Does pretraining help EEG foundation models like LaBraM and CBraMod?

Representation controls · matched frozen encoders

Does pretraining help EEG foundation models like LaBraM and CBraMod?

Sometimes. CBraMod and EEGPT gain under both readout settings on both tasks; LaBraM gains on mental workload and does not on MI/rest. Eleven further foundation models are compared in the same frozen recipe.

Reviewed aggregate JSON ↓Methods & limits ↓

Measured on: ds003810; EEGMAT; BETA; ds006593; TMNRED / ds005383; EESM19 scalp subset · Methods: EEGNet; LaBraM; CBraMod; CCA; REVE; LUNA; BrainOmni; CodeBrain; EEGMamba; ST-EEGFormer; eeg-fm-masking; ERP-FM; SingLEM; ZUNA; EEGPT

Short answer

Sometimes, depending on the encoder and the task: with frozen encoders and the fixed readout, pretrained CBraMod beat the mean of its three random initializations by +6.39 pp of balanced accuracy on MI/rest (10 people) and by +5.90 pp on mental workload (36 people), pretrained EEGPT beat its own by +4.25 pp and +6.45 pp on the same two tasks, and pretrained LaBraM beat its own by +8.12 pp on mental workload, each with a paired 95% interval above zero. On MI/rest, LaBraM's difference was −2.42 pp (interval crosses zero), and −3.56 pp under the train-selected readout (interval wholly below zero). Neither benchmark is in LaBraM's, CBraMod's or EEGPT's published pretraining list (checked 2026-10-04, EEGPT's 2026-10-12). The evidence is for these two tasks and frozen encoders. Eleven further foundation models were run in the same frozen recipe without a random-initialization control. On sleep staging (20 people) REVE Large, ST-EEGFormer Base and ST-EEGFormer Large (76.6%, 77.3%, 79.6%) lie above every published row, spectral ridge at 72.2% included, as does EEGPT, whose interval clears spectral ridge's by only 0.10 pp; on BETA none lies above training-free CCA (63.1%); on EEGMAT the frozen readouts and LoRA order the encoders differently.

LaBraM · MI / restFixed readout · −5.94 to +1.03 pp

−2.42 pp

LaBraM · MI / restTrain-selected readout · −7.00 to −0.19 pp

−3.56 pp

CBraMod · MI / restFixed readout · +3.06 to +9.50 pp

+6.39 pp

CBraMod · MI / restTrain-selected readout · +4.53 to +11.11 pp

+7.89 pp

EEGPT · MI / restFixed readout · +0.75 to +7.44 pp

+4.25 pp

EEGPT · MI / restTrain-selected readout · +0.72 to +7.08 pp

+3.92 pp

LaBraM · Mental workloadFixed readout · +4.97 to +11.40 pp

+8.12 pp

LaBraM · Mental workloadTrain-selected readout · +4.74 to +11.42 pp

+7.99 pp

CBraMod · Mental workloadFixed readout · +2.31 to +9.57 pp

+5.90 pp

CBraMod · Mental workloadTrain-selected readout · +3.12 to +10.43 pp

+6.73 pp

EEGPT · Mental workloadFixed readout · +2.61 to +10.59 pp

+6.45 pp

EEGPT · Mental workloadTrain-selected readout · +4.40 to +10.40 pp

+7.38 pp

Pretrained minus random: MI/rest, 10 people; mental workload, 36 people. Difference in participant-mean balanced accuracy, in percentage points (pp): pretrained encoder minus the mean of three random initializations, same frozen architecture and splits. Dot: the difference; line: paired participant bootstrap 95% interval, printed under each label; dashed line: zero, no difference. Dark rows use the fixed readout, rust rows the train-selected readout.

Paired attribution contrasts

Pretrained minus random, same readout

Each constructor-random architecture is averaged over three initialization seeds. Pretrained and random encoders share the same splits, pooling and ridge readout protocol.

