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

In three linesVoidPadding introduces a dedicated [VOID] token for padding in masked diffusion language models (MDLMs), freeing [EOS] for semantic termination. On Dream-7B-Instruct, it improves mathematical reasoning and code generation benchmarks by +17.84 points over baseline and +6.95 over RainbowPadding, reducing NFE by 55.7%.

Summary generated by Claude — human-verified