Official repository for paper From General LLM to Specialized Beamforming for Low-Altitude ISAC Networks: A LoRA-Based Multi-Expert Framework.
Integrated sensing and communication (ISAC) achieves both radar sensing and data transmission, yet the design of ISAC beamforming often leads to inherently non-convex optimization problems. Recently, large language models (LLMs), as representatives of general artificial intelligence (AI), have exhibited remarkable capabilities in mathematical reasoning and problem-solving. Empowering ISAC with LLM has thus attracted considerable attention as a promising direction. However, most existing studies simply treat a pretrained LLM as a black-box solver without further adaptation, failing to adapt a general-purpose model into a task-oriented expert. In this paper, we pioneer the use of LLMs to construct an end-to-end ISAC optimizer. We design a multi-expert framework that enables adaptation to diverse ISAC scenarios. Leveraging the strength of LLM in natural language (NL) understanding, we reformulate communication scenarios into NL descriptions, thereby eliminating the need for preprocessing modules and explicit mathematical derivations. Furthermore, inspired by the mixture-of-experts paradigm, we employ low-rank adaptation to specialize the LLM for ISAC optimization, enabling it to generalize across diverse scenarios. We develop a two-stage training framework. In the supervised fine-tuning stage, the model learns to generate solutions with a structured output format. Then, reinforcement learning further refines the outputs to ensure constraint feasibility and numerical optimality. Extensive experiments demonstrate that our approach achieves superior performance, significantly outperforming existing AI-based methods.
We will open our coda after our paper is accepted.