📄 中文摘要
扩散语言模型通过在统一的细化规则下进行迭代去噪生成文本。然而，实际中不同的词元以不同的速度稳定，导致了大量冗余的细化，进而促使对去噪过程的细化控制。现有方法通常基于固定解码过程下的瞬时、步级信号来评估细化的必要性。与此不同，词元是否收敛是由其在未来细化轨迹中的预测变化来定义的。此外，改变细化规则会重塑未来的细化轨迹，这反过来又决定了细化规则的制定，使得细化控制本质上是动态的。该研究提出了一种新的细化调控方法，旨在提高扩散语言模型的解码效率。
加速扩散语言模型解码的渐进细化调控
📄 English Summary
Progressive Refinement Regulation for Accelerating Diffusion Language Model Decoding
Diffusion language models generate text through iterative denoising under a uniform refinement rule applied to all tokens. However, in practice, tokens stabilize at different rates, leading to substantial redundant refinement and motivating the need for refinement control over the denoising process. Existing approaches typically assess the necessity of refinement based on instantaneous, step-level signals within a fixed decoding process. In contrast, the convergence of a token is defined by how its prediction changes along its future refinement trajectory. Moreover, altering the refinement rule reshapes future refinement trajectories, which in turn determines how refinement rules should be formulated, making refinement control inherently dynamic. This research proposes a new refinement control method aimed at enhancing the decoding efficiency of diffusion language models.