Difference in participant-mean balanced accuracy. Intervals are paired participant bootstraps, conditional on three random initializations.
Task / model	Fixed alpha 100	Train-selected alpha
LaBraMMI / rest · n=10	−2.42 pp−5.94 to +1.03 pp−2.42 pp pretraining contrast	−3.56 pp−7.00 to −0.19 pp−3.56 pp pretraining contrast
CBraModMI / rest · n=10	+6.39 pp+3.06 to +9.50 pp+6.39 pp pretraining contrast	+7.89 pp+4.53 to +11.11 pp+7.89 pp pretraining contrast
EEGPTMI / rest · n=10	+4.25 pp+0.75 to +7.44 pp+4.25 pp pretraining contrast	+3.92 pp+0.72 to +7.08 pp+3.92 pp pretraining contrast
LaBraMMental workload · n=36	+8.12 pp+4.97 to +11.40 pp+8.12 pp pretraining contrast	+7.99 pp+4.74 to +11.42 pp+7.99 pp pretraining contrast
CBraModMental workload · n=36	+5.90 pp+2.31 to +9.57 pp+5.90 pp pretraining contrast	+6.73 pp+3.12 to +10.43 pp+6.73 pp pretraining contrast
EEGPTMental workload · n=36	+6.45 pp+2.61 to +10.59 pp+6.45 pp pretraining contrast	+7.38 pp+4.40 to +10.40 pp+7.38 pp pretraining contrast

Measured scores

Scores under both readout settings

Head selection uses only inner training participants.

Fixed head · alpha 100 · participant-mean balanced accuracy.
Task / model	Pretrained encoder	Random encoder mean	Paired difference
LaBraMMI / rest	53.5%	55.9%3 initializations	−2.42 pp−5.94 to +1.03 pp
CBraModMI / rest	62.9%	56.5%3 initializations	+6.39 pp+3.06 to +9.50 pp
EEGPTMI / rest	58.3%	54.1%3 initializations	+4.25 pp+0.75 to +7.44 pp
LaBraMMental workload	64.6%	56.5%3 initializations	+8.12 pp+4.97 to +11.40 pp
CBraModMental workload	62.3%	56.4%3 initializations	+5.90 pp+2.31 to +9.57 pp
EEGPTMental workload	65.7%	59.2%3 initializations	+6.45 pp+2.61 to +10.59 pp

Train-selected head · alpha 100 / 10,000 / 1 · participant-mean balanced accuracy.
Task / model	Pretrained encoder	Random encoder mean	Paired difference
LaBraMMI / rest	52.9%	56.5%3 initializations	−3.56 pp−7.00 to −0.19 pp
CBraModMI / rest	62.7%	54.9%3 initializations	+7.89 pp+4.53 to +11.11 pp
EEGPTMI / rest	57.0%	53.1%3 initializations	+3.92 pp+0.72 to +7.08 pp
LaBraMMental workload	64.4%	56.5%3 initializations	+7.99 pp+4.74 to +11.42 pp
CBraModMental workload	68.4%	61.7%3 initializations	+6.73 pp+3.12 to +10.43 pp
EEGPTMental workload	66.0%	58.6%3 initializations	+7.38 pp+4.40 to +10.40 pp

EEGPT’s rows: its pretrained rows under the fixed head are its core-matrix cells on these two protocols, the same predictions, so the notes on those cells hold here. Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. ds003810 is recorded at 125 Hz; its windows reach the encoder upsampled to 256 Hz, so they carry nothing above 62.5 Hz.

Fixed classical control

52.3%

Log-covariance ridge · MI / rest

95% interval 48.5%–56.4%

Fixed classical control

61.9%

Log-covariance ridge · Mental workload

95% interval 56.2%–67.6%

Training-seed sensitivity

EEGNet across three training seeds

Five existing EEGNet protocols were rerun with three fixed MPS seeds. The range is a small-sample check of training sensitivity, not a population interval; the published point estimates stand.

Balanced accuracy across three training seeds. Sample SD and full range are percentage points.
Protocol	Three runs	Mean	Sample SD	Full range
Semantic ERPn=30 · 3,600 trials	60.50% / 60.42% / 60.33%	60.42%	0.08 pp	0.17 pp60.33%–60.50%
P300n=21 · 2,520 trials	55.79% / 50.36% / 52.06%	52.74%	2.78 pp	5.44 pp50.36%–55.79%
Scalp sleep stagingn=20 · 3,000 trials	57.33% / 54.43% / 51.60%	54.46%	2.87 pp	5.73 pp51.60%–57.33%
BETA · 4 selected channelsn=70 · 11,200 trials	44.77% / 44.42% / 45.03%	44.74%	0.30 pp	0.61 pp44.42%–45.03%
BETA · 8 selected channelsn=70 · 11,200 trials	56.54% / 56.73% / 56.99%	56.75%	0.23 pp	0.46 pp56.54%–56.99%

Dataset authors, versions, licenses and source records for all four seed-sensitivity datasets are listed under Data use & privacy → Core and deployment sources.

