A Survey on LLM Edge-Intelligence: Recent Advances and Open Challenges

초록

Large language models (LLMs) are increasingly deployed on edge devices to reduce latency, bandwidth, and privacy risks. This survey reviews recent techniques that enable efficient edge-LLM inference and fine-tuning: (1) memoryefficient model architectures, (2) edge-aware inference orchestration, and (3) privacy-preserving fine-tuning. Representative works are analyzed for their algorithmic contributions and empirical gains in memory, latency, and accuracy. We identify three core challenges: (1) generalizing compression policies across diverse LLM families; (2) designing lightweight, online orchestration that jointly optimizes latency, bandwidth, memory, and energy under dynamic conditions; and (3) ensuring privacy-preserving, adaptive fine-tuning without catastrophic forgetting. The paper concludes with a roadmap for unified, end-to-end frameworks that balance resource constraints, performance, and security in practical Edge-LLM deployments.

키워드