Progressive Refinement Regulation for Accelerating Diffusion Language Model Decoding
This work provides a method to accelerate text generation for users of diffusion language models, offering strong specific gains in decoding speed.
The paper addresses the inefficiency of diffusion language models, which apply a uniform refinement rule to all tokens despite varying stabilization rates. They propose Progressive Refinement Regulation (PRR), a framework that learns a token-wise controller to regulate refinement, significantly accelerating decoding while maintaining generation quality.
Diffusion language models generate text through iterative denoising under a uniform refinement rule applied to all tokens. However, tokens stabilize at different rates in practice, leading to substantial redundant refinement and motivating refinement control over the denoising process. Existing approaches typically assess refinement necessity from instantaneous, step-level signals under a fixed decoding process. In contrast, whether a token has converged is defined by how its prediction changes along its future refinement trajectory. Moreover, changing the refinement rule reshapes future refinement trajectories, which in turn determine how refinement rules should be formulated, making refinement control inherently dynamic. We propose \emph{Progressive Refinement Regulation} (PRR), a progressive, trajectory-grounded refinement control framework that derives a token-level notion of empirical convergence progress from full decoding rollouts. Based on this signal, PRR learns a lightweight token-wise controller to regulate refinement via temperature-based distribution shaping under a progressive self-evolving training scheme. Experiments show that PRR substantially accelerates diffusion language model decoding while preserving generation quality.