SimSD: Simple Speculative Decoding Closes the Speed Gap for Diffusion Language Models

arXiv·high signal

Presents SimSD, a speculative decoding method for diffusion large language models (dLLMs) that bridges the inference speed gap with autoregressive models. Unlike prior work requiring separate draft models or expensive verification trees, SimSD uses a simple approach within the diffusion framework itself with code released. As dLLMs like MDLM and SEDD gain traction as alternatives to autoregressive generation, this makes them viable for latency-sensitive deployment.