WorksComputer Science

Backbone-Informed Adaptation after EEG Pretraining: A Capacity-Controlled Study

Mingyang Jiang, Yamin Li, Fan Ma, Daniel Moyer, Hua Xu

Published Nov 04, 2026Fresh — added to the Hub within the last 30 daysv136 min read

AI contributions

Not recorded for this paper

View details

Ancillary data

1 code link

View

Abstract

Electroencephalography (EEG) foundation models learn representations that combine information across channels and time. This structure motivates adapters that model the same relationships during downstream learning. However, it is not clear whether such a design outperforms other candidate adapters. Moreover, in order to arrive at a fair comparison between candidate adapters, we also need to consider their trainable parameters. Here, we ask whether a backbone-informed (native) adapter makes better use of a given capacity by comparing it against a token-wise multilayer perceptron control that transforms each feature vector independently. We approximately match their trainable parameter counts and keep the pretrained encoder fixed. Across two EEG foundation models and tasks covering emotion recognition, sleep staging, and EEG event classification, the native adapter shows no consistent advantage. For one backbone, differences between the designs are small relative to variation across runs. For the other, the token-wise control performs better on emotion recognition and EEG event classification. On sleep staging, the native adapter's advantage over the token-wise control is reduced when we also train the component that processes sequences of EEG segments. The relative performance of the two adapters depends on the backbone, task, and downstream training protocol. These findings motivate evaluating backbone-informed adapters against capacity-matched alternatives.

Code availability: https://github.com/0oJimmyo0/EEGxPlore-TMLR

Introduction

Electroencephalography (EEG) records brain electrical activity across channels and time, supporting decoding tasks such as emotion recognition, sleep staging, and EEG event classification. Self-supervised EEG foundation models learn representations from unlabeled recordings for reuse across downstream tasks ([13, 11, 26]). Because pretraining is not tailored to every downstream task, these features may still need task-specific adjustments. Downstream adaptation either updates pretrained weights through full fine-tuning or trains fewer task-specific parameters using parameter-efficient methods ([16, 10]). We focus on adapters that operate after a fixed pretrained encoder. In this setting, the way in which pretrained features are transformed is a central design choice.

Existing EEG adaptation methods have explored different ways to modify pretrained representations, including emphasizing task-relevant features ([10]), modeling sensor relationships ([24]), and using class prototypes ([19]). Another design choice is whether an adapter should explicitly follow the channel and temporal organization of the pretrained encoder. Mixing information across channels and short time segments, or temporal patches, may help adapt these representations to downstream tasks, but the encoder already models relationships across these dimensions. It is therefore unclear whether an adapter that follows this organization is preferable to one that transforms each feature vector independently. Improvements over a frozen-feature classifier alone cannot resolve this question, because adding an adapter also introduces trainable parameters. This motivates comparing the two designs at similar trainable capacity.

To examine this question, we compare two designs that receive the same pretrained features and precede the same classifier architecture (Figure 1). Our backbone-informed design is a native-axis attention adapter that mixes information along the channel and temporal-patch axes of the pretrained representation. The token-wise multilayer perceptron (MLP) control instead transforms each feature vector, or token, independently. We approximately match their trainable adapter parameter counts and keep the pretrained encoder fixed. This comparison evaluates the complete designs, including differences in their operators and token-mixing behavior.

Primary controlled comparison showing a frozen pretrained representation $H \in \mathbb{R}^{B \times C \times S \times D}$ passed to Native adaptation and Token-wise residual control branches, producing $H + \Delta H_N$ and $H + \Delta H_{NM}$, respectively, followed by shared downstream evaluation and the paired effect $\Delta =$ Native $-$ Token-wise. — Figure 1. Primary controlled comparison. A common pretrained representation is passed to an additional task-supervised native-axis attention adapter and a parameter-matched token-wise control. Within each pair, checkpoint, split, seed, classifier architecture and training protocol, and validation rule are held fixed; the paired effect is native minus token-wise.

We conduct our study with two representative EEG foundation models, CBraMod and LaBraM. CBraMod applies channel and temporal attention to separate feature subspaces, while LaBraM produces patch tokens that can be arranged into a channel-by-patch grid ([11, 26]), providing two feature organizations for constructing the adapters. The primary comparison (RQ1) asks whether the native adapter outperforms the capacity-matched token-wise control. RQ2 provides context by asking whether the native adapter improves over the frozen-backbone probe, a classifier trained on the same fixed features without an adapter. RQ3 is a secondary analysis of whether the native adapter adds value when the backbone itself is trainable.

Under the evaluated protocols, CBraMod differences are small relative to seed variation. LaBraM favors the token-wise control on emotion recognition (FACED) and EEG event classification (TUEV). On sleep staging (ISRUC), the native adapter’s balanced-accuracy advantage over the token-wise control shrinks when we train the downstream sequence encoder, which combines information across consecutive EEG segments (§ 6.2). A branch-preserving MLP performs similarly to the token-wise control across the LaBraM comparisons. We also compare full fine-tuning with and without the native adapter across four settings: CBraMod on FACED, TUEV, and SEED-V, and LaBraM on PhysioNet-MI. Adding the adapter yields no consistent benefit. These results motivate comparing backbone-informed adapters with parameter-matched alternatives.

Related Work

EEG representations and feature organization

The structure of a pretrained EEG encoder determines the features and organization available to a downstream adapter. Self-supervised models learn reusable representations through different pretraining approaches ([13, 28, 20, 31]). Electrode-location encoding, retrieval-based spatial learning, and alternating spatial–temporal attention organize spatial and temporal information ([21, 5, 4]), while in frequency-balanced pretraining, the learned representation is informed by the frequency content ([29]). Our study asks how to adapt these features once pretraining is complete. LaBraM’s patch representations and CBraMod’s criss-cross channel–temporal attention ([11, 26]) provide two feature organizations for constructing and comparing post-backbone adapters.

Adaptation designs and structural priors

Parameter-efficient methods offer different ways to modify a pretrained network. Bottleneck adapters add down-projection, nonlinearity, and up-projection modules within the network ([8]), while LoRA learns low-rank weight updates ([9]). Localized LoRA restricts structured updates to selected weight-matrix blocks ([2]). For EEG foundation models, [16] compare full fine-tuning with parameter-efficient alternatives, including LoRA. These methods provide broader tuning references for our comparison, which fixes adapter placement after the backbone.

A separate design choice concerns what guides adaptation. Jeon et al. [10] focus pretrained EEG features on information relevant to downstream tasks. Suzumura et al. [24] use a graph-based adapter to learn spatial relationships among EEG sensors while keeping a pretrained temporal backbone fixed. Ma et al. [19] guide lightweight adaptation with structured class prototypes under limited supervision. These methods use task, sensor, or supervision information to guide adaptation. Our question is whether following the backbone’s channel and temporal-patch axes improves performance over a token-wise alternative with similar adapter capacity.

Token mixing need not use attention: MLP-Mixer uses cross-token MLPs, while the PoolFormer instantiation of MetaFormer uses pooling [25, 30]. Our token-wise control transforms tokens independently. The comparison evaluates complete adapter designs rather than isolating the mixer operator.

Evaluation protocols and controlled comparisons

The relative performance of different adapters needs to be interpreted in the context of the evaluation protocol. Literature on EEG foundation models indicates that models, preprocessing, data splitting, the unit of evaluation, and probing versus fine-tuning are relevant choices [14, 18], and related neuroimaging work shows that preprocessing can affect prediction and representation similarity [17]. Further, evaluation across tasks, datasets, and distribution shifts provides a more complete picture of generalizability [27, 1, 15]. For the adapter comparison, our study holds the checkpoint, split, classifier protocol, and training recipe fixed within each primary pair. We select each training run’s checkpoint using only validation performance, and compare the adapters’ held-out test results from runs with the same random seed.

Adaptation Framework

The native adapter and token-wise control add residual transformations to the post-backbone representation. CBraMod’s native branches use separate feature halves, whereas LaBraM’s branches use the full token embedding. Both adapters preserve the feature-grid shape and feed the same classifier architecture.

Representation format

We denote the post-backbone feature grid by

H \in\mathbb{R}^{B \times C \times S \times D}, \tag*{(1)}

where $B$ is the batch size, $C$ is the channel dimension, $S$ is the temporal-patch dimension, and $D$ is the feature dimension. The adapter preserves this shape. It constructs axis-specific sequences from the grid, applies an interaction mixer, and reshapes the output to the original format.

CBraMod format

CBraMod uses parallel channel and temporal attention Wang et al. [26]. The implementation used here applies channel attention to the first $D/2$ features and temporal attention to the second $D/2$ . Our native adapter follows this branch partition:

H = [H^{\mathrm{ch}} \parallel H^{\mathrm{patch}}], \qquad H^{\mathrm{ch}}, H^{\mathrm{patch}} \in\mathbb{R}^{B \times C \times S \times D/2}. \tag*{(2)}

The channel branch reshapes each temporal patch into a sequence over channels,

H^{\mathrm{ch}} \longrightarrow\mathbb{R}^{(BS) \times C \times D/2}, \tag*{(3)}

Backbone-specific native adapters and token-wise residual control — Figure 2. Backbone-specific native adapters and the token-wise control used in RQ1. CBraMod and LaBraM follow their respective representation formats; the token-wise control applies an independent transformation without channel- or patch-axis mixing. § 4.3 reports adaptation-module parameter counts and matching tolerances.

whereas the patch branch reshapes each channel into a sequence over temporal patches,

H^{\mathrm{patch}} \longrightarrow\mathbb{R}^{(BC)\times S\times D/2}. \tag*{(4)}

The channel-only and patch-only variants enable the corresponding branch; the channel-plus-patch variant enables both branches independently.

