LLM-Guided Dual-Timescale Meta-Regulated MARL for Intelligent Low-Altitude Economy Networks

Abstract

Intelligent low-altitude economy networks demand reliable, low delay, and energy-efficient unmanned aerial vehicle (UAV) collaboration under highly dynamic wireless conditions and stringent resource constraints. This setting induces a multi-objective coordination problem with non-trivial trade-offs. To address it, we propose a large language model (LLM)-guided meta-regulation framework for multi-agent reinforcement learning (MARL). The framework adopts a dual-timescale architecture: MARL agents conduct fast-timescale policy optimization, while a slow-timescale meta-controller adaptively tunes reward weights to steer learning toward desirable trade-offs among reliability, delay, and energy consumption. In addition, we introduce a meta early-exit mechanism to balance training stability, efficiency, and peak performance. Extensive simulations show that the proposed method delivers faster convergence and smoother learning dynamics than fixed-weight MARL baselines, improves task success rate and delay, and remains robust as the UAV swarm scales. Overall, LLM-guided meta-regulation offers an effective and flexible solution for multi-objective coordination in intelligent low-altitude networking scenarios.