Multimodal Large Language Model-Aided Environment-Aware Channel Prediction and Beamforming

The acquisition of accurate channel state information (CSI) in future 6G with ultra-massive multiple-input-multiple-output (MIMO) systems is crucial for designing optimal transmission strategies such as beamforming, but it is significantly challenged by prohibitive pilot overhead and highly dynamic scattering environments. While multimodal sensing provides rich environmental information, existing small-scale artificial intelligence (AI) models lack the robustness and generalization required for practical deployment. To address these limitations, this paper proposes a novel cloud-edge hierarchical intelligence framework based on multimodal large language models (MLLMs) that integrates semantic environmental understanding with real-time decision-making. The cloud-based MLLM performs structured reasoning on scatterer attributes and channel states through specialized modules for multi-modal scatterer fusion and channel relationship modeling, conducting Chain-of-Thought (CoT) reasoning and generating high-level strategic policies. At the edge, a dynamic scattering cluster tracking method is proposed in complex propagation environments and a lightweight Reinforcement Learning (RL) agent enables robust beam prediction, while a progressive self-correction mechanism ensures continuous adaptation through performance monitoring and model fine-tuning. Simulation results show that by leveraging semantic environmental understanding, the proposed framework achieves a substantially lower CSI prediction NMSE compared to traditional and unimodal MLLM-based methods, while achieving higher spectral efficiency.