EEG-PRIME: A New Foundation Model for Cross-Dataset EEG Decoding

EEG-PRIME, a novel two-stage EEG foundation model, has been developed by researchers to tackle the issue of inadequate generalization in EEG decoding across various subjects and datasets. This model integrates masked pretraining with prototype-aligned instruction tuning, facilitating instruction-aware and subject-invariant decoding in different brain-computer interface (BCI) frameworks. In the pretraining stage, an EEG encoder acquires transferable representations via masked reconstruction enhanced by frequency-cutoff spectral augmentation. The instruction tuning phase employs task-semantic, dataset-specific, and subject-invariant conditioning, which adjusts a Q-Former through Layer-wise Query Modulation. Utilizing frozen text embeddings of class labels as prototypes, the model predicts using cosine similarity across diverse label spaces. Evaluations on several datasets revealed enhanced performance in cross-dataset multi-task decoding. The paper can be found on arXiv with the identifier 2608.13072.