DC-Leap Accelerates Diffusion LLMs Up to 53× on MBPP Without Any Training

arXiv 2607.20467·medium signalXBlueskyLinkedInCopy link

Jiao, Wu, Sun and colleagues identify Joint Probability Dependence Error as the reason diffusion LLM parallel decoding is stuck with overly conservative confidence thresholds, forcing redundant denoising iterations. DC-Leap adds a Dynamic Contiguous Verification strategy that folds strictly-ordered causal constraints into parallel decoding to neutralize JPDE, plus a draft-guided mechanism that leaps forward across multiple tokens to supply look-ahead context while keeping bidirectional attention's structural benefit. The reported speedups reach 53.19× on MBPP for long-sequence generation and 105.02× combined with KV-cache at comparable generation quality — entirely training-free, which makes it applicable to existing dLLM checkpoints.