Hybrid optimization and DRL Framework for Edge LLMs

Deploying large language models (LLMs) and deep learning architectures at the network edge faces substantial computational, latency, and energy constraints. This paper introduces a hybrid optimization and distributed reinforcement learning (DRL) framework designed to coordinate multiple lightweight subsystems of an edge-based LLM. The proposed approach integrates a multi-stage optimization pipeline (architecture search, knowledge distillation, pruning, and quantization) with a distributed multi-agent DRL controller (MADDPG). The objective is to balance precision, latency, and energy efficiency dynamically across edge nodes. Experimental results on benchmark vision and NLP datasets demonstrate significant improvements in performance compared to stand-alone optimization techniques, while maintaining near-baseline accuracy.