities: Masked Diffusion Language Models (MDLM) excel at semantic understanding and zero-shot generalization, whereas Uniform-noise Diffusion Language Models (UDLM) achieve strong few-step generation quality, yet neither attains balanced performance across both dimensions. To address this, we propose XDLM, which bridges the two paradigms via a stationary noise kernel. XDLM offers two key contributions: (1) it provides a principled theoretical unification of MDLM and UDLM, recovering each paradigm as a special case; and (2) an alleviated memory bottleneck enabled by an algebraic simplification of the posterior probabilities. Experiments demonstrate that XDLM advances the Pareto frontier between understanding capability and generation quality. Quantitatively, XDLM surpasses UDLM by 5.4 points on zero-shot text benchmarks and outperforms MDLM in few-step image generation (FID 54.1 vs. 80.8). When scaled to tune an 8B-parameter large language model, XDLM achieves 15.0 MBPP in just 32 steps, effectively doubling the baseline performance. Finally, analysis of training dynamics reveals XDLM's superior potential for long-term scaling. Code is available at https://github.com/MzeroMiko/XDLM
format Preprint
id arxiv_https___arxiv_org_abs_2602_01362
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Balancing Understanding and Generation in Discrete Diffusion Models
Liu, Yue
Zhao, Yuzhong
Xie, Zheyong
Ye, Qixiang
Jiao, Jianbin
Hu, Yao
Cao, Shaosheng
Liu, Yunfan
Computation and Language
In discrete generative modeling, two dominant paradigms demonstrate divergent capabilities: Masked Diffusion Language Models (MDLM) excel at semantic understanding and zero-shot generalization, whereas Uniform-noise Diffusion Language Models (UDLM) achieve strong few-step generation quality, yet neither attains balanced performance across both dimensions. To address this, we propose XDLM, which bridges the two paradigms via a stationary noise kernel. XDLM offers two key contributions: (1) it provides a principled theoretical unification of MDLM and UDLM, recovering each paradigm as a special case; and (2) an alleviated memory bottleneck enabled by an algebraic simplification of the posterior probabilities. Experiments demonstrate that XDLM advances the Pareto frontier between understanding capability and generation quality. Quantitatively, XDLM surpasses UDLM by 5.4 points on zero-shot text benchmarks and outperforms MDLM in few-step image generation (FID 54.1 vs. 80.8). When scaled to tune an 8B-parameter large language model, XDLM achieves 15.0 MBPP in just 32 steps, effectively doubling the baseline performance. Finally, analysis of training dynamics reveals XDLM's superior potential for long-term scaling. Code is available at https://github.com/MzeroMiko/XDLM
title Balancing Understanding and Generation in Discrete Diffusion Models
topic Computation and Language
url https://arxiv.org/abs/2602.01362