LaBraM format

LaBraM produces a sequence of patch tokens that can be reorganized into a channel-by-patch grid [11]. Its native adapter uses the full embedding dimension. The channel and patch branches construct sequences as

H \longrightarrow\mathbb{R}^{(BS)\times C\times D} \qquad\text{or} \qquad H \longrightarrow\mathbb{R}^{(BC)\times S\times D}, \tag*{(5)}

respectively. The primary LaBraM runs use low-rank attention branches; Appendix § 9.§ 2 records the architecture and initialization settings.

Native attention branch

For each axis included in the adapter, the native attention branch extracts the corresponding sequence from $H$ , applies a bottleneck attention block along that sequence, and places the resulting update back into the original feature grid. For an enabled interaction axis $a$ , let $R_a$ denote the associated residual branch. Let $\Phi_a$ denote the backbone-specific extraction operator that selects the relevant feature subspace and forms the sequence for axis $a$ , and let $\Psi_a$ denote the corresponding restoration operator. For CBraMod, $\Phi_a$ selects the corresponding $D/2$ subspace; for LaBraM, it operates on the full $D$ -dimensional embedding. The branch applies layer normalization, projects the features to a bottleneck dimension, performs self-attention over the selected axis, and projects the result back:

X_a = \Phi_a(H), \tag*{(6)}

Z_a = \operatorname{Down}_a(\operatorname{LN}(X_a)), \tag*{(7)}

U_a = \operatorname{MHA}_a(Z_a,Z_a,Z_a), \tag*{(8)}

R_a(H) = \Psi_a(\operatorname{Up}_a(U_a)). \tag*{(9)}

The adapted representation is

H_{\mathrm{adapted}} = H + \sum_{a\in\mathcal{A}} \lambda_a R_a(H), \tag*{(10)}

where $\mathcal{A}$ is the set of enabled axes and $\lambda_a$ denotes the effective coefficient induced by the learned branch scale and configured residual multiplier. For the CBraMod primary configurations, the output projections are zero-initialized, so the residual update is initially zero and $H_{\mathrm{adapted}} = H$ at initialization. The primary LaBraM configuration uses a small learned initial branch scale without zero-initializing the output projection; its initialization settings are reported in Appendix § 9.§ 2. The bottleneck reduces trainable capacity relative to updating the backbone.

Token-wise control

The token-wise control applies an independent bottleneck transformation to each channel–patch token:

R_{\mathrm{tw}}(H) = \operatorname{Up}(\operatorname{GELU}(\operatorname{Down}(\operatorname{LN}(H)))). \tag*{(11)}

We adjust the token-wise bottleneck width to match the native adapter’s trainable parameter count within the tolerance in § 4.3. RQ1 compares these complete attention and MLP designs, so operator choice and axis mixing vary together.

Experimental Protocol

RQ1 tests whether the native adapter outperforms the capacity-matched token-wise control. RQ2 tests its benefit over a frozen-backbone probe, and RQ3 its incremental benefit during full fine-tuning.

Datasets

We conduct our analysis on five public downstream datasets (Table 1): FACED (emotion recognition) ([3]), ISRUC (sleep staging) ([12]), TUEV (EEG event classification) ([22, 7]), PhysioNet-MI (motor imagery) ([6, 23]), and SEED-V (emotion recognition) ((Liu et al., 2022)).

Dataset	Task	$C$	$S$	Classes	Comparisons	Split unit
FACED	Emotion	32	10	9	RQ1, RQ2	Subject/source file
ISRUC	Sleep staging	6	30	5	RQ1, RQ2	Subject
TUEV	Event classification	16	5	6	RQ1, RQ2	Record/prefix
PhysioNet-MI	Motor imagery	64	4	4	RQ2	Subject
SEED-V	Emotion	62	1	5	RQ2 (CBraMod)	Within-subject trial

Table 1. Datasets and frozen-backbone comparisons. RQ1 compares the native adapter with the capacity-matched token-wise control; RQ2 compares it with the frozen-backbone probe. $C$ denotes channels and $S$ temporal patches. TUEV record/prefix splits do not establish subject disjointness.

FACED, ISRUC, and TUEV each provide multiple channels and temporal patches, so the native adapter can mix features along both axes. For each dataset and backbone, we evaluate the native adapter, capacity-matched token-wise control, and frozen-backbone probe across three seeds. These six settings support both the adapter-design comparison (RQ1) and the probe comparison (RQ2).

PhysioNet-MI extends RQ2 to motor imagery for both backbones. It has native adapter and probe runs but no capacity-matched token-wise pair, so it does not support RQ1. SEED-V has only one temporal patch per example ( $S = 1$ ), leaving no cross-patch interaction to model. Its CBraMod channel-only adapter and probe runs test whether adaptation helps when only channel mixing is available. The full-fine-tuning comparisons (RQ3) cover CBraMod on FACED, TUEV, and SEED-V, and LaBraM on PhysioNet-MI.

We follow the CBraMod data-splitting protocol Wang et al., [26] for all downstream datasets. Specifically, FACED, ISRUC, and PhysioNet-MI use subject-disjoint splits; SEED-V uses within-subject trial splits. TUEV uses record/prefix splits with unverified subject disjointness. Appendix § 9.§ 1 gives split counts and preprocessing details. Note that the documentation for the released CBraMod weights does not specify whether TUEV was excluded from the TUEG pretraining data Wang et al., [26]. We therefore do not interpret CBraMod’s TUEV results as evidence of transfer to a dataset unseen during pretraining.

Conditions and evaluation

For the frozen-backbone comparisons, we train the classifier while keeping the pretrained encoder fixed. The probe uses no adapter, whereas the native adapter and token-wise control are trained alongside the classifier. For broader performance context, we include a generic bottleneck adapter and LoRA, which learns low-rank updates to backbone weights. We also evaluate upper-layer tuning, which updates the final backbone layers, and dense fine-tuning, which updates the full backbone (see details in Appendix § 10).

Following the released CBraMod implementation Wang et al., [26], we use flattened-feature MLP heads for FACED/TUEV and a projection and Transformer head for ISRUC. For LaBraM, we follow its mean-pooling and linear-classifier design Jiang et al., [11] and add a sequence encoder before the classifier for ISRUC. Table 8 reports the head architectures and trainable parameter counts.

Backbone	Dataset	Downstream head	Trainable params
CBraMod	FACED	Flatten; 64,000 $\to$ 2,000 $\to$ 200 $\to$ 9 MLP; ELU and dropout 0.1 after the two hidden layers	128,404,009
CBraMod	ISRUC	Per-epoch flatten; 36,000 $\to$ 512 projection and GELU; one Transformer layer ( $d = 512$ , 4 heads, feed-forward width 2,048, dropout 0.1); 512 $\to$ 5 classifier	21,587,461
CBraMod	TUEV	Flatten; 16,000 $\to$ 1,000 $\to$ 200 $\to$ 6 MLP; ELU and dropout 0.1 after the two hidden layers	16,202,406
LaBraM	FACED	Mean patch pooling; frozen LayerNorm; 200 $\to$ 9 classifier	1,809
LaBraM	ISRUC	Per-epoch mean pooling; frozen LayerNorm; newly initialized, frozen Transformer layer ( $d = 200$ , 4 heads, feed-forward width 800); 200 $\to$ 5 classifier	1,005
LaBraM	TUEV	Mean patch pooling; frozen LayerNorm; 200 $\to$ 6 classifier	1,206

Table 8. Downstream heads in the six primary frozen-backbone comparisons. Counts include trainable downstream-head parameters and exclude adapters. Head architecture and trainability are shared by native, token-wise, and probe conditions within each comparison. Linear-layer dimensions are input-to-output.

For ISRUC, both primary protocols freeze the pretrained epoch encoder. CBraMod trains the downstream sequence encoder within its classifier head. The primary LaBraM–ISRUC runs freeze a newly initialized downstream sequence encoder. A separate follow-up trains the sequence encoder and classifier in all four conditions while retaining the frozen epoch encoder (§ 6.2). Appendix § 9.§ 2 gives the primary configurations.

We run each primary experimental condition with three training seeds. For each run, we select the checkpoint with the highest Cohen’s $\kappa$ on the validation set and report its performance on the held-out test set. We use validation balanced accuracy for selector sensitivity. Held-out test balanced accuracy and macro-F1 are primary metrics, with test kappa and weighted-F1 secondary.

Capacity control and paired comparisons

Adding adapters introduces both a feature transformation and additional trainable parameters. Therefore, performance gains over a frozen-backbone probe may reflect both factors and do not by themselves establish whether native-axis attention is preferable to a token-wise transformation. RQ1 compares these designs with closely matched parameter counts, using the same frozen pretrained features and downstream classifier architecture.

We fix the native attention bottleneck dimension at 64 and adjust the token-wise bottleneck width to match its parameter count: 148 for CBraMod and 213 for LaBraM. Matching covers the complete adaptation module, including projections, normalization, and learned residual scales. We exclude the downstream head because its architecture and trainability are identical within each pair. Let $P_{\mathrm{native}}$ and $P_{\mathrm{tw}}$ denote the two adapter parameter counts. We choose widths with a relative mismatch of at most five percent:

\delta_P = \frac{\left|P_{\mathrm{native}} - P_{\mathrm{tw}}\right|}{P_{\mathrm{native}}} \le0.05. \tag*{(12)}

The resulting counts are 59,610 versus 59,949 for CBraMod and 85,810 versus 86,014 for LaBraM, giving mismatches of 0.569% and 0.238%, respectively. These counts apply to each backbone’s FACED, ISRUC, and TUEV comparisons.

Within each comparison, we use the same residual multiplier, adapter dropout, output-initialization policy, optimizer, learning-rate schedule, and classifier protocol (Appendix § 9.§ 2). We use one training recipe per setting without separate tuning for each adapter. We also replace attention with token-wise MLPs within the native branches, approximately matching parameter counts. This tests whether the replacement changes performance relative to the token-wise control (§ 6.1).