Frozen-probe comparison

Further foundation encoders, frozen

Eleven further EEG foundation models — sixteen encoder checkpoints — were swapped into the published LaBraM and CBraMod frozen recipe, with the same windows, people, folds, ridge head and scoring, and no random-initialization control.

EEGPT sits in the sleep and BETA tables too, from its own release: it ran in the same runs as the published LaBraM and CBraMod rows with only the encoder swapped, and its random-initialization control is in the sections above.

A row is above or below another only where their 95% intervals do not overlap; with 126 frozen cells from the further encoders and EEGPT’s eight more, an occasional non-overlap is expected by chance.

Sleep staging: four new encoders above every published row

On the core matrix’s sleep protocol (EESM19 scalp subset, 20 people, six channels, chance 20.0%), REVE Large, ST-EEGFormer Base and ST-EEGFormer Large lie entirely above every published row — spectral ridge, the published non-foundation row with the highest point estimate (72.2%, 70.7%–73.8%), the frozen LaBraM and CBraMod rows, and EEGNet. So does EEGPT, by a narrow margin: the lower end of its interval, 73.9%, clears the upper end of spectral ridge’s by 0.10 pp. Sleep is the only one of the seven scored protocols where a new row clears the published non-foundation row with the highest point estimate. The other new rows overlap the published ones or lie below them.

Published rows on sleep staging (core matrix): Spectral ridge 72.2% (70.7%–73.8%) · LaBraM 71.1% (69.0%–73.2%) · CBraMod 71.3% (68.7%–73.7%) · EEGNet 56.5% (54.9%–58.3%)

Sleep staging · frozen encoder and ridge head · balanced accuracy with its descriptive 95% interval.
Encoder	Balanced accuracy	Against LaBraM, CBraMod and spectral ridge
REVE Base	74.6%72.9%–76.4%	Overlaps all three
REVE Large	76.6%75.2%–78.0%	Above all three
LUNA Base	67.5%65.3%–69.5%	Overlaps LaBraM and CBraMod; below spectral ridge
LUNA Large	72.4%70.4%–74.0%	Overlaps all three
BrainOmni Base	67.7%65.8%–69.5%	Overlaps LaBraM and CBraMod; below spectral ridge
CodeBrain	70.7%68.8%–72.7%	Overlaps all three
EEGMamba	64.1%61.6%–66.4%	Below all three
ST-EEGFormer Base	77.3%75.6%–78.8%	Above all three
ST-EEGFormer Large	79.6%78.2%–81.0%	Above all three
eeg-fm-masking MAE (r=9 cm, L=2)	67.9%65.4%–70.4%	Overlaps LaBraM and CBraMod; below spectral ridge
ERP-FM Base	71.0%69.5%–72.6%	Overlaps all three
SingLEM	35.3%33.7%–37.0%	Below all three
ZUNA 1.1Model card: research use only, not for diagnosis or clinical use.	60.1%57.8%–62.3%Exposure unknown	Below all three
EEGPTSame runs as the published rows	75.5%73.9%–77.1%	Above LaBraM (intervals 0.67 pp apart), CBraMod (intervals 0.17 pp apart) and spectral ridge (intervals 0.10 pp apart)

Against the published rows: LaBraM, CBraMod and spectral ridge, the published non-foundation row with the highest point estimate, by whether the 95% intervals overlap. Unless a row is marked, its dataset is not in the model authors’ published pretraining list (checked 2026-10-04); each model’s sources are on its method page. EEGPT’s row gives the gap between the two intervals beside each relation; its authors’ list was checked on 2026-10-12.

BETA: above CBraMod, overlapping or below training-free CCA

On BETA’s 40-target SSVEP with eight electrodes (70 people, chance 2.5%), REVE Large and eeg-fm-masking MAE (r=9 cm, L=2) lie above the published frozen CBraMod row (33.7%), as do the three masking-ablation siblings shown below. No new row lies above standard CCA (63.1%), which needs no training at all. ST-EEGFormer’s BETA cells are in its authors’ pretraining list, and they still lie below CBraMod.

