Dream-7B-Instruct

VoidPadding decouples padding and termination tokens in masked diffusion language models

VoidPadding introduces a dedicated [VOID] token for padding in masked diffusion language models (MDLMs), separating it from the [EOS] semantic termination role that existing models conflate. The authors show this dual-role conflation causes EOS overflow under large-block decoding, and their fix enables early stopping and adaptive canvas expansion at inference time. On Dream-7B-Instruct, VoidPadding improves a four-task mean across math reasoning and code generation by +17.84 points over the baseline and reduces decoding NFE by 55.7%.