For each metric and seed, we subtract condition $B$ ’s test score from condition $A$ ’s, comparing runs with the same backbone, dataset, and training regime. We report the mean and standard deviation of these paired differences across the three seeds, together with the number of seeds for which $A$ outperforms $B$ . These summaries describe the three observed reruns, without population-level significance claims. Aggregates across settings are descriptive and do not treat the settings as independent samples.

Results

Primary matched-adapter comparison (RQ1)

Table 2 reports test balanced accuracy and macro-F1 for the six primary comparisons.

Backbone	Dataset	Probe	Native	Token-wise	RQ1 $\Delta$	RQ2 $\Delta$
CBraMod	FACED		.3777 ± .0023	.3910 ± .0039	.3879 ± .0031	+.0030 ± .0045
	ISRUC		.7562 ± .0033	.7519 ± .0065	.7522 ± .0024	−.0003 ± .0052
	TUEV		.4656 ± .0315	.4548 ± .0223	.4580 ± .0083	−.0032 ± .0194
LaBraM	FACED		.1589 ± .0003	.1613 ± .0097	.1951 ± .0028	−.0338 ± .0116
	ISRUC†		.6515 ± .0021	.7453 ± .0003	.7114 ± .0044	+.0340 ± .0047
	TUEV		.3658‡	.3916 ± .0182	.4307 ± .0010	−.0391 ± .0174
CBraMod	FACED		.3789 ± .0024	.3907 ± .0038	.3874 ± .0015	+.0033 ± .0041
	ISRUC		.7516 ± .0055	.7485 ± .0084	.7484 ± .0021	+.0001 ± .0080
	TUEV		.4813 ± .0192	.4720 ± .0216	.4763 ± .0052	−.0044 ± .0171
LaBraM	FACED		.1499 ± .0003	.1406 ± .0049	.1940 ± .0024	−.0533 ± .0070
	ISRUC		.6354 ± .0019	.7322 ± .0002	.6977 ± .0059	+.0345 ± .0058
	TUEV		.3483‡	.3676 ± .0260	.4329 ± .0010	−.0653 ± .0250

Table 2. Primary held-out test performance and paired effects. Entries are mean ± standard deviation over three seeds. RQ1 is native − token-wise and RQ2 is native − probe; positive values favor native. † LaBraM–ISRUC uses the frozen-sequence protocol. ‡ LaBraM–TUEV RQ2 effects use a common probe reference (Appendix § 10.§ 7).

For RQ1, CBraMod’s mean effects range from −0.004 to +0.003 across the two primary metrics and are smaller than the corresponding across-seed standard deviations. Neither adapter shows a consistent advantage across the three CBraMod settings. For LaBraM, the effects are larger and setting-dependent. The native adapter is weaker on FACED by 0.034 balanced-accuracy points and 0.053 macro-F1 points and on TUEV by 0.039 and 0.065 points, respectively. The primary ISRUC comparison favors the native adapter by 3.40 and 3.45 percentage points, but freezes a newly initialized sequence encoder. Training that encoder reduces these differences to 0.23 and 0.22 points, respectively (§ 6.2). ISRUC therefore shows protocol sensitivity rather than a general native-adapter advantage.

Under the primary frozen-sequence protocol, all three LaBraM seeds favor the native adapter on ISRUC; all three favor the token-wise control on FACED and TUEV (Figure 3).

Figure 3. Seed-level RQ1 effects for the six primary comparisons. Hollow points show the three paired seed effects, diamonds show their means, and horizontal intervals show across-seed standard deviations. Positive values favor the native adapter. LaBraM–ISRUC uses the frozen-sequence protocol; § 6.2 reports the sequence-trained follow-up.

Validation trajectories provide additional context for these selected-checkpoint comparisons (Appendix Figure 6). Appendix § 10.§ 5 examines checkpoint-selection sensitivity separately.

Probe improvement and adapter preference (RQ2)

For RQ2, the native adapter improves both metrics on CBraMod–FACED and LaBraM–ISRUC. LaBraM–TUEV also has positive mean effects, with greater seed variation. The LaBraM–FACED effect is near zero in balanced accuracy and negative in macro-F1; CBraMod–ISRUC and CBraMod–TUEV have negative means. Appendix Table 17 reports additional native-adapter versus probe comparisons on PhysioNet-MI and SEED-V.

Improvement over the probe does not establish which adapter design is preferable. On LaBraM–TUEV, the native adapter has positive mean effects over the probe but negative effects against the token-wise control on both metrics (Figure 4). LaBraM–ISRUC favors the native adapter under both contrasts in the primary frozen-sequence protocol, while CBraMod’s RQ1 effects remain near zero.

Mean RQ2 and RQ1 effects across three paired seeds — Figure 4. Mean RQ2 (native − probe; horizontal) and RQ1 (native − token-wise; vertical) effects in percentage points across three paired seeds. The shaded lower-right region indicates mean improvement over the probe but lower performance than the token-wise control; outlined LaBraM–TUEV points illustrate this for both metrics. Uncertainty is reported in Table 2 and Figure 3. C/L denote CBraMod/LaBraM. † LaBraM–ISRUC freezes a newly initialized sequence encoder; § 6.2 reports the follow-up. LaBraM–TUEV RQ2 uses the reevaluated common probe reference (Table 2).

The absolute scores also matter for interpretation. On LaBraM–FACED, balanced accuracy ranges from .159 to .195 for the probe, native adapter, and token-wise control, while LoRA and Upper-2 reach .282 and .297 on the same split (Appendix Table 11). LoRA and Upper-2 provide broader performance context rather than additional capacity-matched adapter comparisons.

Secondary trainable-backbone comparison (RQ3)

RQ3 compares full fine-tuning with and without the native adapter, updating the pretrained encoder and classifier in both conditions (Appendix Table 15). Adding the adapter improves mean balanced accuracy and macro-F1 on CBraMod–TUEV, CBraMod–SEED-V, and LaBraM–PhysioNet-MI, but reduces both metrics on CBraMod–FACED.

On TUEV, the balanced-accuracy gain is $1.27 \pm3.70$ percentage points and the macro-F1 gain is $2.49 \pm8.58$ points, with standard deviations exceeding the mean gains. Adding the native adapter therefore provides gains in some settings, without a consistent benefit across the four comparisons. Unlike RQ1, these comparisons do not include a token-wise control.

Robustness and Diagnostic Analyses

Targeted follow-up experiments test sensitivity to the branch operator and downstream sequence-encoder training. The FACED/ISRUC branch-preserving MLP and LaBraM–ISRUC sequence-encoder follow-ups used configurations and contrasts fixed before those runs. Validation trajectories, selector comparisons, class-level analyses, and residual-update measurements diagnose existing primary results.

Branch-preserving MLP control

To test the sensitivity of the ranking to the branch operator, we replace native attention with a branch-local multilayer perceptron (MLP). This control preserves branch extraction but transforms tokens independently, with widths adjusted to match parameter counts.

For LaBraM, against the branch-preserving MLP, the native adapter remains weaker on FACED and TUEV and stronger on ISRUC under the primary frozen-sequence protocol, on all three seeds for both metrics (Table 3). The two MLP controls have similar mean performance. FACED and ISRUC provide initialization-matched tests; historical TUEV branch controls use seed-specific adapter initialization. CBraMod–TUEV effects remain small relative to seed variability (Appendix § 10.§ 3).

Table 3: LaBraM branch-preserving MLP comparison at validation- $\kappa$ -selected checkpoints (mean $\pm$ SD, three paired seeds). Native: native adapter; Branch: branch-preserving MLP; Token: token-wise control. BA: balanced accuracy; F1: macro-F1. † TUEV Branch uses seed-specific adapter initialization; all other conditions use adapter seed 12345. ISRUC uses a newly initialized frozen sequence encoder (Table 7).

Dataset	Metric	Native	Branch	Token	Native − Branch	Branch − Token
FACED	BA	$.1613 \pm.0097$	$.1996 \pm.0034$	$.1951 \pm.0028$	$-.0384 \pm.0113$	$+.0046 \pm.0055$
FACED	F1	$.1406 \pm.0049$	$.1947 \pm.0032$	$.1940 \pm.0024$	$-.0540 \pm.0080$	$+.0007 \pm.0027$
ISRUC	BA	$.7453 \pm.0003$	$.7136 \pm.0049$	$.7114 \pm.0044$	$+.0317 \pm.0051$	$+.0023 \pm.0065$
ISRUC	F1	$.7322 \pm.0002$	$.7003 \pm.0055$	$.6977 \pm.0059$	$+.0319 \pm.0057$	$+.0026 \pm.0072$
TUEV†	BA	$.3916 \pm.0182$	$.4311 \pm.0033$	$.4307 \pm.0010$	$-.0395 \pm.0208$	$+.0003 \pm.0043$
TUEV†	F1	$.3676 \pm.0260$	$.4336 \pm.0030$	$.4329 \pm.0010$	$-.0660 \pm.0288$	$+.0007 \pm.0039$

Table 3.

Sequence-encoder training on ISRUC

The primary LaBraM–ISRUC protocol freezes a newly initialized downstream sequence encoder. We repeat all four conditions with this encoder trained jointly with the classifier and any adapter, while keeping the pretrained epoch encoder frozen. The follow-up uses the same three seeds, subject splits, 30-epoch budget, and validation- $\kappa$ selector; details and seed-level results are in Appendix § 10.8.

The sequence-trained protocol improves all four conditions (Table 4). All three adapters outperform the no-adapter baseline on every seed for both metrics. The native adapter has the highest means, but its advantage over the token-wise control is substantially smaller than in the primary frozen-sequence comparison; macro-F1 changes sign on one seed. The native adapter exceeds the branch-preserving MLP on all three seeds, by 0.37 BA and 0.30 F1 points on average. The magnitude and consistency of the ISRUC preference are therefore sensitive to the sequence-encoder training protocol.

