Foundation Model Guides Dual-Branch EEG Decoding

The arXiv preprint arXiv:2605.00857 (submitted Apr 21, 2026) presents FUSED, a Foundation-guided Source-free EEG Decoding framework by Peiliang Gong et al. The paper targets source-free domain adaptation (SFDA) for cross-subject EEG decoding and reports empirical results across three EEG paradigms: motor imagery, emotion recognition, and SSVEP.

Per the arXiv preprint, FUSED integrates a large-scale EEG Foundation Model (FM) with a compact Specialist Model (SM) using a dual-branch co-adaptation architecture. The authors introduce linear and prototype views on both branches to enable cross-branch pseudo-label generation. They add a Consensus Filtering Mechanism that leverages the FM's inherent stability to select high-quality unlabeled samples and a Two-Stage Pseudo-Label Refinement scheme that applies cross-branch arbitration to suppress error accumulation. Finally, the pipeline performs mutual information maximization to calibrate FM decision boundaries, followed by knowledge distillation from FM to SM-a described calibrate-then-distill sequence. The paper reports that these components collectively yield consistent state-of-the-art performance on the evaluated benchmarks, according to the preprint.

Editorial analysis: In the broader SFDA and BCI literature, foundation models pretrained on large-scale physiological data are increasingly seen as sources of stable priors that can improve generalization when target labels are unavailable. Observers note that using a heavy FM to guide a lightweight specialist mirrors patterns in other modalities where distillation plus selective pseudo-labeling improves privacy-preserving adaptation.

For practitioners: look for the authors' public code and pretrained FM checkpoints, detailed compute and dataset descriptions, and external replication on diverse cohorts. Reproducibility, FM availability, and robustness to recording variability will determine practical adoption.