Large Language Model-Driven Channel Prediction in Cell-Free mMIMO Systems

The channel state information (CSI) acquisition plays a pivotal role in cell-free (CF) massive multiple-input-multi-output (mMIMO) systems. However, conventional pilot-based channel estimation incurs prohibitive overhead costs as user density and mobility increase. To address this, we propose a multi-slot alternating estimation–prediction (MAEP) framework, which leverages temporal correlation to predict future CSI directly and thereby drastically reduce pilot overhead. The efficacy of the proposed framework hinges on prediction accuracy. Inspired by the remarkable modeling capabilities of large language models (LLMs) and their demonstrated efficacy in cross-modal applications, we introduce an LLM-driven channel predictor termed frequency-temporal alignment with LLM (FTAlign-LLM). FTAlign-LLM bridges the modality gap between CSI and the LLM’s feature space through three key components:(i) a multi-scale CSI attention (MSCA) network for extracting rich spatiotemporal features across frequency and delay domains, (ii) a frequency–temporal feature fusion (FTFF) network that fuses these features and aligns them with the LLM’s feature space, and (iii) the utilization of parameter-efficient fine-tuning for LLM adaptation. Extensive results demonstrate that FTAlign-LLM significantly outperforms benchmarks in prediction accuracy. Concurrently, the MAEP framework achieves substantial improvements in sum spectral efficiency, particularly when a large number of access points are deployed in CF mMIMO systems.