ZeroMAG: Zero-Shot Multimodal Adapter Generation for Plug-and-Play EEG Foundation Models

Yubo Wang; Jingying Ma; Xinliang Zhou; Yangxuan Zhou; Jiquan Wang; Sha Zhao; Yiyuan Yang; Yi Ding

cs.LGOct 2, 2026

ZeroMAG: Zero-Shot Multimodal Adapter Generation for Plug-and-Play EEG Foundation Models

Authors: Yubo Wang, Jingying Ma, Xinliang Zhou, Yangxuan Zhou, Jiquan Wang, Sha Zhao, Yiyuan Yang, Yi Ding,

Organizations: Brown University · Stanford University · National University of Singapore · Zhejiang University ·

Abstract

EEG foundation models (EFMs) capture reusable knowledge from large-scale EEG data, while many EEG recordings also include companion physiological signals that provide complementary information beyond the EEG-only interface. The challenge is to preserve this pretrained knowledge while extending the EFM to heterogeneous multimodal recordings through an adaptation inferred from unlabeled target data. We introduce ZeroMAG, a zero-shot multimodal adapter generation framework that extends a frozen EEG encoder and prediction head using unlabeled target recordings, without target labels or target-side optimization. The target datasets are held out from all model training and selection in the ZeroMAG pipeline. ZeroMAG organizes companion modalities around a configuration-invariant adapter, constructs a modality-subject-task condition from unlabeled recordings and task context, and generates adapter weights in a function-constrained latent space learned from source adapters. Across six held-out target datasets and three EFM backbones, ZeroMAG improves balanced accuracy by 7.22 percentage points over EEG-only inference and 4.89 points over direct weight regression, while coming within 0.50 points of supervised multimodal adaptation on average. Ablations further show that removing functional supervision from either representation learning or conditional generation degrades generated-adapter performance, confirming the contribution of both components.

Figures & tables

Figure 1: Multimodal EEG landscape and representative results. (a) Physiological information captured by common companion signals. (b) Publication trends, task-wise multimodal prevalence, and companion modality availability, based on prior surveys ( Lee et al., 2025 ; Yeung & Chu, 2022 ) . (c) Balanced accuracy on six held-out target datasets, comparing three EFM backbones with and without ZeroMAG against SleepFM and PhysioOmni.

Figure 2: Overview of ZeroMAG. The framework combines a configuration-invariant multimodal adapter, target-derived condition construction, and function-constrained conditional generation for zero-shot adapter synthesis.

CBraMod
Method	ISRUC-S3	HMC	MASS	SEED-V	EEGMAT	FoG
EEG-only	$75.83$ $\pm 0.76$	$72.78$ $\pm 0.50$	$74.29$ $\pm 0.58$	$40.62$ $\pm 1.24$	$73.38$ $\pm 2.48$	$76.61$ $\pm 1.35$
Supervised tuning	$84.04$ $\pm 0.88$	$80.81$ $\pm 0.41$	$82.41$ $\pm 0.68$	$45.44$ $\pm 1.58$	$81.46$ $\pm 1.93$	$85.02$ $\pm 1.21$
Mean source weights	$74.31$ $\pm 0.82$	$72.93$ $\pm 0.49$	$75.91$ $\pm 0.63$	$39.83$ $\pm 1.13$	$73.52$ $\pm 0.92$	$76.73$ $\pm 0.78$
Nearest source weights	$76.63$ $\pm 0.61$	$73.18$ $\pm 0.52$	$76.35$ $\pm 0.51$	$40.68$ $\pm 1.37$	$74.63$ $\pm 0.81$	$77.52$ $\pm 0.73$
Direct MLP	$76.89$ $\pm 0.73$	$73.43$ $\pm 0.55$	$77.48$ $\pm 0.65$	$40.99$ $\pm 1.05$	$75.21$ $\pm 1.02$	$78.03$ $\pm 0.62$

Table 1: Balanced accuracy (%, mean

\pm

std over five seeds) on six held-out target datasets. Within each EEG-backbone group, ZeroMAG and the label-free baselines use the same frozen encoder and prediction head, without target labels or target-side optimization. Supervised multimodal fine-tuning is a separate reference using target training labels and all available modalities. Bold entries indicate the best target-label-free result in each group.