Table 4: LaBraM–ISRUC sequence-encoder sensitivity (mean $\pm$ SD, three paired seeds). Both protocols freeze the pretrained epoch encoder and train the classifier and any adapter. The baseline has no adapter; only the follow-up trains the newly initialized sequence encoder. BA: balanced accuracy; F1: macro-F1.

Condition	Frozen sequence encoder BA	Frozen sequence encoder F1	Trainable sequence encoder BA	Trainable sequence encoder F1
No-adapter baseline	$.6515 \pm.0021$	$.6354 \pm.0019$	$.7581 \pm.0016$	$.7502 \pm.0030$
Native adapter	$.7453 \pm.0003$	$.7322 \pm.0002$	$.7713 \pm.0036$	$.7641 \pm.0037$
Token-wise control	$.7114 \pm.0044$	$.6977 \pm.0059$	$.7689 \pm.0065$	$.7619 \pm.0078$
Branch-preserving MLP	$.7136 \pm.0049$	$.7003 \pm.0055$	$.7676 \pm.0041$	$.7611 \pm.0048$

Table 4.

Class-specific TUEV behavior

The LaBraM difference is concentrated in generalized periodic epileptiform discharges (GPED) and eye movement (EYEM), where native F1 is lower by 0.114 and 0.190, respectively; CBraMod effects are comparatively small (Figure 5). Absolute values are in Appendix § 9.§ 3. Macro-F1 gives each class equal weight, so GPED and EYEM contribute substantially to the gap despite their different test-set supports. The class-level results locate the performance differences without identifying their causes.

TUEV class-level F1 differences (native − token-wise) — Figure 5. TUEV class-level F1 differences (native − token-wise; percentage points), averaged over three seeds at validation- $\kappa$ -selected checkpoints. Outlines mark LaBraM’s largest mean deficits. Parentheses give test-set class support; both backbones share a zero-centered scale. Appendix Table 9 reports across-seed variation.

Discussion

Implications for downstream adaptation

The matched comparisons distinguish the value of adding adaptation capacity from the choice of adapter. A native adapter can improve over a frozen-backbone probe while remaining weaker than the token-wise control, as on LaBraM–TUEV. The branch-preserving MLP provides complementary evidence: retaining the branch layout without attention yields similar mean performance to the token-wise control.

The ISRUC follow-up shows that adapter comparisons are sensitive to downstream training conditions. Under the sequence-trained protocol, the native adapter’s advantage becomes much smaller (§ 6.2). The larger frozen-sequence effect therefore provides protocol-specific evidence rather than a general preference for the native adapter.

CBraMod’s small effects remain unresolved rather than establishing equivalence. Both adapters produce measurable residual updates, so their similar performance is not simply due to either remaining inactive (Appendix § 10.9). Pretrained feature interactions and downstream classifier mixing may affect the value of additional adapter-level interactions, but these comparisons do not isolate their contributions, or those of adapter operators and optimization.

Two follow-up comparisons could clarify these unresolved contributions. A parameter-matched axis-mixing MLP would retain cross-token interaction without attention, complementing the branch-preserving control. On ISRUC, varying sequence-encoder trainability while holding dropout behavior and other settings fixed would help distinguish their contributions to the protocol sensitivity.

Limitations

We evaluate two backbones across six primary backbone–dataset settings and three training seeds, with adapters placed after the pretrained encoder. These runs do not cover all sources of initialization uncertainty. Adapter pairs share optimization settings; other insertion points and separate tuning remain untested.

RQ1 compares complete attention-based and token-wise MLP adapter designs, so it does not isolate axis mixing. The native adapter follows each backbone’s axes and feature partitions, but we do not test alternative arrangements or whether CBraMod’s final feature halves retain distinct channel and temporal roles. Adapter implementations, classifier heads, and optimization also differ across backbones. CBraMod’s flattened-feature heads, for example, already mix information across the representation, which may reduce the benefit of further adapter-level mixing. Differences across backbones therefore characterize the evaluated pipelines and cannot be attributed to backbone identity alone.

The primary LaBraM–ISRUC protocol freezes a newly initialized sequence encoder; the follow-up trains it while retaining the frozen pretrained epoch encoder (Appendix § 10.8). This sensitivity analysis changes both sequence-encoder training and its dropout behavior. CBraMod–TUEV pretraining overlap is unresolved. LaBraM–TUEV uses a common probe reference with unverified independent training replication (Appendix § 10.§ 7). Available held-out predictions limit selector-sensitivity analysis (Appendix § 10.§ 5).

Conclusion

In this study, we examined whether downstream adaptation should follow the channel and temporal structure of pretrained EEG features. Across six primary matched comparisons, the native adapter shows no consistent advantage over the token-wise control. The preferred design varies across backbone–task settings and downstream training protocols, with the sleep-staging comparison particularly sensitive to sequence-encoder training. Pretrained feature organization provides a starting point for adapter design. Evaluating backbone-informed adaptation requires comparisons that account for both trainable capacity and downstream training conditions; improvement over a frozen-backbone probe alone does not establish a preference for the structured design.

Primary Per-Seed RQ1 Results

Table 5 reports the held-out test values underlying the primary RQ1 aggregate effects. All entries use the checkpoint selected by validation Cohen’s $\kappa$ . Effects are native minus token-wise and are paired within seed.

Backbone	Dataset	Seed	Native BA	Token-wise BA	$\Delta$ BA	Native F1	Token-wise F1	$\Delta$ F1
CBraMod	FACED	42	.3952	.3873	+.0079	.3951	.3873	+.0078
CBraMod	FACED	1024	.3875	.3853	+.0023	.3879	.3860	+.0019
CBraMod	FACED	3407	.3902	.3913	−.0011	.3891	.3891	+.0001
CBraMod	ISRUC	42	.7541	.7547	−.0006	.7472	.7506	−.0035
CBraMod	ISRUC	1024	.7446	.7500	−.0054	.7408	.7464	−.0056
CBraMod	ISRUC	3407	.7570	.7520	+.0050	.7575	.7482	+.0092
CBraMod	TUEV	42	.4426	.4485	−.0059	.4638	.4719	−.0081
CBraMod	TUEV	1024	.4805	.4631	+.0174	.4964	.4821	+.0143
CBraMod	TUEV	3407	.4413	.4624	−.0211	.4557	.4749	−.0193
LaBraM	TUEV	42	.4127	.4317	−.0190	.3959	.4340	−.0381
LaBraM	TUEV	1024	.3806	.4297	−.0491	.3448	.4322	−.0874
LaBraM	TUEV	3407	.3815	.4308	−.0493	.3622	.4325	−.0703
LaBraM	FACED	42	.1515	.1981	−.0466	.1352	.1965	−.0614
LaBraM	FACED	1024	.1708	.1947	−.0239	.1444	.1936	−.0491
LaBraM	FACED	3407	.1614	.1924	−.0310	.1423	.1918	−.0495
LaBraM	ISRUC	42	.7456	.7085	+.0372	.7323	.6943	+.0381
LaBraM	ISRUC	1024	.7451	.7165	+.0286	.7323	.7045	+.0278
LaBraM	ISRUC	3407	.7453	.7092	+.0361	.7320	.6943	+.0377

Table 5. Per-seed primary RQ1 test results at validation- $\kappa$ -selected checkpoints. BA denotes balanced accuracy and F1 macro-F1; $\Delta$ is native minus token-wise performance in 0–1 score units.

Experimental Details

This appendix gives dataset construction and training configurations for the comparisons in § 4.

We applied the primary comparison criteria retrospectively to existing runs. For the subsequent LaBraM FACED/ISRUC matched comparisons and CBraMod–ISRUC optimization-matched controls, we fixed configurations before training. We also fixed configurations and contrasts before the FACED/ISRUC branch-preserving controls and LaBraM–ISRUC sequence-encoder follow-up.

We use the public CBraMod checkpoint.² The original study reports standard and TUEV-excluded models [26], but the release model card does not identify which variant these weights represent.

Dataset preprocessing and input construction

We follow the CBraMod preprocessing protocol [26]. Within each paired comparison, both adapter conditions use the same preprocessed inputs and splits. Table 6 gives input shapes, split definitions, and example counts.

Dataset	Input shape	Split definition	Train / Val. / Test
FACED	(32, 10, 200)	Subjects 000–079 / 080–099 / 100–122	6,720 / 1,680 / 1,932
ISRUC	(20, 6, 30, 200)	Subjects 1–80 / 81–90 / 91–100	3,559 / 468 / 435
TUEV	(16, 5, 200)	Record/prefix split; original evaluation set used for testing	68,712 / 15,220 / 29,421
SEED-V	(62, 1, 200)	Within-subject trials 0–4 / 5–9 / 10–14	34,432 / 42,960 / 40,352
PhysioNet-MI	(64, 4, 200)	70 / 19 / 20 subjects	6,300 / 1,734 / 1,803

Table 6. Model inputs and dataset splits. Counts denote ten-second segments for FACED, 20-epoch sequences for ISRUC, five-second records for TUEV, one-second windows for SEED-V, and four-second epochs for PhysioNet-MI. ISRUC input shapes include the sequence dimension.

FACED. We use the processed 32-channel recordings for nine-class emotion recognition ([3]). We resample each processed trial to 6,000 samples and divide it into three non-overlapping ten-second segments. Each segment contains ten one-second patches with 200 samples per channel. Subjects are disjoint across the training, validation, and test sets.

ISRUC. We use six bipolar channels for five-class sleep staging ([12]), in the order F3-A2, C3-A2, O1-A2, F4-A1, C4-A1, O2-A1. The preprocessing script takes 200 Hz signals as input, applies a 0.3–35 Hz bandpass filter and a 50 Hz notch filter, and constructs 30-second epochs containing 6,000 samples per channel. We group consecutive epochs into non-overlapping sequences of 20 and discard incomplete epochs and sequences at the end of each recording. Each epoch retains its sleep-stage label.

