LLM-Empowered MAPPO for Embodied Cognitive Satellite–Terrestrial Networks With RSMA

Abstract

This paper introduces an embodied agentic AI framework that integrates large language models (LLMs) with multi-agent reinforcement learning (MADRL) to enable adaptive control in cognitive satellite-terrestrial networks (CSTNs). The framework embeds LLM-based cognitive modules into network entities, transforming them into autonomous agents capable of semantic perception, reasoning, and collaborative decision-making. To address key CSTN challenges such as dynamic interference, complex resource allocation, and heterogeneous quality-of-service (QoS) demands, we employ LLMs to interpret high-level operational intents, augmented by retrieval-augmented generation (RAG) for accessing domain knowledge. This enables each agent to adaptively configure rate-splitting multiple access (RSMA)-based protocols, derive key performance metrics (e.g., outage probability, age of information), and formulate a constrained long-term energy efficiency optimization problem. To solve this problem, we propose an LLM-enhanced multi-agent proximal policy optimization (LEMAPPO) algorithm for joint power and rate allocation. The LLM enhances MAPPO through action guidance and reward function design, thereby improving learning efficiency and policy robustness. Simulations demonstrate that the proposed algorithm achieves substantial energy efficiency gains while satisfying reliability and timeliness constraints, outperforming existing benchmarks. Specifically, it outperforms standard MAPPO by up to 28.5% in energy efficiency under stringent outage constraints, and achieves 27.3% higher efficiency than MAPPO in multi-user scenarios.