This survey addresses the gap between the theoretical parallel generation advantage of diffusion large language models (dLLMs) and their practical inference speed. It introduces a unified latency decomposition framework designed to disentangle algorithmic, architectural, and system-level factors that impact real-world deployment.
The authors categorize acceleration techniques along three axes: algorithmic innovations, architectural and system optimizations, and inference-time scaling. The work provides guidelines for reproducible benchmarking to address the difficulty of rigorous comparisons in existing studies.
By analyzing these trade-offs, the survey highlights open challenges for realizing the full potential of parallel generation in dLLMs.