Component	Variant	B-ACC $\uparrow$	$\Delta$ $\downarrow$
Condition	Task + modality only	$62.32$ $\pm 0.41$	$11.09$ $\pm 0.72$
	Shuffled subject condition	$64.17$ $\pm 0.53$	$9.24$ $\pm 0.98$
Generator	Direct weight MLP	$62.34$ $\pm 0.67$	$11.07$ $\pm 0.88$
	Modular latent MLP	$65.23$ $\pm 0.62$	$8.18$ $\pm 0.83$
	Modular latent DiT (w/o MoE)	$68.34$ $\pm 0.21$	$5.07$ $\pm 0.71$
Adapter	Shared modality encoder	$69.25$ $\pm 0.33$	$4.16$ $\pm 0.81$

Table 2: Ablation results with CBraMod, averaged over all nine held-out datasets (% B-ACC, mean

\pm

std over five seeds). Ablated results are shaded blue, and

\Delta

denotes the decrease from Full ZeroMAG in the same run. Ablations use a separate run under the same protocol, so Full differs from the Table 1 average.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
$X^{\mathrm{EEG}}$	EEG input to the frozen encoder.
$\mathcal{M}$	Complete set of supported companion modalities.
$A\subseteq\mathcal{M}$	Companion modalities available in a particular recording.
$X^{A}$	Collection $\{X^{m}:m\in A\}$ of available companion inputs.
$R=\{r_{i}\}$	Recording samples, with $r_{i}=(X_{i}^{\mathrm{EEG}},X_{i}^{A})$ .
$R^{\mathrm{fit}}$	Source samples whose labels are used to fit a reference adapter.

Appendix

Table A.1: Signal and adapter notation. Dimensions are symbolic here; numerical settings are listed in Table H.1 .

Symbol	Meaning
$e(\cdot;\eta)$	Stochastic block encoder, including the cross-block transformer.
$\bar{e}(\cdot;\eta)$	Deterministic encoder path using posterior means.
$\mu_{b},\ell_{b}$	Mean and log variance of the Gaussian for block $b$ ; $\ell_{b}$ is introduced here for the posterior expansion.
$U_{b,\mathrm{pre}}$	Sampled block token before contextualization.
$E_{b},\mathrm{CBT}$	Block identity embedding and cross-block transformer.
$U,U_{0},U_{t}$	Contextualized sample, canonical latent, and noise-corrupted canonical latent.

Appendix

Table A.2: Latent, condition, and optimization notation.

Component	Reference fitting	Stage 1	Stage 2	Deployment
EEG encoder and head	Fixed	Fixed	Fixed	Fixed
Reference adapter $\omega^{*}$	Updated	Fixed	Fixed	Not required
Encoder $\eta$	Not used	Updated	Fixed	Not required
Decoder $\psi$	Not used	Updated	Fixed	Fixed
Condition encoder $\xi$	Not used	Not used	Updated	Fixed
Denoiser $\theta$	Not used	Not used	Updated	Fixed

Appendix

Table A.3: Parameter updates in the ZeroMAG pipeline. “Fixed” describes parameter values, not a zero input Jacobian.

CBraMod
Method	CAP	DREAMER	BCI-IV
EEG-only	$53.86$ $\pm 0.83$	$43.91$ $\pm 1.04$	$37.82$ $\pm 1.31$
Supervised tuning	$68.16$ $\pm 0.64$	$60.20$ $\pm 0.98$	$51.11$ $\pm 0.89$
Mean source weights	$53.68$ $\pm 0.97$	$43.45$ $\pm 1.35$	$37.54$ $\pm 1.70$
Nearest source weights	$54.78$ $\pm 0.55$	$45.08$ $\pm 1.23$	$39.26$ $\pm 0.91$
Direct MLP	$55.93$ $\pm 0.65$	$46.05$ $\pm 0.74$	$41.00$ $\pm 1.61$

Appendix

Table C.1: B-ACC (%, mean

\pm

std over five seeds) on CAP, DREAMER, and BCI-IV. Protocol and layout as in main Table 1 .

Target	Task	Holdout
ISRUC-S3	Sleep staging	Dataset
HMC	Sleep staging	Dataset
MASS	Sleep staging	Dataset
SEED-V	Emotion recognition	Dataset
EEGMAT	Cognitive-load estimation	Dataset
FoG	Freezing-of-gait detection	Task

Appendix

Table C.2: Task coverage and holdout type of the target datasets.

CBraMod
Method	ISRUC-S3	HMC	MASS
EEG-only	$75.83\pm 0.76$	$72.78\pm 0.50$	$74.29\pm 0.58$
Supervised reference	$84.04\pm 0.88$	$80.81\pm 0.41$	$82.41\pm 0.68$
Mean source weights	$74.31\pm 0.82$	$72.93\pm 0.49$	$75.91\pm 0.63$
Nearest source weights	$76.63\pm 0.61$	$73.18\pm 0.52$	$76.35\pm 0.51$
Direct MLP	$76.89\pm 0.73$	$73.43\pm 0.55$	$77.48\pm 0.65$

Appendix

Table C.3: Sleep staging: balanced accuracy (%, mean

\pm

reported SD over five seeds). Bold identifies ZeroMAG in each backbone group.

