Foundation Model for Intelligent Wireless Communications

Abstract: The evolution toward intelligent next-generation wireless systems promises unprecedented spectral efficiency and reliability but is hindered by a paradigm of narrow and data-hungry AI models. Breaking from this constraint, this work introduces WiFo-2, a revolutionary wireless foundation model that establishes a new state of the art for extensive channel state information (CSI)-based tasks. Uniquely architected as a sparse mixture of experts, WiFo-2 effectively manages heterogeneous data and tasks while enabling highly efficient inference. It is pretrained on a massive and diverse dataset of 11.6 billion CSI points, which enables the acquisition of profound and generalizable channel knowledge. WiFo-2 demonstrates remarkable zero-shot capabilities, not only matching but surpassing the full-shot performance of task-specific baselines on unseen configurations, all while providing reliable confidence estimates. Furthermore, the model achieves exceptional performance on eight key downstream tasks with minimal fine-tuning. A functional hardware prototype demonstrates its real-world deployment feasibility and significant system gains, highlighting WiFo-2's superiority and paving the way for a paradigm shift in AI-based wireless systems.