LWM-Temporal: Sparse Spatio-Temporal Attention for Wireless Channel Representation Learning

A team from Arizona State University and Meta has introduced LWM-Temporal, a new member of the Large Wireless Models (LWM) family designed specifically for spatiotemporal wireless channel representation. The model operates in the angle-delay-time domain and introduces a novel Sparse Spatio-Temporal Attention (SSTA) mechanism that restricts interactions to physically plausible neighborhoods, reducing attention complexity by an order of magnitude while preserving geometry-consistent dependencies. This architecture allows the model to efficiently capture how wireless signals evolve as users and objects move through space.

LWM-Temporal is pretrained using a self-supervised, physics-informed masking curriculum that emulates realistic occlusions, pilot sparsity, and measurement impairments commonly found in real wireless environments. Experimental results demonstrate consistent improvements over strong baselines in channel prediction tasks across multiple mobility regimes, particularly under long prediction horizons and when fine-tuning data is limited. The model's ability to learn transferable spatiotemporal representations highlights the importance of geometry-aware architectures for wireless AI applications, potentially enabling more efficient 5G/6G network optimization and beamforming.