Published rows on BETA at eight electrodes (core matrix): Standard CCA 63.1% (57.2%–69.0%) · Spectral ridge 50.7% (45.6%–55.8%) · LaBraM 10.8% (9.5%–12.2%) · CBraMod 33.7% (29.8%–37.6%) · EEGNet 55.8% (50.1%–61.3%)

SSVEP, BETA, eight electrodes · frozen encoder and ridge head · balanced accuracy with its descriptive 95% interval.
Encoder	Balanced accuracy	Against LaBraM, CBraMod and standard CCA
REVE Base	37.5%33.3%–41.9%	Above LaBraM; overlaps CBraMod; below standard CCA
REVE Large	52.7%47.7%–57.8%	Above LaBraM and CBraMod; overlaps standard CCA
LUNA Base	8.6%7.7%–9.5%	Below all three
LUNA Large	10.0%8.9%–11.1%	Overlaps LaBraM; below CBraMod and standard CCA
BrainOmni Base	21.7%18.6%–25.0%	Above LaBraM; below CBraMod and standard CCA
CodeBrain	11.9%10.5%–13.3%	Overlaps LaBraM; below CBraMod and standard CCA
EEGMamba	40.2%35.7%–44.7%	Above LaBraM; overlaps CBraMod; below standard CCA
ST-EEGFormer Base	17.1%15.5%–18.8%in the authors' published pretraining list	Above LaBraM; below CBraMod and standard CCA
ST-EEGFormer Large	22.6%20.1%–25.3%in the authors' published pretraining list	Above LaBraM; below CBraMod and standard CCA
eeg-fm-masking MAE (r=9 cm, L=2)	54.8%49.4%–60.0%	Above LaBraM and CBraMod; overlaps standard CCA
ERP-FM Base	18.4%16.3%–20.5%	Above LaBraM; below CBraMod and standard CCA
SingLEM	26.6%23.7%–29.5%	Above LaBraM; below CBraMod and standard CCA
ZUNA 1.1Model card: research use only, not for diagnosis or clinical use.	11.9%10.8%–13.0%Exposure unknown	Overlaps LaBraM; below CBraMod and standard CCA
EEGPTSame runs as the published rows	20.3%18.2%–22.6%	Above LaBraM (intervals 5.97 pp apart); below CBraMod (intervals 7.24 pp apart) and standard CCA (intervals 34.64 pp apart)

Against the published rows: LaBraM, CBraMod and standard CCA, the published non-foundation row with the highest point estimate (its interval overlaps EEGNet’s), by whether the 95% intervals overlap. Unless a row is marked, its dataset is not in the model authors’ published pretraining list (checked 2026-10-04); each model’s sources are on its method page. EEGPT’s row gives the gap between the two intervals beside each relation; its authors’ list was checked on 2026-10-12.

EEGPT on BETA: PO5 and PO6 use channel embeddings that EEGPT’s pretraining never trained: they are two of the four names in its 62-name channel dictionary (AF7, AF8, PO5, PO6) that its 58-electrode pretraining montage leaves out. The other six channels here, and all four of the 4-channel subset, use trained embeddings. For context: BETA is not in EEGPT’s published pretraining list, but the Tsinghua 40-target SSVEP benchmark (Wang et al., 2017) is, as TSU. BETA’s authors describe BETA as an extension of that benchmark, from the same group with the same 40-target layout and its own recordings; whether anyone took part in both is not stated.

EEGMAT: frozen readouts and LoRA disagree

Nine of the new encoders were also adapted on EEGMAT, mental arithmetic against rest (36 people, 5 participant-disjoint folds, three seeds, chance 50.0%), with the recipe of the published LaBraM adaptation: a head trained on the frozen encoder against rank-4 LoRA with the same head, five epochs, no tuning. LoRA’s budget differs by architecture, from 32,768 to 428,032 added trainable parameters.

The ridge readout of the frozen-probe table, the trained head and LoRA do not put the encoders in the same order. LoRA’s change excludes zero for REVE Base, REVE Large, LUNA Base, SingLEM and ZUNA 1.1; it includes zero for BrainOmni Base, CodeBrain, EEGMamba and ST-EEGFormer Base.

