26 December 2025 in IEEE Communications Surveys & Tutorials
Foundation models (FMs) are recognized as a transformative breakthrough that has started to reshape the future of artificial intelligence (AI) across both academia and industry. The integration of FMs into wireless networks is expected to enable the development of general-purpose AI agents capable of handling diverse network management requests and highly complex wireless-related tasks involving multi-modal data. Inspired by these ideas, this work discusses the utilization of FMs, especially multi-modal FMs in wireless networks. We focus on two important types of tasks in wireless network management: prediction tasks and control tasks. In particular, we first discuss FMs-enabled multi-modal contextual information understanding in wireless networks. Then, we explain how FMs can be applied to prediction and control tasks, respectively. Following this, we introduce the development of wireless-specific FMs from two perspectives: available datasets for development and the methodologies used. Finally, we conclude with a discussion of the challenges and future directions for FM-enhanced wireless networks.
7 April 2026 in IEEE Transactions on Communications
Artificial intelligence (AI), and more specifically deep learning techniques, have demonstrated strong capabilities in processing wireless signals, enabling automatic modulation classification (AMC). However, existing AI-based AMC methods often produce unreliable predictions and lack robustness to out-of-distribution (OOD) inputs, which limits their deployment in real-world scenarios. This study aims to address the gap in simulations to real deployments by fitting predictions to OOD scenarios. We propose an uncertainty-aware AMC framework based on supervised contrastive learning (SupCon). The proposed framework aims to enhance classification reliability, particularly under OOD conditions. In this framework, a ResNet-based representation learning model that consists of a feature extractor and a projection head is first trained using a combination of SupCon loss and cross-entropy (CE) loss to produce class-discriminative embeddings. By decoupling the shared representation learning model from task-specific classifiers, the framework enables a modular and computationally efficient uncertainty estimation strategy, avoiding redundant computation during inference. The learned embeddings are then processed by a set of class-wise binary classifiers that provide both classification and sample-wise uncertainty estimates. A rejection mechanism is incorporated to improve decision reliability by discarding uncertain predictions. We evaluate the proposed framework on the RadioML 2018 dataset. Experimental results show that our approach significantly improves representation quality and classification reliability compared to conventional supervised learning. Specifically, the classification accuracy increases from 63.7% to 93.6% under in-distribution conditions, and from 30.8% to 92.5% under 50% OOD contamination, while maintaining over 85% recall on accepted predictions.