Accelerating Masked Diffusion Large Language Models: A Survey of Efficient Inference Techniques

This survey, arXiv:2607.12829, examines efficient inference techniques for masked diffusion large language models (dLLMs). It argues that parallel generation does not automatically produce practical end-to-end speedups because latency also depends on diffusion-aware caching, reuse, architecture, and system behavior. The paper introduces a unified latency decomposition framework intended to separate these factors in real deployments. It organizes acceleration methods into algorithmic innovations, architectural and system optimizations, and inference-time scaling, then proposes reproducible benchmarking guidelines and identifies open challenges for making parallel generation operationally useful.