CBraMod
Method	SEED-V	EEGMAT	FoG
EEG-only	$40.62\pm 1.24$	$73.38\pm 2.48$	$76.61\pm 1.35$
Supervised reference	$45.44\pm 1.58$	$81.46\pm 1.93$	$85.02\pm 1.21$
Mean source weights	$39.83\pm 1.13$	$73.52\pm 0.92$	$76.73\pm 0.78$
Nearest source weights	$40.68\pm 1.37$	$74.63\pm 0.81$	$77.52\pm 0.73$
Direct MLP	$40.99\pm 1.05$	$75.21\pm 1.02$	$78.03\pm 0.62$

Appendix

Table C.4: Emotion, cognitive load, and gait: balanced accuracy (%, mean

\pm

reported SD over five seeds). Bold identifies ZeroMAG in each backbone group.

Figure C.1: Gains in balanced accuracy over EEG-only and Direct MLP, in percentage points. Both panels use the same scale. Values are differences of means from main Table 1 ; no paired uncertainty or significance is inferred.

Backbone	Target	$\Delta^{\mathrm{EEG}}$	$\Delta^{\mathrm{MLP}}$	$G^{\mathrm{sup}}$	Recovery (%)
CBraMod	ISRUC-S3	7.76	6.70	0.45	94.52
CBraMod	HMC	7.51	6.86	0.52	93.52
CBraMod	MASS	7.74	4.55	0.38	95.32
CBraMod	SEED-V	4.21	3.84	0.61	87.34
CBraMod	EEGMAT	7.61	5.78	0.47	94.18
CBraMod	FoG	7.86	6.44	0.55	93.46

Appendix

Table C.5: Arithmetic derived from main Table 1 ; no additional experimental runs. Recovery is

100(z-e)/(s-e)

for each row.

Figure C.2: Means and reported

\pm 1

SD for four methods within each target. Marker shapes identify backbones. Horizontal limits vary by target to show within-target differences; all values are balanced accuracy (%).

Figure D.1: Training and deployment within one separately trained backbone instantiation. Stage 1 produces the encoder and decoder. Stage 2 uses a fixed canonical latent and propagates auxiliary-loss gradients through the frozen decoder. Deployment uses unlabeled calibration data and fixed model parameters.

Figure F.1: Target-side wall-clock time measured programmatically: condition construction plus adapter generation versus supervised adaptation, under the same hardware and protocol. The 7.5

\times

ratio uses the reported 2- and 15-minute measurements. No timing variance or component-wise timing breakdown is inferred.

Component	Variant	B-ACC	Decrease
Condition	Task + modality only	$62.32\pm 0.41$	$11.09\pm 0.72$
	Shuffled subject	$64.17\pm 0.53$	$9.24\pm 0.98$
Generator	Direct weight MLP	$62.34\pm 0.67$	$11.07\pm 0.88$
	Modular latent MLP	$65.23\pm 0.62$	$8.18\pm 0.83$
	Latent DiT, no MoE	$68.34\pm 0.21$	$5.07\pm 0.71$
Adapter	Shared encoder	$69.25\pm 0.33$	$4.16\pm 0.81$

Appendix

Table G.1: Component ablations from main Table 2 . B-ACC and decrease are reported in percentage points as mean

\pm

standard deviation over five seeds. The aggregation uses CBraMod over all nine target datasets, as in main Table 2 .

Figure G.1: Performance decrease under component ablation. Markers show the decreases reported in main Table 2 relative to full ZeroMAG (B-ACC

73.41\pm 0.71

). Whiskers reproduce the table’s reported standard deviations of the decreases; they are not newly calculated confidence intervals. Variants are grouped by condition, generator, and adapter. These are separate interventions, not additive contributions. The aggregation covers all nine target datasets, as in main Table 2 .

Setting	B-ACC (%)	Stage 1 recon. error	Generated weight error	Prediction KL
Full	$73.41\pm 0.71$	1.00	1.00	1.00
w/o Stage 1 functional	$72.26\pm 0.76$	0.94	1.05	1.31
w/o Stage 2 functional	$71.83\pm 0.79$	1.00	1.08	1.48
w/o both functional	$70.54\pm 0.86$	0.93	1.14	1.82

Appendix

Table G.2: Functional-loss ablation. Parameter-error and KL columns are normalized to their respective Full values; lower is better. Reconstruction error evaluates the Stage 1 output against the reference adapter. Generated-weight error evaluates the final generated adapter against the same reference. KL evaluates the final generated adapter.

Group	Setting	Value
Representation	Active parameter blocks / latent shape	$3M+1$ / $(3M+1)\times 128$
	CBT layers / hidden width / heads	$2$ / $256$ / $8$
Decoder	Experts / routing	$4$ / learned soft top- $2$
	Expert MLP widths	$128\rightarrow 256\rightarrow 128$
Both stages	Optimizer / learning rate / weight decay	AdamW / $10^{-4}$ / $10^{-4}$
	Batch size	$64$ reference-adapter samples

Appendix

Table H.1: Implementation settings.

