DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models
DEFT proposes a plug-and-play adapter for EEG foundation models to improve cross-subject classification. It disentangles mixed pretrained tokens into a task-oriented content branch and a subject-consistent style branch. By performing downstream prediction only on the content branch, it prevents subject variation from burdening the classifier, leading to better robustness in cross-subject decoding.