TUEV. We use six-class EEG event labels ([22, 7]). We resample recordings to 200 Hz, apply a 0.3–75 Hz bandpass filter and a 60 Hz notch filter, and construct 16 differential channels in the order FP1-F7, F7-T3, T3-T5, T5-O1, FP2-F8, F8-T4, T4-T6, T6-O2, followed by FP1-F3, F3-C3, C3-P3, P3-O1, FP2-F4, F4-C4, C4-P4, P4-O2. Each example spans five seconds, including two seconds of context on either side of the annotated event. We divide sorted training-record prefixes into 80% training and 20% validation groups and use the original evaluation set for testing. The input files do not specify a TUEV release version.

SEED-V. We use 62 EEG channels for five-class emotion recognition ((Liu et al., 2022)). We resample recordings to 200 Hz, apply a 0.3–75 Hz bandpass filter, and segment each annotated trial into non-overlapping one-second windows. We split trials within each subject and session into training, validation, and test sets before constructing the windows. Each example contains one temporal patch, giving the $S = 1$ boundary condition.

PhysioNet-MI. We use the four motor-imagery classes from 64-channel recordings ([6, 23]). We interpolate channels marked as bad, apply an average reference, a 0.3 Hz high-pass filter, and a 60 Hz notch filter, then resample to 200 Hz. We extract four-second epochs from the task annotations, exclude rest events, and retain 800 samples per channel. Subjects are disjoint across splits.

Backbone input formatting. We divide amplitudes by 100 for FACED, TUEV, SEED-V, and PhysioNet-MI and leave ISRUC amplitudes unchanged. Each temporal patch contains 200 samples. Both backbones receive the same preprocessed examples and splits. LaBraM maps FACED and PhysioNet-MI channels to their model positions; for TUEV, it uses sequential positional slots 1–16.

Primary experimental configurations

Table Table 7 gives the primary RQ1 configurations, shared across seeds 42, 1024, and 3407. Adapter parameter counts exclude the shared downstream head. The epoch budgets apply to both adapter conditions and the probe; Figure 6 shows validation trajectories over these budgets.

Backbone	Dataset	Condition	$r$	Heads	Adapter LR	Head LR	Batch	Training epochs	Adapter params
CBraMod	FACED	Native	64	4	$5 \times10^{-4}$	$5 \times10^{-4}$	64	50	59,610
CBraMod	FACED	Token-wise	148	–	$5 \times10^{-4}$	$5 \times10^{-4}$	64	50	59,949
CBraMod	ISRUC	Native	64	4	$2 \times10^{-5}$	$3.536 \times10^{-4}$	8	20	59,610
CBraMod	ISRUC	Token-wise	148	–	$2 \times10^{-5}$	$3.536 \times10^{-4}$	8	20	59,949
CBraMod	TUEV	Native	64	4	$5 \times10^{-5}$	$2 \times10^{-4}$	8	20	59,610
CBraMod	TUEV	Token-wise	148	–	$5 \times10^{-5}$	$2 \times10^{-4}$	8	20	59,949
LaBraM	FACED	Native	64	4	$7 \times10^{-4}$	$7 \times10^{-4}$	32	80	85,810
LaBraM	FACED	Token-wise	213	–	$7 \times10^{-4}$	$7 \times10^{-4}$	32	80	86,014
LaBraM	ISRUC†	Native	64	4	$2 \times10^{-4}$	$2 \times10^{-4}$	16	30	85,810
LaBraM	ISRUC†	Token-wise	213	–	$2 \times10^{-4}$	$2 \times10^{-4}$	16	30	86,014
LaBraM	TUEV	Native	64	4	$5 \times10^{-4}$	$5 \times10^{-4}$	64	15	85,810
LaBraM	TUEV	Token-wise	213	–	$5 \times10^{-4}$	$5 \times10^{-4}$	64	15	86,014

Table 7. Primary frozen-backbone configurations, shared across the three seeds. † The newly initialized sequence encoder is frozen in all three conditions.

Figure 6. Validation trajectories for the six primary comparisons. Lines show native minus token-wise validation Cohen’s $\kappa$ for three seeds (gray) and their mean (blue); positive values favor native. Each panel shows its own recorded training budget, matching Table 7. LaBraM–ISRUC uses the primary frozen-sequence protocol; § 6.2 reports sensitivity to sequence-encoder training.

All primary conditions use AdamW with base weight decay 0.05 and label smoothing 0.1. Adapters use dropout 0 and residual multiplier $\gamma= 1$ . We select checkpoints by validation Cohen’s $\kappa$ .

All adaptation modules initialize the residual scale at $\alpha= 0.01$ . LaBraM uses low-rank native attention branches and a full token-wise bottleneck control. Output projections are zero-initialized for CBraMod and not zero-initialized for LaBraM.

Here $r$ denotes the bottleneck dimension. CBraMod uses model/classifier dropout 0.1 and a per-iteration cosine schedule with minimum learning rate $10^{-6}$ . For LaBraM, the model dropout is 0.0, stochastic-depth rates are 0.1 for FACED/ISRUC and 0.2 for TUEV, and the schedules use ten, one, and five warm-up epochs for FACED, ISRUC, and TUEV, respectively. Frozen-backbone evaluation disables stochastic depth.

Both ISRUC sequence Transformers use GELU and pre-layer normalization. The LaBraM sequence encoder contains 482,600 parameters. In the sequence-trained follow-up, it and the classifier contribute 483,605 trainable parameters, with sequence dropout 0.1 active during training.

The CBraMod–ISRUC residual scales and remaining adapter parameters use weight decay 0.05 in both conditions, for all three seeds.

In the primary LaBraM–ISRUC comparison, the newly initialized sequence encoder remains frozen; training updates only the 200-to-5 classifier (1,005 parameters) and any adapter. CBraMod trains its sequence encoder. Appendix § 10.8 reports the LaBraM follow-up with a trainable sequence encoder.

TUEV class-level values

Table Table 9 gives the absolute class-level values underlying Figure 5.

Backbone	Class	$N$	Native R	Token-wise R	$\Delta$ R	Native F1	Token-wise F1	$\Delta$ F1
CBraMod	SPSW	567	.039	.039	+.000 (.000)	.065	.061	+.003 (.003)
CBraMod	GPED	4677	.353	.356	−.003 (.012)	.465	.468	−.003 (.015)
CBraMod	PLED	1998	.310	.323	−.014 (.040)	.369	.385	−.017 (.033)
CBraMod	EYEM	329	.585	.585	+.000 (.053)	.697	.700	−.002 (.046)
CBraMod	ARTF	2204	.542	.539	+.002 (.014)	.364	.368	−.004 (.008)
CBraMod	BCKG	19646	.901	.906	−.005 (.008)	.872	.876	−.004 (.004)
LaBraM	SPSW	567	.000	.000	+.000 (.000)	.000	.000	+.000 (.000)
LaBraM	GPED	4677	.345	.466	−.121 (.143)	.427	.541	−.114 (.110)
LaBraM	PLED	1998	.469	.461	+.008 (.046)	.360	.386	−.026 (.034)
LaBraM	EYEM	329	.077	.185	−.108 (.094)	.118	.308	−.190 (.134)
LaBraM	ARTF	2204	.582	.562	+.020 (.060)	.420	.460	−.040 (.051)
LaBraM	BCKG	19646	.877	.910	−.033 (.006)	.881	.902	−.021 (.007)

Table 9. TUEV class-level results at validation- $\kappa$ -selected checkpoints. $N$ denotes test examples and R recall. Recall and F1 are means over three seeds; $\Delta$ is native minus token-wise, with sample standard deviations in parentheses. Scores and effects are in 0–1 units.

Supporting Results and Robustness Analyses

This appendix contains contextual absolute-performance comparisons, the secondary trainability analysis, checkpoint-selection comparisons, supporting dataset results, and the ISRUC sequence-encoder follow-up.

Contextual adaptation comparisons

Table 10 compares held-out test performance across adaptation methods within each backbone–dataset setting. We select checkpoints by validation Cohen’s $\kappa$ .

(a) Balanced accuracy
Backbone	Dataset	Probe	Native	Token-wise	Generic	LoRA	Upper	Dense
CBraMod	FACED	.378±.002	.391±.004	.388±.003	.390±.003	.417±.007	.403±.005	.572±.011
CBraMod	ISRUC	.756±.003	.752±.006	.752±.002	—	—	—	—
CBraMod	TUEV	.466±.031	.455±.022	.458±.008	.461±.012	.555±.009	.557±.031	.620±.024
LaBraM	TUEV	.365±.008	.392±.018	.431±.001	.437±.013	.572±.023	.541±.020	—
(b) Macro-F1
Backbone	Dataset	Probe	Native	Token-wise	Generic	LoRA	Upper	Dense
CBraMod	FACED	.379±.002	.391±.004	.387±.002	.389±.004	.418±.008	.404±.005	.579±.012
CBraMod	ISRUC	.752±.005	.748±.008	.748±.002	—	—	—	—
CBraMod	TUEV	.481±.019	.472±.022	.476±.005	.474±.007	.527±.006	.518±.020	.565±.037
LaBraM	TUEV	.348±.007	.368±.026	.433±.001	.432±.012	.540±.004	.525±.017	—

Table 10. Contextual held-out test performance for settings with broader reference coverage. Entries are mean ± standard deviation over three seeds at the validation-selected checkpoint; dashes indicate unavailable conditions. Generic denotes a generic bottleneck adapter, Upper denotes upper-layer fine-tuning, and Dense denotes full fine-tuning. ‡ LaBraM–TUEV uses a common probe reference (Appendix § 10.§ 7).

