Locality-aware Diffusion Language Modeling
Published in arXiv preprint, 2026, 2026
Studies the trainability of Masked Diffusion Language Models against autoregressive LLMs across structured generation tasks. Proposes two locality-aware blockwise diffusion architectures (Scatter and Jigsaw) bridging AR-style local ordering and Diffusion-style global iterative refinement.
Recommended citation: Yuxiang Wang, et al. "Locality-aware Diffusion Language Modeling." arXiv:2604.24832, 2026.
Download Paper