The clearest case is SingLEM: both frozen readouts reach chance (51.5% with the ridge head, 50.9% with the trained head), while LoRA reached 60.6%.

EEGMAT · balanced accuracy with its descriptive 95% interval · the paired change over the same people and folds, with how many people moved which way · trainable parameters of the head alone and of the LoRA arm.
Encoder	Frozen, ridge head (frozen-probe table)	Frozen, trained head	LoRA rank 4 + head	LoRA minus trained head	Trainable parameters · head · LoRA arm
LaBraMLaBraM, already published	64.6%60.5%–68.6%	56.6%54.2%–59.0%	64.1%60.2%–67.8%	+7.5 pp+4.8 to +10.1 pp30 improved · 6 worse · 0 unchangedThe interval excludes zero.	402 · 38,802
REVE Base	68.8%64.1%–73.2%	64.0%60.4%–67.7%	66.3%61.9%–70.7%	+2.3 pp+0.1 to +4.6 pp21 improved · 13 worse · 2 unchangedThe interval excludes zero.	1,026 · 181,250
REVE Large	69.0%63.3%–74.7%	65.6%61.1%–70.1%	69.2%63.4%–74.9%	+3.6 pp+0.5 to +6.8 pp21 improved · 13 worse · 2 unchangedThe interval excludes zero.	2,434 · 430,466
LUNA Base	60.7%56.9%–64.4%	58.2%54.5%–61.7%	60.0%55.9%–63.9%	+1.8 pp+0.1 to +3.3 pp23 improved · 11 worse · 2 unchangedThe interval excludes zero.	514 · 37,378
BrainOmni Base	61.6%58.4%–64.9%	64.5%59.6%–69.2%	64.8%59.6%–69.7%	+0.3 pp−1.4 to +1.8 pp18 improved · 14 worse · 4 unchangedThe interval includes zero: no change is established.	16,386 · 106,498
CodeBrain	63.0%59.5%–66.5%	66.5%62.3%–70.6%	67.1%62.6%–71.6%	+0.6 pp−0.7 to +1.9 pp18 improved · 14 worse · 4 unchangedThe interval includes zero: no change is established.	7,602 · 58,802
EEGMamba	60.3%57.0%–63.4%	60.4%55.7%–65.2%	60.8%56.3%–65.3%	+0.4 pp−0.7 to +1.5 pp18 improved · 14 worse · 4 unchangedThe interval includes zero: no change is established.	7,602 · 90,930
ST-EEGFormer Base	64.5%59.0%–70.0%	63.2%59.7%–66.7%	65.1%61.0%–69.4%	+1.9 pp−0.2 to +4.1 pp20 improved · 16 worse · 0 unchangedThe interval includes zero: no change is established.	1,538 · 148,994
SingLEM	51.5%49.2%–53.8%Interval reaches chance level	50.9%49.5%–52.3%Interval reaches chance level	60.6%56.7%–64.4%	+9.7 pp+6.1 to +13.3 pp30 improved · 6 worse · 0 unchangedThe interval excludes zero.	1,218 · 33,986
ZUNA 1.1Exposure unknownModel card: research use only, not for diagnosis or clinical use.	53.3%49.5%–57.1%Interval reaches chance level	51.2%49.8%–52.7%Interval reaches chance level	56.9%53.7%–60.1%	+5.7 pp+2.5 to +8.7 pp28 improved · 7 worse · 1 unchangedThe interval excludes zero.	1,218 · 296,130

The others were not adapted: LUNA Large (not run by design: LUNA Large is frozen probes only in this evaluation’s specification); ST-EEGFormer Large (not run by design: ST-EEGFormer Large is frozen probes only); eeg-fm-masking MAE (r=9 cm, L=2), eeg-fm-masking JEPA (r=9 cm, L=2), eeg-fm-masking MAE (r=one channel, L=1) and eeg-fm-masking JEPA (r=one channel, L=1) (not run by design: eeg-fm-masking is a frozen-probe masking ablation); ERP-FM Base (not run by design: ERP-FM is frozen probes only and EEGMAT arithmetic is outside the ERP design).

REVE Base and Large: a size comparison

REVE Large’s point estimate is at or above REVE Base’s on all seven scored protocols; their intervals do not overlap on BETA at eight electrodes, BETA at four electrodes, the P300 target ERP and the semantic target ERP. Both checkpoints were scored on the same people, but each interval is marginal and no paired test was run, so a gap is a description, not an established effect of size.