The generic bottleneck adapter uses the same token-wise residual MLP form as the token-wise control: layer normalization, a down-projection, GELU, and an up-projection, with a learned residual scale. Its bottleneck width is fixed at 64 (26,265 trainable adapter parameters), whereas the primary token-wise control uses width 148 for CBraMod or 213 for LaBraM to approximately match the native adapter’s capacity. The Generic columns therefore report a smaller-capacity version of the same architecture.

LaBraM cross-dataset context

Table 11 reports LaBraM–FACED and LaBraM–ISRUC matched runs. Values are validation-kappa-selected test results averaged over the three seeds. The three primary conditions (probe, native, and token-wise) use a frozen backbone and the same dataset-specific training protocol.

Dataset	Condition	Regime	BA	Macro-F1	$n$	Qualification
FACED	Frozen probe	Frozen	.159 $\pm$ .000	.150 $\pm$ .000	3	Probe
FACED	Native C+P, rank 64	Frozen	.161 $\pm$ .010	.141 $\pm$ .005	3	Primary
FACED	Token-wise $b$ = 213	Frozen	.195 $\pm$ .003	.194 $\pm$ .002	3	Primary control
FACED	LoRA	LoRA	.282 $\pm$ .032	.280 $\pm$ .032	3	Context
FACED	Upper-2	Upper layer	.297 $\pm$ .001	.297 $\pm$ .002	3	Context
ISRUC	Frozen probe	Frozen	.652 $\pm$ .002	.635 $\pm$ .002	3	Probe
ISRUC	Native C+P, rank 64	Frozen	.745 $\pm$ .000	.732 $\pm$ .000	3	Primary
ISRUC	Token-wise $b$ = 213	Frozen	.711 $\pm$ .004	.698 $\pm$ .006	3	Primary control
ISRUC	Native C+P, rank 32	Trainable	.792 $\pm$ .006	.785 $\pm$ .008	3	Context
ISRUC	LoRA	LoRA	.692 $\pm$ .004	.677 $\pm$ .006	3	Context
ISRUC	Upper-2	Upper layer	.794 $\pm$ .005	.785 $\pm$ .004	3	Context

Table 11. LaBraM test results on FACED and ISRUC (mean $\pm$ standard deviation over three seeds). BA denotes balanced accuracy, C+P channel-plus-patch adaptation, and $b$ the token-wise bottleneck width. Upper-2 tunes the last two backbone blocks. The ISRUC primary probe, native, and token-wise conditions freeze a newly initialized sequence encoder. Paired effects are reported in the main text.

Table 11: LaBraM test results on FACED and ISRUC (mean $\pm$ standard deviation over three seeds). BA denotes balanced accuracy, C+P channel-plus-patch adaptation, and $b$ the token-wise bottleneck width. Upper-2 tunes the last two backbone blocks. The ISRUC primary probe, native, and token-wise conditions freeze a newly initialized sequence encoder. Paired effects are reported in the main text.

Branch-preserving MLP comparison

The branch-preserving MLP preserves native channel/patch branch extraction and replaces attention with independent token transformations. Its widths are 147 for CBraMod and 106 for LaBraM, compared with native rank 64. The realized parameter mismatches are 0.144% and 0.471%, respectively. Table 3 reports the three-way LaBraM comparison; Table 12 gives the six new FACED/ISRUC seed results.

Dataset	Seed	B BA	A − B BA	B − T BA	B F1	A − B F1	B − T F1
FACED	42	.1990	−.0475	+.0009	.1978	−.0626	+.0013
FACED	1024	.1966	−.0258	+.0019	.1913	−.0469	−.0022
FACED	3407	.2033	−.0419	+.0109	.1949	−.0526	+.0031
ISRUC	42	.7089	+.0367	+.0004	.6942	+.0381	−.0003
ISRUC	1024	.7133	+.0318	−.0032	.7016	+.0306	−.0028
ISRUC	3407	.7187	+.0266	+.0095	.7050	+.0270	+.0107

Table 12. New branch-preserving MLP controls on LaBraM FACED/ISRUC. B denotes Branch, A denotes the native adapter, and T denotes the token-wise control. Each row uses its validation- $\kappa$ -selected checkpoint.

FACED/ISRUC branch controls use the primary learning rates, batches, schedules, 80/30-epoch budgets, adapter seed 12345, and classifier trainability. ISRUC retains the newly initialized frozen sequence encoder. LaBraM–TUEV branch controls use adapter seeds 30042, 31024, and 33407; its native and ordinary token-wise conditions use 12345.

Table 12: New branch-preserving MLP controls on LaBraM FACED/ISRUC. B denotes Branch, A denotes the native adapter, and T denotes the token-wise control. Each row uses its validation- $\kappa$ -selected checkpoint.

Table 13: Branch-preserving MLP comparison on TUEV. Entries are mean $\pm$ standard deviation over three seeds. The effect is native adapter minus branch-preserving MLP; negative values favor the branch-preserving MLP. MLP MLP denotes the branch-preserving MLP control.

Backbone	Native BA	MLP BA	$\Delta$ BA	Native F1	MLP F1	$\Delta$ F1	BA MLP >	F1 MLP ≥
CBraMod	.4548 $\pm$ .0223	.4637 $\pm$ .0168	−.0089 $\pm$ .0315	.4720 $\pm$ .0216	.4781 $\pm$ .0063	−.0061 $\pm$ .0235	2/3	2/3
LaBraM	.3916 $\pm$ .0182	.4311 $\pm$ .0033	−.0395 $\pm$ .0208	.3676 $\pm$ .0260	.4336 $\pm$ .0030	−.0660 $\pm$ .0288	3/3	3/3

Table 13. Branch-preserving MLP comparison on TUEV. Entries are mean $\pm$ standard deviation over three seeds. The effect is native adapter minus branch-preserving MLP; negative values favor the branch-preserving MLP. MLP MLP denotes the branch-preserving MLP control.

On TUEV, the branch-preserving MLP outperforms the native adapter on all three LaBraM seeds and on two of three CBraMod seeds for both metrics.

Secondary trainability analysis

The four full-fine-tuning comparisons train all backbone and classifier weights, with an adapter added only in the native condition. Table 14 gives the optimizer settings.

Table 14: Configurations for the four RQ3 full-fine-tuning pairs, shared across seeds 42, 1024, and 3407. Native rank is 64; C/P denote channel/patch axes. Learning rates are peak values. Both conditions use AdamW, cosine schedules, label smoothing 0.1, and validation- $\kappa$ selection. WD denotes base weight decay.

Backbone	Dataset	Axes	Epochs	Batch	Backbone/upper LR	Head/adapter LR	WD
CBraMod	FACED	C+P	50	64	$10^{-4}/10^{-4}$	$5 \times10^{-4}/5 \times10^{-4}$	.05
CBraMod	TUEV	C+P	20	8	$10^{-4}/2 \times10^{-4}$	$2 \times10^{-4}/2 \times10^{-4}$	.05
CBraMod	SEED-V	C	40	32	$10^{-4}/10^{-4}$	$10^{-4}/10^{-4}$	.03
LaBraM	PhysioNet-MI	C+P	40	64	$5 \times10^{-4}/5 \times10^{-4}$	$5 \times10^{-4}/5 \times10^{-4}$	.05

Table 14.

CBraMod upper-group rates apply to the last two backbone blocks; FACED uses one common backbone rate. CBraMod–TUEV exempts the adapter scalar from weight decay. LaBraM uses five warmup epochs and exempts biases and one-dimensional parameters from weight decay.

Table 15: RQ3 incremental native-adapter effects under frozen and trainable-backbone regimes (mean $\pm$ SD over three seeds). Frozen effects are native minus probe; trainable effects are full fine-tuning with the native adapter minus full fine-tuning without the adapter. Dashes indicate unavailable comparisons. ‡ LaBraM–TUEV frozen effects use the common probe reference (Appendix § 10.§ 7). BA denotes balanced accuracy and F1 macro-F1; effects are in 0–1 score units.

Backbone	Dataset	Frozen		Trainable
		BA	F1	BA	F1
CBraMod	FACED	+0.0133 $\pm$ 0.0061	+0.0118 $\pm$ 0.0061	−0.0064 $\pm$ 0.0083	−0.0073 $\pm$ 0.0091
CBraMod	ISRUC	−0.0043 $\pm$ 0.0050	−0.0032 $\pm$ 0.0060	–	–
CBraMod	TUEV	−0.0108 $\pm$ 0.0511	−0.0093 $\pm$ 0.0406	+0.0127 $\pm$ 0.0370	+0.0249 $\pm$ 0.0858
LaBraM	TUEV‡	+0.0258 $\pm$ 0.0182	+0.0193 $\pm$ 0.0260	–	–
LaBraM	PhysioNet-MI	+0.0139 $\pm$ 0.0057	+0.0054 $\pm$ 0.0085	+0.0159 $\pm$ 0.0074	+0.0164 $\pm$ 0.0064
CBraMod	SEED-V	−0.0069 $\pm$ 0.0036	−0.0074 $\pm$ 0.0042	+0.0077 $\pm$ 0.0068	+0.0085 $\pm$ 0.0068

Table 15.

Checkpoint-selection differences

Table 16 compares checkpoint epochs selected by validation Cohen’s $\kappa$ and balanced accuracy across three seeds.

Table 16: Checkpoint-epoch differences between validation Cohen’s $\kappa$ and balanced-accuracy selection. “Same” counts matching selections among three seeds; median and mean shifts are absolute epoch differences.

Setting	Condition	Same	Median shift	Mean shift
CBraMod–FACED	Native	2/3	0.0	4.0
CBraMod–FACED	Token-wise	2/3	0.0	1.7
CBraMod–TUEV	Native	0/3	3.0	4.0
CBraMod–TUEV	Token-wise	0/3	7.0	6.0
LaBraM–TUEV	Native	0/3	6.0	5.3
LaBraM–TUEV	Token-wise	0/3	3.0	3.0

Table 16.

