Proceedings of the Annual Meeting of the Cognitive Science Society

EEG-based auditory attention decoding (AAD) seeks to identify which speaker a listener attends to in multi-talker "cocktail party" settings. A central challenge is cross-subject generalization: neural responses vary substantially across individuals, inducing distribution shifts across subjects; consequently, models trained in subject-dependent or mixed-subject regimes may latch onto subject-specific cues that hinder transfer and interpretation. To address this, we propose a band-gated multi-band framework that decomposes EEG into low- and high-frequency pathways and adaptively fuses them at the sample level to learn attention-discriminative representations while accommodating inter-individual spectral variability. We further introduce an identity-disentangled objective that leverages confidencefiltered pseudo-labels to perform alignment in an auxiliary bottleneck space, encouraging a more subject-invariant bottleneck representation while mitigating subject-specific variability. Evaluated on KUL, DTU, and AVED under leave-one-subject-out protocols and two decision-window settings, our approach achieves the best or highly competitive performance against strong baselines. Analyses of the learned gate and representation geometry provide qualitative support for the roles of adaptive band reweighting and identity suppression in improving robustness in cross-subject AAD. Code is available at https://github.com/siyingtao/BDGI_for_AAD.