The nagging problem in EEG decoding is that no two brains look alike, so a model trained on other people often flops on you, and practical systems still recalibrate per person from scratch. EEG foundation models were supposed to help, but even big pretrained ones need real adaptation before they work on a new task; you can't just use them as frozen feature extractors.

This paper splits adaptation into two channels. A global adapter learns what's common across everyone, while subject-specific adapters soak up each person's individual quirks, and stacked LoRA combines both. Across three motor-imagery datasets and several backbones, the stacked version won most comparisons. The nuance worth keeping: the balance shifts with your population. A shared adapter suffices for a big, diverse crowd, while individual adapters earn their keep when subjects vary more.

From the abstract, so read the paper for per-backbone results and how the split was tuned.