METIS

A brain-signal foundation model from Peking University that aligns neural recordings with language so one model serves many analyses without task-specific retraining. METIS is pretrained on what the paper calls the largest and most diverse brain-signal corpus assembled, about 70,000 hours across 20 datasets, through a unified language–signal alignment framework, and reports zero-shot gains of about 20.9% over prior approaches. It addresses the two failures the authors see in current practice: end-to-end models that need retraining per task, and pretrained encoders that still depend on heavy fine-tuning.