Supporting and boundary RQ2 comparisons

Table 17: Frozen-probe effects by backbone and dataset for supporting and boundary conditions. Effects are native minus probe (mean $\pm$ SD over three seeds, in 0–1 score units). BA denotes balanced accuracy and F1 macro-F1; BA + and F1 + count seeds with positive effects. Effects for the primary settings are in Table 2.

Backbone	Dataset	BA	F1	BA +	F1 +
CBraMod	PhysioNet-MI	$+0.0014 \pm0.0102$	$+0.0010 \pm0.0087$	1/3	1/3
LaBraM	PhysioNet-MI	$+0.0139 \pm0.0057$	$+0.0054 \pm0.0085$	3/3	2/3
CBraMod	SEED-V	$-0.0069 \pm0.0036$	$-0.0074 \pm0.0042$	0/3	0/3

Table 17.

Per-seed RQ2 effects

Table 18 reports the numerical seed-level values for the primary frozen-probe comparisons. The native adapter and probe are evaluated at their respective validation-kappa-selected checkpoints; the effect is native minus probe.

Backbone	Dataset	Seed	Balanced accuracy			Macro-F1
			Native	Probe	$\Delta$	Native	Probe	$\Delta$
CBraMod	FACED	42	.3952	.3751	+.0201	.3951	.3762	+.0189
CBraMod	FACED	1024	.3875	.3791	+.0085	.3879	.3796	+.0083
CBraMod	FACED	3407	.3902	.3790	+.0113	.3891	.3808	+.0083
CBraMod	ISRUC	42	.7541	.7543	−.0002	.7472	.7466	+.0006
CBraMod	ISRUC	1024	.7446	.7544	−.0098	.7408	.7509	−.0101
CBraMod	ISRUC	3407	.7570	.7600	−.0030	.7575	.7575	+.0000
CBraMod	TUEV	42	.4426	.4986	−.0561	.4638	.4927	−.0289
CBraMod	TUEV	1024	.4805	.4360	+.0445	.4964	.4591	+.0373
CBraMod	TUEV	3407	.4413	.4623	−.0210	.4557	.4921	−.0364
LaBraM	TUEV	42	.4127	.3658	+.0469	.3959	.3483	+.0476
LaBraM	TUEV	1024	.3806	.3658	+.0149	.3448	.3483	−.0035
LaBraM	TUEV	3407	.3815	.3658	+.0157	.3622	.3483	+.0138
LaBraM	FACED	42	.1515	.1587	−.0072	.1352	.1497	−.0146
LaBraM	FACED	1024	.1708	.1593	+.0115	.1444	.1502	−.0058
LaBraM	FACED	3407	.1614	.1587	+.0027	.1423	.1497	−.0074
LaBraM	ISRUC	42	.7456	.6539	+.0917	.7323	.6375	+.0948
LaBraM	ISRUC	1024	.7451	.6502	+.0949	.7323	.6346	+.0977
LaBraM	ISRUC	3407	.7453	.6504	+.0948	.7320	.6341	+.0979

Table 18.

Table 18: Per-seed RQ2 held-out test results for the primary settings. LaBraM–TUEV uses a common probe reference, as explained below. RQ2 effects are native minus probe.

Separate evaluation of three saved LaBraM–TUEV probe checkpoints reproduces identical classification scores. We therefore report a common probe reference; independent training replication remains unverified, and these scores do not establish low seed variability.

LaBraM–ISRUC sequence-encoder sensitivity

We repeat the native, token-wise, branch-preserving MLP, and no-adapter conditions with the newly initialized sequence encoder trained jointly with the classifier and any adapter. The pretrained encoder remains frozen in evaluation mode. We retain the subject splits and inputs in Appendix § 9, the three training seeds, and the 30-epoch budget. Training the sequence encoder also activates its dropout (0.1), so this follow-up tests the sequence-training protocol.

We pair sequence-encoder and classifier initialization within each seed and initialize adapters with seed 12345 independently of the training RNG. Native and token-wise adapter counts match Table 7; the branch-preserving MLP contains 86,214 parameters. Matching excludes the shared sequence encoder and classifier (Table 8).

We retain the LaBraM–ISRUC batch size, learning rate, warmup, and regularization settings in Appendix B.2, applying the same peak learning rate $2 \times10^{-4}$ to every trainable group. AdamW exempts biases and one-dimensional parameters from weight decay; the cosine schedule has minimum learning rate $10^{-6}$ . We select the first checkpoint attaining the maximum validation Cohen’s $\kappa$ and compute test BA and macro-F1 from its confusion matrix. Table 19 and Table 20 report all 12 runs and their paired effects.

Condition	Seed	Epoch	BA	F1
No-adapter baseline	42	25	0.7580	0.7477
No-adapter baseline	1024	23	0.7598	0.7536
No-adapter baseline	3407	23	0.7567	0.7493
Native adapter	42	23	0.7672	0.7598
Native adapter	1024	23	0.7739	0.7660
Native adapter	3407	26	0.7726	0.7665
Token-wise control	42	27	0.7614	0.7529
Token-wise control	1024	23	0.7728	0.7659
Token-wise control	3407	23	0.7725	0.7669
Branch-preserving MLP	42	26	0.7633	0.7558
Branch-preserving MLP	1024	23	0.7714	0.7650
Branch-preserving MLP	3407	23	0.7680	0.7625

Table 19. LaBraM–ISRUC follow-up test results with a trainable sequence encoder. Epoch is the validation- $\kappa$ -selected training epoch (one-based). BA: balanced accuracy; F1: macro-F1.

Contrast	BA	BA +	Macro-F1	F1 +
Native − Token-wise	$+0.234 \pm0.306$	3/3	$+0.218 \pm0.414$	2/3
Native − Branch	$+0.368 \pm0.103$	3/3	$+0.299 \pm0.174$	3/3
Branch − Token-wise	$-0.134 \pm0.320$	1/3	$-0.080 \pm0.366$	1/3
Native − Baseline	$+1.312 \pm0.349$	3/3	$+1.392 \pm0.283$	3/3

Table 20. Paired effects with a trainable LaBraM–ISRUC sequence encoder. Entries are mean $\pm$ sample SD in percentage points; positive counts describe the three observed seeds. Contrasts were fixed before the follow-up runs.

CBraMod residual-update diagnostic

We load the nine primary native-adapter checkpoints and their nine token-wise counterparts, selected by validation Cohen’s $\kappa$ , and measure residual updates in evaluation mode without retraining. Both conditions use the same validation splits and summary procedure. For each validation feature tensor $H$ , we compute $\lVert\Delta H\rVert_F / \lVert H\rVert_F$ , where $\Delta H$ includes the learned branch scales and residual multiplier $\gamma= 1$ . The full validation splits contain 1,680 FACED examples, 9,360 ISRUC epochs from 468 sequences, and 15,220 TUEV examples. Table 21 reports learned scales and per-example ratio summaries.

Dataset	Seed	Condition	$\alpha_C$	$\alpha_S$	$\alpha_T$	Ratio mean	Ratio SD
FACED	42	Native	0.0308	0.0279	–	0.6723	0.0433
FACED	42	Token-wise	–	–	0.0376	0.7350	0.0458
FACED	1024	Native	0.0304	0.0291	–	0.6841	0.0416
FACED	1024	Token-wise	–	–	0.0368	0.7278	0.0440
FACED	3407	Native	0.0307	0.0281	–	0.6680	0.0426
FACED	3407	Token-wise	–	–	0.0390	0.7290	0.0435
ISRUC	42	Native	0.0354	0.0386	–	0.1076	0.0412
ISRUC	42	Token-wise	–	–	0.0369	0.0686	0.0283
ISRUC	1024	Native	0.0389	0.0423	–	0.1431	0.0557
ISRUC	1024	Token-wise	–	–	0.0415	0.0885	0.0365
ISRUC	3407	Native	0.0367	0.0351	–	0.1663	0.0645
ISRUC	3407	Token-wise	–	–	0.0416	0.0990	0.0402
TUEV	42	Native	0.0590	0.0642	–	0.5645	0.0767
TUEV	42	Token-wise	–	–	0.1032	0.6250	0.0797
TUEV	1024	Native	0.0404	0.0425	–	0.5371	0.0793
TUEV	1024	Token-wise	–	–	0.0628	0.5824	0.0831
TUEV	3407	Native	0.0391	0.0417	–	0.5680	0.0784
TUEV	3407	Token-wise	–	–	0.0728	0.6096	0.0840

Table 21. CBraMod residual updates for nine primary native–token-wise pairs at their validation- $\kappa$ -selected checkpoints. Learned residual scales are $\alpha_C$ and $\alpha_S$ for the native channel and temporal-patch branches, and $\alpha_T$ for the token-wise control; dashes indicate inapplicable scales. Ratio mean and SD summarize $\|\Delta H\|_F/\|H\|_F$ across validation examples, with ISRUC measured per sleep epoch. These are feature-update magnitudes, not performance or causal effects.

Both designs produce non-negligible residual updates; their magnitudes alone do not explain the performance differences between adapters.

Validation trajectories

Figure 6 reports validation Cohen’s $\kappa$ trajectories for the six primary comparisons. Appendix § 10.§ 5 compares the checkpoint epochs selected by validation Cohen’s $\kappa$ and balanced accuracy.

