DC-Leap: Training-Free Acceleration of dLLMs via Draft-Guided Contiguous Leaping Decoding

While parallel decoding is central to the efficiency of Diffusion Large Language Models (dLLMs), current strategies are often hindered by overly conservative confidence thresholds. These thresholds, necessitated by the Joint Probability Dependence Error (JPDE), result in redundant denoising iterations and suboptimal inference speeds. To overcome this, we propose DC-Leap, a training-free framework that enables reliable acceleration of dLLMs in the moderate-confidence regime. DC-Leap introduces a Dynamic Contiguous Verification strategy that integrates strictly-ordered causal constraints into the parallel decoding process. By progressively validating token dependencies, this mechanism effectively neutralizes the JPDE, enabling reliable acceleration with comparable performance. Furthermore, DC-Leap incorporates the draft-guided decoding mechanism, where the draft helps extend the context by leaping forward across multiple tokens, providing look-ahead context and retaining the structural benefits of bidirectional attention during inference. Extensive experiments on standard benchmarks demonstrate that DC-Leap achieves substantial speedups, up to **53.19$\times$** on MBPP for long-sequence generation, and up to **105.02$\times$** when combined with KV-Cache with comparable generation quality. Code is available at .

Parallel decoding strategies for Diffusion Large Language Models (dLLMs) speed up generation by predicting multiple tokens simultaneously. However, to mitigate potential errors, these strategies use strict confidence thresholds, which force them to discard many potentially correct tokens, thereby bottlenecking the efficiency. Our work introduces DC-Leap, a training-free strategy to accelerate parallel decoding. We identify that much of the inefficiency stems from failing to properly account for the relationship between tokens decoded in a single forward process. DC-Leap addresses this by enforcing a sequential verification process that validates tokens within a dynamic window. Additionally, we use a draft-guided mechanism that allows the model to be aware of the future context, ensuring more stable predictions. By implementing these changes, DC-Leap increases the generation speed of dLLMs significantly across three models and five benchmarks without compromising accuracy. This makes dLLMs significantly more efficient for practical, real-world use.