DMax: Aggressive Parallel Decoding for dLLMs — part 3/3

Cleaned-text code points [31061, 31630). Bibliography omitted here; source.md includes it.

In this paper, we present DMax, a novel paradigm for efficient diffusion language models that mitigates error accumulation for parallel decoding. DMax enables aggressive decoding parallelism while preserving the accuracy of the original model. We introduce two key components of our approach, namely On-Policy Uniform Training and Soft Parallel Decoding, and demonstrate their effectiveness through extensive experiments on diverse benchmarks. Our results establish a strong new baseline for parallel decoding in dLLMs and suggest a promising new direction for dLLMs.