Open-Set and Cross-Domain Generalization of BATok

Establish the open-set and cross-domain generalization of BATok for instruction-conditioned electromagnetic spectrum understanding beyond the datasets and signal distributions evaluated in the paper.

Background

BATok is evaluated on modulation-recognition, structured-detection, and language-conditioned signal-grounding tasks using the HisarMod2019, EMSpec-Sim, SpaceNet, and simulated multi-signal data represented in EMSpec-Instruct. These experiments demonstrate performance within the evaluated closed-set and dataset-specific settings.

The paper does not verify whether BATok and its associated VLM integrations can recognize previously unseen signal classes or transfer reliably across domains, datasets, acquisition conditions, or signal distributions. Consequently, open-set recognition and cross-domain robustness remain unresolved aspects of the proposed framework.