Accelerating Inference in Diffusion Large Language Models: A Survey

Diffusion Large Language Models (dLLMs) have emerged as a promising non-autoregressive paradigm, offering the ability to capture bidirectional context and the potential for parallel generation. However, in practice, current open-source dLLMs often incur high inference costs. These costs primarily stem from two critical bottlenecks: the high computational redundancy caused by the difficulty of applying KV caching to dynamic bidirectional representations, and the generation quality degradation inherent in parallel decoding. Motivated by the critical need to resolve these specific bottlenecks, this survey provides a comprehensive overview of inference acceleration techniques for dLLMs by proposing a unified taxonomy that categorizes existing approaches, reviewing techniques within each category, analyzing their underlying design trade-offs between speed and generation fidelity, and highlighting future directions. This survey aims to provide a structured roadmap for researchers and practitioners seeking to develop faster and more reliable diffusion large language models.