REVE Base and Large on every scored protocol: the table

Frozen encoder and ridge head · balanced accuracy with its descriptive 95% interval · Large minus Base as the export states it, and whether the two intervals overlap.
Protocol	REVE Base	REVE Large	Large minus Base	Intervals
Motor imagery & rest10 people	66.2%61.8%–70.6%	70.2%64.0%–76.1%	+4.0 pp	Overlap
SSVEP · 8 channels70 people	37.5%33.3%–41.9%	52.7%47.7%–57.8%	+15.2 pp	Do not overlap
SSVEP · 4 channels70 people	25.5%22.4%–28.8%	46.4%41.7%–51.2%	+20.8 pp	Do not overlap
Arithmetic & rest36 people	68.8%64.1%–73.2%	69.0%63.3%–74.7%	+0.2 pp	Overlap
P300 target ERP21 people	50.5%48.5%–52.3%Interval reaches chance level	54.6%52.7%–56.5%	+4.1 pp	Do not overlap
Semantic target ERP30 people	53.7%51.8%–55.4%	59.7%57.5%–61.8%	+6.0 pp	Do not overlap
Sleep staging20 people	74.6%72.9%–76.4%	76.6%75.2%–78.0%	+2.0 pp	Overlap

The other two size pairs do not show this pattern: LUNA’s Base and Large intervals separate only on sleep staging, ST-EEGFormer’s only on BETA at eight electrodes and BETA at four electrodes, where its cells are in the authors’ pretraining list.

Masking ablation: not separable on six of seven protocols

eeg-fm-masking released four checkpoints of one study: masking framework (MAE or JEPA) crossed with mask geometry (the paper-recommended r = 9 cm with L = 2, or a single channel with L = 1), with the same architecture, corpus and recipe. The paper-recommended MAE checkpoint is a matrix row in the tables above; the other three are its ablation. On six of the seven scored protocols every pair of the four overlaps. On sleep staging one pair separates — JEPA at r = 9 cm against MAE on one channel — and it changes the framework and the geometry at once, so it isolates neither.

The four eeg-fm-masking checkpoints on every scored protocol: the table

Frozen encoder and ridge head · balanced accuracy with its descriptive 95% interval.
Protocol	MAE · r = 9 cm, L = 2 (matrix row)	JEPA · r = 9 cm, L = 2	MAE · one channel, L = 1	JEPA · one channel, L = 1	Pairs whose intervals separate
Motor imagery & rest	68.1%64.3%–72.3%	63.2%58.2%–67.9%	64.9%60.3%–69.4%	63.3%59.0%–67.4%	None: every pair overlaps
SSVEP · 8 channels	54.8%49.4%–60.0%	49.6%44.5%–54.8%	53.4%48.1%–58.6%	51.0%45.8%–56.1%	None: every pair overlaps
SSVEP · 4 channels	47.1%42.1%–52.1%	43.4%38.7%–48.3%	46.7%41.9%–51.5%	45.4%40.6%–50.2%	None: every pair overlaps
Arithmetic & rest	63.6%59.1%–68.1%	61.6%56.4%–66.5%	64.3%60.2%–68.2%	62.5%58.5%–66.2%	None: every pair overlaps
P300 target ERP	49.5%48.1%–51.0%At or below chance level	51.6%50.4%–52.7%	51.1%49.2%–53.0%Interval reaches chance level	50.0%48.7%–51.5%Interval reaches chance level	None: every pair overlaps
Semantic target ERP	54.8%52.7%–56.9%	55.8%53.9%–57.6%	54.9%53.4%–56.4%	55.0%53.4%–56.6%	None: every pair overlaps
Sleep staging	67.9%65.4%–70.4%	70.3%68.0%–72.7%	65.2%63.0%–67.3%	68.3%65.7%–70.8%	JEPA at r = 9 cm vs. MAE on one channel

Limits of the further encoders

One fixed recipe, not a ceiling

The frozen probe is the published linear head with only the encoder swapped; the adaptation is one untuned five-epoch recipe on one task.

Small cohorts, descriptive intervals

Motor imagery has 10 people, P300 21, sleep staging 20. The intervals are participant bootstraps that ignore cross-validation dependence and carry no multiplicity correction.

