Model adaptation · new people; next day held

New people, next day: which part of a pretrained model should you update?

Which part of a pretrained encoder to update for people it has not seen, on a task it is adapted to: the head only, the last block, or a low-rank adapter, each from the same checkpoint and five-epoch recipe. The next-day version has run; its source is held.

Measured on: EEGMAT · Methods: LaBraM

Short answer

On new people, same task, zero labels from the test person — LaBraM Base on mental arithmetic versus rest, 36 people in five participant-disjoint folds — updating the last block with the head reached 65.7% balanced accuracy and rank-4 LoRA 64.1%, against 56.6% for a short head-only fit; the paired interval for LoRA minus last block crosses zero. On the same people and folds, the core matrix's frozen LaBraM with a ridge head scored 64.6%. The next-day experiment ran, but its source is held, so its figures are unpublished.

New people, same task, zero labels from the test person

Adapting a foundation model: head only, last block or LoRA?

LaBraM Base on EEGMAT, mental arithmetic versus rest: 36 people in five participant-disjoint folds. The three update rules start from the same head, see the same batches in the same order and train for the same five epochs; three seeds are averaged within each person before the cohort mean.

Head only · encoder frozen

56.6%
56.6% balanced accuracy

Macro F1 54.6% · 402 trainable parameters

95% interval 54.2%–59.0%

Last block + head

65.7%
65.7% balanced accuracy

Macro F1 62.3% · 482,882 trainable parameters

95% interval 61.6%–69.7%

LoRA rank 4 + head

64.1%
64.1% balanced accuracy

Macro F1 60.5% · 38,802 trainable parameters

95% interval 60.2%–67.8%

For scale: the core matrix's frozen LaBraM, on the same people and folds but read out by a closed-form ridge head on standardized features, scored 64.6% (core matrix release, experiments.json). The head-only arm here is a short five-epoch head on unscaled features — not the best a frozen encoder can do — and the two are not a paired comparison.

+7.5 ppLoRA minus head only, same people

LoRA beat the short head-only fit; it did not beat the last block

30 of 36 people improved and 6 got worse; paired interval +4.8 to +10.1 pp. Updating the last block instead gained +9.1 pp (+6.0 to +12.2 pp). LoRA minus last block was −1.6 pp, interval −4.4 to +1.0 pp: it crosses zero, so neither is shown to be better. 17 people did better with LoRA, 19 with the last block.

Change in participant-mean score for the same 36 people under both rules · descriptive person-bootstrap 95% interval · people helped, harmed and unchanged.
ComparisonBalanced accuracyMacro F1
LoRA − head only+7.5 pp+4.8 to +10.1 pp30 helped · 6 harmed · 0 unchanged+5.9 pp+2.7 to +9.0 pp22 helped · 14 harmed · 0 unchanged
Last block − head only+9.1 pp+6.0 to +12.2 pp30 helped · 5 harmed · 1 unchanged+7.7 pp+4.4 to +11.2 pp27 helped · 9 harmed · 0 unchanged
LoRA − last block−1.6 pp−4.4 to +1.0 pp17 helped · 19 harmed · 0 unchanged−1.9 pp−5.0 to +1.1 pp15 helped · 21 harmed · 0 unchanged

On macro F1 the gains are smaller, and LoRA helps fewer people over head only than it does on balanced accuracy.

What each rule costs

Training time is the sum over 15 fits (5 folds × 3 seeds) on one Apple-silicon Mac, PyTorch MPS backend, excluding loading and evaluation.
Update ruleTrainable parametersTraining time, 15 fitsBalanced accuracy, seed by seed
Head only · encoder frozen40251.5 s56.9% · 56.4% · 56.5%
Last block + head482,88265.3 s65.7% · 65.7% · 65.7%
LoRA rank 4 + head38,802128.2 s64.4% · 65.1% · 62.7%

LoRA trains far fewer weights than the last block but took longer here: its adapters sit in all twelve blocks, so gradients flow back through most of the encoder, and this implementation is an effective-weight parametrization, not an optimized low-rank kernel.

Promised on the roadmap and not delivered by this batch: peak memory (only lower-bound samples were recorded, so none is published) and a comparison at a fixed computation budget (every rule trained for the same five epochs, not the same compute).