References

[1]Hubert Banville, Stéphane d’Ascoli, Simon Dahan, Jérémy Rapin, Marlène Careil, Yohan Benchetrit, Jarod Lévy, Saarang Panchavati, Antoine Ratouchniak, Mingfang Zhang, Elisa Cascardi, Katelyn Begany, Teon Brooks, and Jean-Rémi King. NeuralBench: A unifying framework to benchmark NeuroAI models. arXiv preprint arXiv:2605.08495, 2026. URL https://arxiv.org/abs/2605.08495.
[2]Babak Barazandeh, Subhabrata Majumdar, Om Rajyaguru, and George Michailidis. Localized LoRA: A structured low-rank approximation for efficient fine-tuning. arXiv preprint arXiv:2506.00236, 2025. URL https://arxiv.org/abs/2506.00236.
[3]Jingjing Chen, Xiaobin Wang, Chen Huang, Xin Hu, Xinke Shen, and Dan Zhang. A large finer-grained affective computing EEG dataset. Scientific Data, 10:740, 2023. doi: 10.1038/s41597-023-02650-w.DOI
[4]Alexandru Dimofte, Glenn Anta Bucagu, Thorir Mar Ingolfsson, Xiaying Wang, Andrea Cossettini, Luca Benini, and Yawei Li. CERBro: Compact encoder for representations of brain oscillations using efficient alternating attention. arXiv preprint arXiv:2501.10885, 2025. URL https://arxiv.org/abs/2501.10885.
[5]Yi Ding, Muyun Jiang, Weibang Jiang, Shuoei Zhang, Xinliang Zhou, Chenyu Liu, Shanglin Li, Yong Li, and Cuntai Guan. BrainPro: Towards large-scale brain state-aware EEG representation learning. arXiv preprint arXiv:2509.22050, 2025. URL https://arxiv.org/abs/2509.22050.
[6]Ary L. Goldberger, Luis A. N. Amaral, Leon Glass, Jeffrey M. Hausdorff, Plamen Ch. Ivanov, Roger G. Mark, Joseph E. Mietus, George B. Moody, Chung-Kang Peng, and H. Eugene Stanley. PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation, 101(23):e215–e220, 2000. doi: 10.1161/01.CIR.101.23.e215.DOI
[7]Amir Harati, Mohammad Golmohammadi, Silvia Lopez, Iyad Obeid, and Joseph Picone. Improved EEG event classification using differential energy. In 2015 IEEE Signal Processing in Medicine and Biology Symposium, pp. 1–4, 2015. doi: 10.1109/SPMB.2015.7405421.DOI
[8]Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attarian, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, pp. 2790–2799, 2019. URL https://proceedings.mlr.press/v97/houlsby19a.html.arxiv.org/abs/1902.00751
[9]Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. URL https://arxiv.org/abs/2106.09685.
[10]Jaehyun Jeon, Seungwoo Jeong, Yeajin Shon, and Heung-Il Suk. Parameter-efficient transfer learning for EEG foundation models via task-relevant feature focusing. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management, pp. 1068–1077. Association for Computing Machinery, 2025. doi: 10.1145/3746252.3761372.DOI
[11]Wei-Bang Jiang, Li-Ming Zhao, and Bao-Liang Lu. Large brain model for learning generic representations with tremendous EEG data in BCI. In The Twelfth International Conference on Learning Representations, 2024. URL https://proceedings.iclr.cc/paper_files/paper/2024/file/47393e8594c82ce8fd83adc672cf9872-Paper-Conference.pdf.arxiv.org/abs/2405.18765
[12]Sirvan Khalighi, Teresa Sousa, Jose Moutinho Santos, and Urbano Nunes. ISRUC-Sleep: A comprehensive public dataset for sleep researchers. Computer Methods and Programs in Biomedicine, 124:180–192, 2016. doi: 10.1016/j.cmpb.2015.10.013.DOI
[13]Demetres Kostas, Stephane Aroca-Ouellette, and Frank Rudzicz. BENDR: Using transformers and a contrastive self-supervised learning task to learn from massive amounts of EEG data. Frontiers in Human Neuroscience, 15:653659, 2021. doi: 10.3389/fnhum.2021.653659.DOI
[14]Gayal Kuruppu, Neeraj Wagh, Vaclav Kremen, Sandipan Pati, Gregory Worrell, and Yogatheesan Varatharajah. EEG foundation models: A critical review of current progress and future directions. arXiv preprint arXiv:2507.11783, 2025. URL https://arxiv.org/abs/2507.11783v3.
[15]Gabriel Jason Lee, Jathurshan Pradeepkumar, and Jimeng Sun. Test-time adaptation for EEG foundation models: A systematic study under real-world distribution shifts. arXiv preprint arXiv:2604.16926, 2026. URL https://arxiv.org/abs/2604.16926.
[16]Na Lee, Konstantinos Barmpas, Yannis Panagakis, Dimitrios Adamos, Nikolaos Laskaris, and Stefanos Zafeiriou. Are Large Brainwave Foundation Models Capable Yet? Insights from Fine-Tuning. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pp. 32878–32888. PMLR, 2025. URL https://proceedings.mlr.press/v267/lee25a.html.
[17]Xinhui Li, Alex Fedorov, Mrinal Mathur, Anees Abrol, Gregory Kiar, Sergey Plis, and Vince Calhoun. Learning pipeline-invariant representation for robust brain phenotype prediction. In Data-centric Machine Learning Research Workshop at the 40th International Conference on Machine Learning, 2023. URL https://dmlr.ai/assets/accepted-papers/114/CameraReady/ICML_DMLR_2023.pdf.
[18]Dingkun Liu, Yuheng Chen, Zhu Chen, Zhenyao Cui, Yaozhi Wen, Jiayu An, Jingwei Luo, and Dongrui Wu. EEG-FM-Compass: Progress, benchmarking, and future directions for EEG foundation models. National Science Review, 2026. doi: 10.1093/nsr/nwag466. URL https://doi.org/10.1093/nsr/nwag466.
[19]Jingying Ma, Feng Wu, Yucheng Xing, Qika Lin, Tianyu Liu, Chenyu Liu, Ziyu Jia, and Mengling Feng. SCOPE: Structured prototype-guided adaptation for EEG foundation models with limited labels. arXiv preprint arXiv:2602.17251, 2026. URL https://arxiv.org/abs/2602.17251v3.
[20]Navid Mohammadi Foumani, Geoffrey Mackellar, Soheila Ghane, Saad Irtza, Nam Nguyen, and Mahsa Salehi. EEG2Rep: Enhancing self-supervised EEG representation through informative masked inputs. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 5544–5555, 2024. doi: 10.1145/3637528.3671600.DOI
[21]Navid Mohammadi Foumani, Soheila Ghane, Nam Nguyen, Mahsa Salehi, Geoffrey I. Webb, and Geoffrey Mackellar. EEG-X: Device-agnostic and noise-robust foundation model for EEG. arXiv preprint arXiv:2511.08861, 2025. URL https://arxiv.org/abs/2511.08861.
[22]Iyad Obeid and Joseph Picone. The Temple University Hospital EEG Data Corpus. Frontiers in Neuroscience, 10:196, 2016. doi: 10.3389/fnins.2016.00196.DOI
[23]Gerwin Schalk. EEG Motor Movement/Imagery Dataset. PhysioNet, 2009. URL https://physionet.org/content/eegmmidb/1.0.0/.
[24]Toyotaro Suzumura, Hiroki Kanezashi, and Shotaro Akahori. Graph adapter of EEG foundation models for parameter efficient fine tuning. arXiv preprint arXiv:2411.16155, 2024. URL https://arxiv.org/abs/2411.16155v1.
[25]Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. MLP-Mixer: An all-MLP architecture for vision. In Advances in Neural Information Processing Systems, volume 34, 2021. URL https://proceedings.neurips.cc/paper/2021/hash/cba0a4ee5ccd02fda0fe3f9a3e7b89fe-Abstract.html.arxiv.org/abs/2105.01601
[26]Jiquan Wang, Sha Zhao, Zhiling Luo, Yangxuan Zhou, Haiteng Jiang, Shijian Li, Tao Li, and Gang Pan. CBraMod: A criss-cross brain foundation model for EEG decoding. In International Conference on Learning Representations, 2025. URL https://proceedings.iclr.cc/paper_files/paper/2025/file/bbbd6d915cb90be21c12548a82d45ecd-Paper-Conference.pdf.arxiv.org/abs/2412.07236
[27]Wei Xiong, Jiangtong Li, Jie Li, Kun Zhu, and Changjun Jiang. EEG-FM-Bench: A comprehensive benchmark for the systematic evaluation and diagnostic analyses of EEG foundation models. In Proceedings of the 43rd International Conference on Machine Learning, volume 306 of Proceedings of Machine Learning Research, pp. 140324–140359. PMLR, 2026. URL https://proceedings.mlr.press/v306/xiong26f.html.arxiv.org/abs/2508.17742
[28]Chaoqi Yang, M. Brandon Westover, and Jimeng Sun. BIOT: Biosignal transformer for cross-data learning in the wild. In Advances in Neural Information Processing Systems, volume 36, 2023. URL https://proceedings.neurips.cc/paper_files/paper/2023/hash/f6b30f3e2dd9cb53bbf2024402d02295-Abstract-Conference.html.DOI
[29]Junjie Yu, Zihan Deng, Jianyu Zhang, Junrong Mu, Jiahui An, Wenxiao Ma, Ziling Lu, Yue Wang, Yan Zhu, Kexin Lou, and Quanying Liu. Understanding and correcting low-frequency bias in EEG foundation model. arXiv preprint arXiv:2608.01898, 2026. URL https://arxiv.org/abs/2608.01898.
[30]Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. MetaFormer is actually what you need for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. URL https://arxiv.org/abs/2111.11418.
[31]Tongtian Yue, Xuange Gao, Shuning Xue, Yepeng Tang, Longteng Guo, Jie Jiang, and Jing Liu. BrainGPT: Unleashing the potential of EEG generalist foundation model via autoregressive pre-training. arXiv preprint arXiv:2410.19779, 2024. URL https://arxiv.org/abs/2410.19779v2.