Broader cross-scenario generalization

Characterize and improve the cross-scenario generalization of the BeamGuard multimodal forecasting-and-control pipeline beyond the DeepSense 6G Scenarios 32 and 33 day--night evaluation, including substantially different road geometries and scenario distributions.

Background

The main experiments train and evaluate BeamGuard primarily on DeepSense 6G Scenarios 32 and 33, which represent paired daytime and nighttime conditions. Additional tests on Scenarios 31 and 34 reveal a substantial degradation in ranked beam-prediction accuracy and communication-level performance under broader scenario shift.

The held-out results indicate that the framework does not yet establish universal generalization across different road settings and scenario distributions. The authors therefore explicitly identify broader scenario generalization as unresolved.