LBP: A Generative AI Framework for Lightweight Multimodal Beam Prediction in 6G Edge Networks
Abstract
Beam prediction under integrated sensing and communication is a fundamental building block for next-generation communication networks. However, achieving reliable beam prediction in practical systems remains challenging due to the conflicting requirements of high accuracy, ultra-low latency, and strict model complexity constraints. Existing methods often address these challenges only partially, as parameter-heavy architectures limit inference efficiency and robustness in dynamic environments. To address these challenges, we propose a generative artificial intelligence framework for lightweight multimodal beam prediction (LBP) in 6G edge networks. At its core, LBP introduces a hierarchical attention mechanism that jointly captures vehicle motion dynamics and global traffic context, enabling accurate beam prediction with substantially reduced model complexity. By fusing vehicle trajectory information with traffic scene perception and integrating a scene-aware module, LBP achieves robust performance across diverse weather and traffic conditions. Extensive experimental results show that LBP outperforms existing baselines with significantly fewer parameters and millisecond-level inference latency, enabling real-time 6G edge deployment.
