Diffusion Language Models (DLMs) present a non‑autoregressive generation paradigm that refines tokens through iterative denoising instead of left‑to‑right decoding. Compared with Transformer‑based autoregressive Large Language Models (LLMs), DLMs can update multiple uncertain tokens in parallel and exploit bidirectional context throughout generation, enabling more flexible quality‑latency trade‑offs beyond fixed sequential decoding.
These properties are especially attractive for mobile edge agents:
- Partial refinement and early exit can dramatically cut response latency;
- Constraint‑guided correction improves robustness under noisy, incomplete, or dynamic contexts;
- Parallel updates reduce reliance on high‑bandwidth communication.
This survey reviews the foundations of DLMs and evaluates their suitability for edge settings under constraints such as latency, memory, energy, bandwidth, privacy, and reliability. Key topics include:
- Resource‑efficient architectures and hardware accelerators;
- Training and inference acceleration techniques like mixed precision, gradient checkpointing, and distributed denoising;
- Model compression (pruning, quantization, distillation) and trade‑offs in edge‑cloud co‑deployment;
- Communication‑aware serving that adapts to IoT and wireless bandwidth fluctuations;
- Evaluation metrics for DLM‑driven agents covering generation quality, response time, and resource usage.
The paper also highlights open challenges: long‑context state management, split inference across devices, trustworthy execution environments, multimodal grounding, and reproducible benchmarking. The goal is to align DLM properties—bidirectionality, parallel refinement, controllability, and quality‑latency elasticity—with system‑level requirements of future mobile edge intelligence.
Review