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

2026-07-14

AI Summary

Although diffusion-based large language models offer potential for parallel generation, their inference efficiency is hindered by the tight coupling of algorithmic, architectural, and system-level factors, making effective acceleration challenging. This work proposes a unified latency decomposition framework that, for the first time, systematically disentangles the multidimensional determinants of inference speed and establishes a comprehensive taxonomy encompassing algorithmic innovations, architecture and system optimizations, and inference-time scaling techniques. By holistically evaluating diffusion-aware caching, parallel scheduling, and deployment strategies, the study quantifies the practical contributions of each approach to end-to-end latency, provides a reproducible benchmarking methodology, clarifies fundamental distinctions and synergies among methods, and offers both theoretical guidance and practical pathways toward efficient deployment, while highlighting key challenges for future research.

Abstract

Diffusion large language models (dLLMs) offer a theoretical advantage in parallel generation over standard autoregressive models. However, parallel generation alone does not guarantee practical speedups. Realizing this efficiency requires specialized inference mechanisms, such as diffusion-aware caching and reuse. Consequently, as inference efficiency becomes a prerequisite for practical deployment, recent research has actively explored acceleration techniques across algorithms, architectures, and systems. However, rigorous comparisons remain difficult, as end-to-end latency stems from intricate trade-offs between algorithmic, architectural, and system-level factors that are often conflated in existing benchmarks. In this survey, we introduce a unified latency decomposition framework for dLLMs to disentangle these factors and analyze their impact on inference speed in real deployments. Guided by this framework, we categorize acceleration techniques along three axes covering algorithmic innovations, architectural and system optimizations, and inference-time scaling. Finally, we provide guidelines for reproducible benchmarking and highlight open challenges for realizing the full potential of parallel generation.

Problem

• masked diffusion LLMs
• inference efficiency
• latency decomposition
• benchmarking
• parallel generation

Innovation

• diffusion large language models
• inference acceleration
• latency decomposition
• parallel generation
• efficient inference