CARP: Consistency-Aware Robust Pretraining for Multi-Task Wireless Foundation Model

Towards building a wireless foundation model, we propose CARP, a consistency-aware robust pretraining model that learns universal and transferable channel state information representations. By integrating unsupervised contrastive learning with targeted variance regularization, CARP simultaneously promotes semantic aggregation and structural details, thus adapting to a broad spectrum of downstream tasks. This design improves performance in two key aspects. First, CARP introduces complementary inductive biases during pretraining to accommodate the heterogeneous representation requirements of downstream tasks. Second, CARP actively filters unstructured noise to enhance physical robustness, ensuring stable performance in practical environments. Furthermore, we provide an information-theoretic interpretation, that this integration is not empirical but optimizing the trade-off between structural invariance and information redundancy, from the Information Bottleneck perspective. Experimentally, numerical results show that CARP significantly improves both classification and regression downstream performance while enhancing resilience under noisy conditions, demonstrating the substantial potential of leveraging channel state consistency for next-generation integrated sensing and communication tasks.