Nine further foundation encoders were also adapted on these same people and folds with two of the rules above — a head trained on the frozen encoder, and rank-4 LoRA — and the same five-epoch recipe, with LoRA budgets that differ by model: their table is on “Does pretraining help?”.

From the saved test scores of the head-only and LoRA arms above: how good their probabilities are — log loss and calibration error on the same people — and when either should decline a decision. Both are on “When should a decoder decline to decide?”.

Next day · status only

Same person, the next day

Does a model fitted on one day still work on a later one, and which part should be updated when it does not? Two sources were prepared for it; for different reasons, neither can be scored in public.

Two recording days: run and replayed, held

BNCI2015-001: 12 people recorded on two days. Each person's model is trained on day A, calibrated with the first 0, 10, 20 or 40 labeled trials of day B under the same three update rules, and tested on the same later day-B trials. Unlike the result above, this design uses the test person's own labels: their day-A trials to train and, except at the zero budget, their first day-B trials to calibrate. It was run and passed an independent replay. The source's editorial hold stands — its catalogue licence is CC BY-NC-ND and its description names no ethics approval — so no figure from it is published.

Same person, later sessions: status only

Stieger longitudinal BCI: one person's 11 sessions were acquired, a subset of the release. The loader and a chronological split work on it: earlier sessions train, later sessions test. With one person, every score would be that person's, so none is published; the source is listed as status only in the release log.

A related question now has a measured answer, for two classical baselines on other data: the same person’s next session, with and without a few labeled trials from it. It is on How much calibration data does an EEG decoder need?.

Methods & limits

Scope of this result

One untuned recipe

Five epochs, AdamW at learning rate 1e-4, batch 32, with no search, validation split, early stopping or checkpoint selection. Another learning rate or training budget could reorder the rules.

New people, same task

The folds hold out people on a task the model is adapted to. Nothing here measures a new recording day, a new headset or a new dataset.

A known dataset

EEGMAT is public. It is not in LaBraM's published pretraining list, checked on 4 October 2026 (source below): what the authors' list shows, not proof that the recordings were never seen.

Descriptive intervals

Person bootstrap after averaging the three seeds within each person. It ignores the dependence that shared cross-validation models create.

LaBraM’s pretraining list, checked 2026-10-04: source ↗

Adapter size

The LaBraM base encoder has 5,819,936 parameters. Rank-4 updates to its twelve fused QKV weights add 38,400 trainable parameters (0.66%); with the 402-parameter head that is the 38,802 above. The adapter tensors take 150 KiB in float32 before metadata.

Engineering checks passed before any real-EEG run: exact adapter save and reload; exact merged-weight equivalence; finite, nonzero adapter gradients in every block; frozen encoder weights unchanged after updates; identical output at zero adapter initialization.

These counts describe the adapter’s size only.

Related work

Other groups already benchmark EEG adapters: OpenEEGBench covers parameter-efficient fine-tuning of EEG foundation models; a systematic study of test-time adaptation reports inconsistent gains under distribution shift; a recent preprint compares self-supervised adaptation under fixed computation budgets.

EEGMAT · LaBraM adaptation, three matched update rules

Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z.

Open Data Commons Attribution License 1.0 · aggregate results only; the adaptation reuses the 2026-09-20 review of these recordings.

Dataset record ↗

Stieger longitudinal BCI · one-person pilot

James R. Stieger, Stephen A. Engel and Bin He · Continuous sensorimotor rhythm based brain computer interface learning in a large population, Scientific Data 8, 98 (2021); data at doi:10.6084/m9.figshare.13123148.v1.

CC-BY-4.0 · status only.

Dataset record ↗

Data source: adaptation-update.json · schema bci-report-adaptation-update-v1.

Data source: evidence-update.json · the adapter counts and checks, from the released engineering check.

A longer write-up of the EEGMAT adaptation results is on Hugging Face: Choosing how to adapt an EEG foundation model ↗

Cite this page

BCI Report (2026). New people, next day: which part of a pretrained model should you update? https://bci.report/topics/model-adaptation/

Figures from releases adaptation-update-20261001 (2026-10-01), evidence-update-20260922 (2026-09-22), research-preview-20260920 (2026-09-20). Cite the upstream datasets as well: their credits are on this page.

Every release is archived on Zenodo: doi:10.5281/zenodo.23123296. BibTeX for the site and its releases → · CITATION.cff ↗