VoidPadding: Let [VOID] Handle Padding in Masked Diffusion Language Models so that [EOS] Can Focus on Semantic Termination

MDLMs generate text by denoising a preallocated masked response canvas, making response-length modeling central to instruction tuning. Existing MDLMs often inherit the autoregressive convention of using repeated $\mathtt{\text{[EOS]}}$ tokens for padding during instruction tuning, giving $\mathtt{\text{[EOS]}}$ a dual role as both a semantic terminator and a padding token. We show that this dual role is a root cause of $\mathtt{\text{[EOS]}}$ overflow under large-block decoding. To decouple these roles, we propose VoidPadding, which introduces $\mathtt{\text{[VOID]}}$ for padding and reserves $\mathtt{\text{[EOS]}}$ for termination. During inference, the learned $\mathtt{\text{[EOS]}}$ signal enables early stopping, while the learned $\mathtt{\text{[VOID]}}$ signal guides adaptive response canvas expansion. On Dream-7B-Instruct, VoidPadding improves the block-size-averaged four-task mean across mathematical reasoning and code generation benchmarks by $+17.84$ points over the original model and $+6.95$ points over RainbowPadding, while reducing decoding NFE by 55.7% on average. Code is available at https://github.com/Haru-LCY/VoidPadding.