Foundation models for EEG aim to learn once from a large corpus of recordings and then transfer to many different clinical and brain-computer interface tasks with a single backbone. The dominant recipe borrows from language modelling: chop the continuous signal into patches or codebook tokens, then train a transformer with masked self-supervision.

The authors object to the first step. Brain activity is continuous rhythm, and slicing it into discrete tokens fragments those rhythms and blurs the fine-grained temporal detail that carries the information. It is a bit like transcribing music by writing down which note is playing every half second: you keep something, and you lose the phrasing that made it music.

Their alternative uses flow matching with a SplitUNet architecture, staying closer to the continuous nature of the waveform rather than forcing it through a vocabulary. Whether it wins on benchmarks is one question, but the critique itself is worth sitting with, because the tokenisation habit was inherited from text rather than chosen for brains.