LLM-Enhanced Deep Reinforcement Learning for Task Offloading in Collaborative Edge Computing

Researchers Hao Guo, Kaixiang Xv, Ziwu Ge, and Lei Yang have introduced LeDRL, a hybrid decision framework that fuses a lightweight large language model (LLM) with self-attention-enhanced deep reinforcement learning (DRL) for dynamic task offloading in collaborative edge computing. Collaborative edge computing relies on geographically distributed nodes to execute tasks, but unpredictable failures and latency demands make real-time offloading decisions challenging. While DRL suffers from sample inefficiency and local optima, and LLMs lack real-time decision-making speed, LeDRL bridges the gap. It uses structured, context-aware prompts that encode node status, task semantics, and link dynamics to generate high-level strategy priors. A self-attention-based alignment module selectively processes these priors for context-aware policy optimization, and a reflective evaluator distills semantic feedback from past trajectories to improve future LLM queries.

Extensive experiments demonstrate LeDRL outperforming baselines in task success rate, convergence speed, and real-time responsiveness across various network scales, achieving over a 17% improvement in success rate. The team also deployed LeDRL on Jetson-based edge devices using their prototype system CoEdgeSys, proving its robustness and feasibility under hardware constraints. The open-source code is available on arXiv. This work shows how combining LLMs' semantic reasoning with DRL's sequential decision-making can enable more reliable and efficient edge computing systems, particularly important for IoT, autonomous systems, and smart infrastructure where low latency and fault tolerance are critical.