Accurate and robust wireless localization is an important enabler for emerging 5G/6G applications, yet the complexity of wireless signals and their sensitivity to environmental changes make cross-scenario localization difficult. Existing data-driven methods often require extensive labeled data and generalize poorly to new environments.

The paper proposes SigMap, a multimodal foundation model for wireless localization. Its cycle-adaptive masking strategy changes the masking pattern according to channel periodicity to learn robust signal representations. A map-as-prompt mechanism then incorporates 3D geographic information through lightweight soft prompts for efficient adaptation across scenarios.

Experiments across multiple localization tasks show state-of-the-art performance and strong zero-shot generalization in unseen environments compared with supervised and self-supervised baselines.