LLM-MM: End-to-End Robust Multimodal Beam Prediction for 6G V2X Networks via MoE-LoRA Adaptation

In 6G vehicle-to-everything (V2X) scenary, precise millimeter-wave beam prediction under dynamic environmental disturbances faces critical challenges of latency-accuracy trade-offs and cross-scenario generalization. It not only requires accurate and prompt determination of the vehicle’s position, but also needs to overcome external environmental disturbances. However, most existing schemes suffer from critical limitations such as poor adaptability in extreme scenarios, prolonged decision latency, and restricted cross-scenario generalization capability. Towards this end, we propose LLM-MM: an end-to-end robust multimodal beam prediction framework for 6G V2X networks via MoE-LoRA Adaptation. Specifically, we first construct a distinctive beam prediction architecture to leverage the powerful inference capabilities of the Large Language Model (LLM), thereby shortening the inference time while ensuring the inference accuracy. Secondly, by integrating the Mixture-of-Experts with Low-Rank Adaptation model, LLM-MM not only ensures its generalization ability across multiple scenarios but also reduces the training cost. Our framework integrates a rigorously evaluated open-source LLM, selected through systematic comparison of multiple candidates to achieve optimal performance. Extensive numerical results demonstrate the advantages of proposed framework from multiple perspectives.