M

is the number of active companion modalities. Auxiliary Stage 2 losses share the same low-noise gate; diffusion loss is always active.

Figure I.1: Reported generalization and source-bank summaries from the main text. Left: recovery of supervised gain for unseen combinations and held-out modality types. Right: retention of the full-bank multimodal gain when using half the source bank. The panels have different reference denominators; no seed-level intervals are inferred.

Method family	Adapted object	Adaptation mechanism
Adapters ( Houlsby et al., 2019 )	Inserted modules	Optimize a small set of task-specific parameters.
LoRA ( Hu et al., 2022 )	Low-rank weight updates	Learn an update while freezing the original weights.
Prompt tuning ( Lester et al., 2021 )	Continuous input prompts	Optimize prompt embeddings for the task.
Tent ( Wang et al., 2021 )	Normalization parameters	Minimize prediction entropy on test inputs.
SHOT ( Liang et al., 2020 )	Target feature extractor	Adapt with a fixed source classifier and unlabeled target data.
Hypernetworks ( Ha et al., 2017 ; Karimi Mahabadi et al., 2021 )	Generated network or adapter weights	Predict parameters from a learned conditioning representation.

Appendix

Table J.1: Adaptation mechanisms in related work. Target-time optimization and conditional parameter generation are separate design choices.

Multimodal Learning

Multimodal learning studies how models combine and reason over different data types, such as images, text, audio, signals, and structured records.

See all 335 papers

Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes
Apr 25, 2026Kuntal Kokate, Bruno Aristimunha, Dung Truong +1Fine-Tuning EEG Decoding
Swartz Center for Computational Neuroscience, Institute for Neural Computation, University of California San Diego, La Jolla, CA 92093, USA · Yneuro, Paris, France · Centre de Recherche Cerveau et Cognition, CNRS, Université Paul Sabatier, Toulouse, France
SCOPE: Structured Prototype-Guided Adaptation for EEG Foundation Models with Limited Labels
Feb 19, 2026Jingying Ma, Feng Wu, Yucheng Xing +5EEG Foundation Models Few-Shot Domain Adaptation
Saw Swee Hock School of Public Health, National University of Singapore, Singapore · Guangzhou Research Translation and Innovation Institute, National University of Singapore, Guangzhou, China · College of Computing and Data Science, Nanyang Technological University, Singapore +3
Parameter-Efficient Self-Supervised Adaptation for EEG-FM under Fixed Computational Budgets
Aug 25, 2026Meghal Dani, Stefanie LiebeEEG Foundation Models Parameter-Efficient Fine-Tuning
University of Tübingen, Germany · Hertie Institute for AI in Brain Health (Hertie AI), Tübingen, Germany · Dept. of Neurology and Epileptology, University Clinic Tübingen, Germany +1

Apr 25, 2026J/K move · Enter open · S save

Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes

Kuntal Kokate, Bruno Aristimunha, Dung Truong, Arnaud Delorme

Swartz Center for Computational Neuroscience, Institute for Neural Computation, University of California San Diego, La Jolla, CA 92093, USA · Yneuro, Paris, France · Centre de Recherche Cerveau et Cognition, CNRS, Université Paul Sabatier, Toulouse, France

Fine-Tuning EEG Decoding Negative Transfer EEG Foundation Models

Open paper AskPDF arXiv

Scaling EEG foundation models requires pooling data across heterogeneous electrode montages, a prerequisite both for larger pretraining corpora and for downstream deployment. We present the first systematic comparison of four channel adaptation methods (Conv1d projection, spherical spline interpolation (SSI), source-space decomposition, and Riemannian re-centering) across five pretrained EEG foundation models (5M--157M parameters), five downstream tasks, and two training regimes with 10--15 random seeds each. We find that rigid-montage models (BENDR, Neuro-GPT) require external adaptation, while flexible models (EEGPT, CBraMod) match or exceed it natively when fine-tuned but benefit from external methods under frozen-encoder deployment. A probe-SFT asymmetry exists: external adaptation can cause severe negative transfer during fine-tuning of flexible models. The optimal method is architecture-dependent (Conv1d for BENDR, SSI/Riemannian for Neuro-GPT, source-space decomposition for depression detection), and 5M-parameter CBraMod outperforms models up to 31

\times

larger on 4/5 datasets, consistent with independent findings that compact EEG-specific architectures can match larger models.

Abstract

Multimodal Learning

Explore similar work

Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes

SCOPE: Structured Prototype-Guided Adaptation for EEG Foundation Models with Limited Labels

Parameter-Efficient Self-Supervised Adaptation for EEG-FM under Fixed Computational Budgets