Family order

Rows keep the export’s family order; the same people under two encoders are not tested as a pair.

Method comparisons

Balanced or class-matched designs (P300, semantic target, motor imagery against rest, sampled sleep epochs) compare methods; they are not detection, false-alarm or latency estimates for real use.

New people, same task and setup

The folds hold out people. Nothing here is cross-day, cross-device or cross-dataset evidence, and nothing is a clinical claim.

Short windows

Every encoder sees the published 1–2 s windows (sleep as 15 two-second segments), usually shorter than its pretraining context; each model’s input handling is in its row footnote on the protocol pages.

Pretraining exposure: the authors’ lists

BETA is in ST-EEGFormer’s published pretraining list and TMNRED in SingLEM’s; ZUNA 1.1 publishes none, so every ZUNA 1.1 cell is unknown; EEGMamba’s list is read from its official code, with medium confidence. Every other cell, LaBraM’s and CBraMod’s included, is not in the authors’ published pretraining list (checked 2026-10-04), and neither are EEGPT’s (checked 2026-10-12) — not proof that the recordings were never seen.

No endorsement

No model’s authors endorse these results. No adapted weights or LoRA deltas were kept or shared, and each model’s weights licence travels with its rows. Timings were measured on a shared GPU and are not published.

Weights licences and terms

REVE Base · REVE Large: REVE Responsible Use License v1.0 (owner-approved 2026-10-04)REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared.Model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7. paper ↗
LUNA Base · LUNA Large: CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared)CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. paper ↗
BrainOmni Base: MIT paper ↗
CodeBrain: Apache-2.0 paper ↗
EEGMamba: MIT paper ↗
ST-EEGFormer Base · ST-EEGFormer Large: MIT (repo LICENSE; README reserves the paper, diagrams and the name)MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). paper ↗
eeg-fm-masking MAE (r=9 cm, L=2) · eeg-fm-masking JEPA (r=9 cm, L=2) · eeg-fm-masking MAE (r=one channel, L=1) · eeg-fm-masking JEPA (r=one channel, L=1): CC-BY-4.0 (cite arXiv:2609.33487)CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). paper ↗
ERP-FM Base: CC BY-NC-SA 4.0 (non-commercial; aggregate scores only)CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. paper ↗
SingLEM: MIT (checkpoint committed in the MIT repository)MIT. paper ↗
ZUNA 1.1: Apache-2.0; model card: research use only, not for diagnosis or clinical useApache-2.0 weights; model card states research use only, not for diagnosis or clinical use. paper ↗

EEGPT: the checkpoint’s licence is unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only.

Every row, with its footnote and exposure source, is on its protocol page (Motor imagery & rest, SSVEP · 8 channels, SSVEP · 4 channels, Arithmetic & rest, P300 target ERP, Semantic target ERP, Sleep staging); the adaptation arms and the comparisons above are in foundation-models-update.json. Release foundation-models-update-20261004.

EEGPT’s rows are in their own section of the same protocol pages, from eegpt-core-update.json, release eegpt-core-update-20261013.

Methods & limits

Limits of the pretraining contrasts

Fixed versus selected heads

The primary head fixes ridge alpha at 100. The sensitivity setting selects among 100, 10,000 and 1 using inner training participants only.

Three seeds, one cohort

The random baseline averages three encoder initializations. The participant bootstrap does not include unrestricted retraining uncertainty.

Task-specific evidence

MI/rest contains 10 participants and 1,200 trials; mental workload contains 36 participants and 2,160 trials.

Exposure of the two benchmarks

Neither ds003810 nor EEGMAT is in LaBraM’s or CBraMod’s published pretraining list, checked on 4 October 2026, or in EEGPT’s, checked on 12 October 2026 (sources below).

Pretraining lists, checked 2026-10-04: LaBraM ↗ · CBraMod ↗ · EEGPT’s, checked 2026-10-12: EEGPT ↗ · EEGPT repository ↗

Frozen CBraMod with a ridge readout is also read on a later session of the same person on WBCIC-SHU, without a random-initialization control: Does a decoder trained on an earlier session still work later?

Frozen EEGPT with a supervised probe is read against matched classical features in one small pilot, on recorded orthosis device errors, without a random-initialization control: What can a small pilot show, and what can it not?