LEAP: Training-Free Method Boosts Parallel Decoding in Diffusion Language Models

A new method called LEAP (Lookahead Early-Convergence Token Detection for Accelerated Parallel Decoding) has been introduced to improve the parallel processing capabilities of Diffusion Language Models (dLLMs). Current dLLMs rely on conditional independence assumptions with high confidence thresholds to maintain accuracy, but these thresholds are overly conservative, limiting scalability. Through token-level statistical analysis, researchers found that many tokens converge to correct predictions early in the denoising process without meeting standard confidence criteria. LEAP is a training-free, plug-and-play technique that uses future context filtering to detect these early-convergent tokens, enabling more efficient parallel decoding without sacrificing accuracy. The method was detailed in a paper published on arXiv (ID: 2605.10980).