ReaLM: Real-Time Channel Prediction with Distilled LLM

In the realm of next-generation wireless communications, real-time and accurate channel state information (CSI) prediction remains a major challenge due to the complex, dynamic, and noise-prone nature of wireless environments. To tackle these difficulties, we propose ReaLM, an efficient framework that adapts and compresses large language models for channel prediction tasks. We fine-tune the Llama model with a time-frequency data fusion feature extraction module, enabling more effective spatiotemporal sequence modeling of CSI. After training, we design a knowledge distillation strategy to reduce the high computational latency and heavy parameter load of large language models, compressing the model by 62% and accelerating inference by 58%. During training, we apply data augmentation and noise injection to improve the model's antiinterference capability. Extensive experiments demonstrate that ReaLM outperforms traditional baselines in low-SNR and crossenvironment scenarios, establishing itself as a robust and efficient solution for 6G real-time channel prediction.