Foundation model learns consistent brain signals despite incomplete data

EEG foundation models trained on masked reconstruction struggle when different masked views of the same signal have minimal overlap, failing to learn consistent representations. DARE-EEG solves this through dual-aligned representation learning: mask alignment uses contrastive learning to constrain representations from multiple masked views of the same sample, while anchor alignment stabilizes features against momentum-updated complete signals. The method also introduces conv-linear-probing to adapt pre-trained models to different electrode configurations and sampling rates without retraining. Experiments across EEG benchmarks show state-of-the-art accuracy with superior cross-dataset portability and lower parameter complexity than existing methods. This is aimed at brain-computer interface researchers and practitioners building generalizable EEG models.