LLM4CoT: Large Language Models Orchestrating Chain-of-Task Optimization via Meta-Adapter in 6G Heterogeneous Networks

The sixth-generation (6G) wireless communication is evolving from traditional terrestrial homogeneous architectures toward integrated space-air-ground-sea heterogeneous networks (HetNets). In such HetNets, network optimization must jointly address crosslayer multitask coordination spanning the physical, link, and network layers, as optimizing each layer independently often fails to achieve globally optimal performance. However, existing studies have largely focused on multitask optimization within individual layers, leaving crosslayer multitask coordination an open challenge. In this paper, we propose a large language model (LLM)-empowered chain-of-task (CoT) optimization framework, termed LLM4CoT, to tackle the cross-layer multi-task optimization, including link quality estimation, access point (AP) selection, and resource allocation. In addition, we design a unified MetaAdapter structure that captures intertask causal dependencies while mitigating error propagation across tasks. Unlike conventional parallel or task-independent adapter designs, our Meta-Adapter introduces a chain-wise loss and a sequential information flow that conditions later decisions on earlier ones. Moreover, parameter-efficient fine-tuning strategies are tailored to LLM4CoT for efficient training and inference. Experimental results demonstrate that LLM4CoT with Meta-Adapter achieves network sum-rate comparable to conventional joint optimization algorithms, and outperforms state-of-the-art lightweight machine learning baselines in both sum-rate and generalization. These results highlight the potential of LLMs for advancing 6G HetNet optimization.