Trajectory-Level Speculative Decoding for Diffusion Language Models

Diffusion-based language models (dLLMs) enable parallel token generation through iterative denoising, but existing decoding strategies collapse to single-token generation under low confidence, severely limiting throughput. Unlike autoregressive models where speculative decoding operates on token sequences in a fixed left-to-right order, dLLMs require speculating over \emph{denoising trajectories}—sequences of multi-token updates with explicit positions and unmasking orders. We develop a trajectory-level speculative framework that constructs draft denoising trajectories via confidence-stratified tree exploration and verifies them through blockwise parallel evaluation with bidirectional attention masking. Our method further introduces inter-block speculation, exploiting diffusion models' bidirectional structure to perform cross-block lookahead. We formally characterize when this approach is exact and identify trajectory drift as the fundamental cost of increased parallelism. Building on Fast-dLLM's dual-cache infrastructure, our framework reduces denoising iterations by 30-40\% and increases tokens-per-step from 2.6 to 4.3, achieving 7-14$\times$ speedup over vanilla dLLMs and 1.3$\times$ over Fast-dLLM with less than 1\% accuracy change across reasoning and code benchmarks.

Modern diffusion-based large language models can generate multiple tokens at once to boost efficiency, unlike traditional models that generate one token at a time. However, these advanced models frequently suffer from performance degradation: they fall back to slow single-token generation when they lack prediction confidence, greatly reducing their overall processing speed and practical usability. Speculative decoding is an effective method to alleviate the low-efficiency issue of LLM generation. However, existing speculative decoding techniques are primarily designed for traditional left-to-right text generation tasks. They cannot adapt to the unique iterative denoising process of diffusion-based language models. To solve this problem, we design a new acceleration framework that predicts and verifies complete model generation trajectories, rather than individual tokens. We adopt a structured exploration strategy to generate reliable draft generation paths and use segmented parallel verification to improve efficiency, while leveraging the model’s bidirectional structural features to enable cross-step preview prediction. We also clarify the exact valid scenarios of our method and identify the core trade-off between parallel speedup and minor trajectory deviation. Extensive tests on reasoning and coding tasks show our approach cuts required model iterations by 30–40% and greatly improves single-step generation volume. It delivers 7–14 times faster speed than original diffusion-based models and 1.3 times faster than the state-of-the-art accelerated method, Fast-dLLM, with nearly no loss in generation accuracy.