Do Personal Adapters Carry Over Across Sessions?

Do Personal Adapters Carry Over Across Sessions?

Xilin Tao and Kani Chen report that personal adapters—small additions fitted to a frozen EEG foundation model—do not carry over equally across the three models they tested. A motor imagery brain-computer interface requires calibration for each new user, who may return on another day; the authors argue that one-session evaluations miss whether an adapter’s gains persist and whether those gains come from the individual rather than the population model. Across 54 two-session subjects, first-day adapters improved second-day performance over the population model for REVE and LaBraM, retaining 61% and 83% of same-day recalibration gains. For CBraMod, the second-day gain was indistinguishable from zero. Same-day recalibration had the highest mean accuracy for all three models, and adding 5 to 20 second-day labels mostly did not match it. The authors conclude that carryover depends on the model and should be tested separately for each model, with personalization compared against both a population reference and adapters fitted to other users. The abstract reports model-specific results, but does not explain what accounts for the difference in carryover.

Publication: 51f4e5a8-1679-47be-8ca6-948d3c4a8460

attribution — pass: The draft correctly attributes the paper and its conclusions to authors Xilin Tao and Kani Chen. Results and recommendations are presented as the authors' reports, not as established fact.

Authors:Xilin Tao, Kani Chen

summary — uncertain: The cross-session findings are faithful, but the reading leaves out the paper's other headline result: in 235 held-out subjects, same-session calibration gave a positive net gain in all nine model and dataset combinations, and this held when population training was quadrupled. Without it, readers may wrongly infer that personal adapters help only sometimes. Repair: add one sentence reporting this result. Central terms EEG, motor imagery and brain-computer interface also go unexplained; they need term-library tooltips (suggest additions if missing) so non-specialists can follow.

same-session calibration produced a positive net gain in all nine model and dataset combinations (means of 1.5 to 5.4 percentage points), and the gain remained positive when population training was quadrupled

date — pass: originalPublished is 2026-09-28, which matches the v1 submission date. It is not the 30 Oct revision date or the November feed date.

Submitted on 28 Sep 2026 (v1)

support — pass: Each claim is directly supported by the abstract, and each quoted excerpt is accurate. The limitations correctly note that the abstract gives no mechanism and covers only three models. The question follows from that unexplained gap and is not a leading claim. The draft implies no unsupported benefits or costs.

whereas the gain for CBraMod was 0.04 points and indistinguishable from zero

media — pass: The selected media is the verified direct link to the arXiv abstract page for this paper, which is the most direct access to the source.