BCIwiki (bciwiki.com) — Adapters trained for each new user raised the mean balanced accuracy of EEG foundation models by 1.5-5.4 percentage points over a shared population model, researchers Xilin Tao and Kani Chen reported in a preprint posted to arXiv on September 28, 2026. The work has not been peer reviewed.
Brain-computer interfaces (BCIs) must be calibrated for every new user, but an apparent personalization gain can simply reflect a stronger population model. To separate the 2 effects, the researchers froze 3 EEG foundation models, CBraMod, REVE and LaBraM, and tested them on 235 held-out subjects from 3 motor-imagery datasets. Each subject's own adapter was compared with the population model and with adapters fitted to other subjects.
Using all labels from the first half of each session, personal adapters beat the population model by 1.5-5.4 points and outperformed exchanged adapters by 2.3-7.3 points in all 9 model-dataset combinations.
The size of the benefit depended on how much the population model was trained. With 4 times the original training budget, median gains stayed positive at 1.0-2.0 points but were smaller for every model, and no population model reached a confirmed plateau.
Getting the benefit cheaply was unreliable. Few-label calibration was consistently non-negative on only 1 dataset, and in CBraMod neither unlabeled context nor meta-learned initialization beat matched controls. The authors recommend judging personalization against both a population reference and exchanged parameters, across population-training budgets.