Large Language Model-enhanced Reinforcement Learning for Low-Altitude Economy Networking

TL;DR

This work addresses the challenge of enabling robust, energy-efficient aerial networking below 1,000 meters (LAENet) by integrating large language models (LLMs) with reinforcement learning (RL). It presents a tutorial and a novel LLM-enhanced RL framework in which LLMs act as information processors, reward designers, decision-makers, and generators to overcome RL limitations in generalization, reward design, and stability. A case study demonstrates that LLM-designed rewards improve learning efficiency and energy performance in a UAV-assisted IoT scenario, achieving up to 7.2% lower final energy for TD3 and significant gains across packet sizes. The paper highlights future directions toward modular LLM-RL agents, memory-enabled planning, and multi-agent LLM collaboration to enable scalable, intelligent aerial networking systems.

Abstract

Low-Altitude Economic Networking (LAENet) aims to support diverse flying applications below 1,000 meters by deploying various aerial vehicles for flexible and cost-effective aerial networking. However, complex decision-making, resource constraints, and environmental uncertainty pose significant challenges to the development of the LAENet. Reinforcement learning (RL) offers a potential solution in response to these challenges but has limitations in generalization, reward design, and model stability. The emergence of large language models (LLMs) offers new opportunities for RL to mitigate these limitations. In this paper, we first present a tutorial about integrating LLMs into RL by using the capacities of generation, contextual understanding, and structured reasoning of LLMs. We then propose an LLM-enhanced RL framework for the LAENet in terms of serving the LLM as information processor, reward designer, decision-maker, and generator. Moreover, we conduct a case study by using LLMs to design a reward function to improve the learning performance of RL in the LAENet. Finally, we provide a conclusion and discuss future work.

Figures (5)

• Figure 1: An illustration of RL, LLM, and applications of LLM-enhanced RL. The number of peer-reviewed publications regarding RL, LLM, and LLM-enhanced RL per year is shown on the left-hand side (the publication data was collected from IEEE Xplore in April 2025).
• Figure 2: An overview of the LLM’s multiple roles in reinforcement learning, including information processor, reward designer, decision-maker, and generator, highlighting its central role in bridging language input and decision-making processes within the LAENet framework.
• Figure 3: UAV-assisted IoT network with LLM-designed reward funtioon for RL in the LAENet. The UAV agent interacts with the environment by selecting actions based on observed states. The LLM generates the reward function based on structured prompt input of role definition and task description. The generated reward function is evaluated through predefined constraints before being applied to policy learning.
• Figure 4: Energy consumption over episodes of different algorithm with manually designed and LLM-generated reward functions.
• Figure 5: Effect of packet size on energy consumption using different reward design methods.