LatentWave: JEPA Pretraining for Wireless Foundation Models

Researchers introduce LatentWave, a wireless foundation model that uses Joint-Embedding Predictive Architecture (JEPA) instead of traditional masked input reconstruction to learn more transferable representations from wireless spectrograms and channel state information. The model demonstrates improved performance across RF signal classification, 5G positioning, beam prediction, and LoS/NLoS classification tasks while supporting variable antenna configurations.

LatentWave represents a meaningful advancement in wireless signal processing by shifting the pretraining paradigm from reconstructive to predictive objectives. Traditional masked-input approaches like WavesFM optimize for pixel-level reconstruction accuracy, which can trap learned representations in low-level signal artifacts rather than capturing the semantic patterns necessary for downstream applications. By operating in latent space and predicting masked regions contextually, JEPA-based pretraining encourages the model to learn more generalizable wireless phenomena.

The innovation addresses a fundamental challenge in wireless AI: building single models that transfer effectively across heterogeneous tasks and hardware configurations. Prior foundation model approaches required task-specific fine-tuning or struggled with variable antenna arrays common in real deployments. LatentWave's per-channel patch embeddings with stochastic sampling during pretraining elegantly solve this by learning antenna-agnostic representations that scale to different hardware setups.

The empirical findings reveal that masking geometry itself functions as a task-dependent inductive bias. Frequency masking improves channel-related tasks by preserving spectral coherence patterns, while region masking maintains signal discriminability for classification. This discovery suggests that pretraining design choices encode implicit task priors, enabling practitioners to optimize foundation models for specific application domains.

The research has moderate industry implications. Wireless signal processing remains computationally expensive and fragmented across specialized models. More efficient, transferable foundation models could accelerate development of 5G/6G applications, spectrum monitoring systems, and radio anomaly detection. However, practical deployment requires validation on production datasets and integration with existing telecom infrastructure.

• →JEPA-based pretraining outperforms masked reconstruction for learning transferable wireless representations in latent space
• →Variable antenna support through stochastic channel sampling enables deployment across heterogeneous wireless hardware without retraining
• →Masking geometry introduces task-dependent inductive biases: frequency masking aids channel tasks while region masking preserves signal classification performance
• →LatentWave shows improvements across four diverse wireless tasks including 5G positioning and beam prediction compared to masked-modeling baselines
• →Foundation model approach reduces need for separate specialized models across RF signal classification, positioning, and channel estimation applications