Disentanglement

ECG attention concentrated in fewer slots, suggesting compact morphological cues dominate. EEG attention was broader and multi-slot, consistent with heterogeneous cortical dynamics. The synthetic perturbation experiment indicated that tokens specialized in different local motifs without explicit supervision. Adding the TCR penalty is interpreted as making tokens less redundant and more specialized. The diversity objective is reported to encourage independent local anomalies to be isolated rather than distributed everywhere.