Edge-Assisted Federated Learning for Large Language Models in IoT Sensor Systems

Federated learning (FL) enables privacy-preserving collaborative training of large-scale AI models, including large language models (LLMs), and is particularly attractive for IoT sensor systems with stringent privacy and resource constraints. However, fine-tuning LLMs in sensor-driven edge environments remains challenging due to limited computation and memory resources on edge devices, as well as heterogeneous and time-varying communication and computational conditions that exacerbate the straggler problem. To address these challenges, we propose HRL-FLLM, a hierarchical client–edge–cloud FL framework driven by deep reinforcement learning (DRL). HRL-FLLM allows edge servers to adaptively coordinate LLM fine-tuning by jointly controlling efficiency-enhancing techniques, including Low-Rank Adaptation, quantization, and sparse communication, thereby reducing computational overhead and communication costs while preserving model performance. We further design a convergence-aware reward function and formulate a training time budget optimization problem, enabling edge servers to learn effective local training time allocation policies under heterogeneous conditions. Extensive experiments show that HRL-FLLM achieves robust convergence and significantly reduces training latency. In large-scale heterogeneous FL scenarios with resource-constrained clients, the proposed framework reduces convergence time by up to 47% compared with conventional cloud-centric FL approaches, demonstrating its effectiveness as a scalable and sustainable edge intelligence solution for IoT sensor systems.