PaperlayerView exact revision on arXiv

Source-linked AI summary

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, Ziyu Jia, Chenyu Liu, Cuntai Guan

arXiv:2610.03546v1cs.LG

TL;DR

Many EEG datasets include complementary physiological signals, but available modalities vary and target labels are often unavailable for adaptation. ZeroMAG generates adapters from unlabeled target recordings while keeping the EEG foundation model frozen. Across six held-out datasets and three backbones, it improves balanced accuracy over label-free baselines and comes within 0.50 points of supervised multimodal adaptation on average.

  • Problem

    Companion signals vary across EEG datasets, and target-specific adaptation is difficult because labels are often unavailable and annotation requires expert review.

    Source: §1 · pp. 1–2

  • Method

    ZeroMAG generates target-specific adapters from unlabeled recordings and task context, using a configuration-invariant adapter and function-constrained latent generation while keeping the EFM fixed.

    Source: §1 · p. 2

  • Results

    Across six held-out datasets and three EFM backbones, ZeroMAG improves balanced accuracy by 7.22 points over EEG-only inference and 4.89 points over Direct MLP, averaging 0.50 points below supervised multimodal adaptation.

    Source: §1 · p. 3

  • Takeaways & Limitations

    ZeroMAG adds multimodal capability to a frozen EFM without target labels or target-side optimization and without repeated multimodal pretraining.

    Source: §6 · p. 10

    Evaluation uses dataset-disjoint targets with separate calibration and query subsets; only unlabeled calibration signals and a supported task inform adapter generation.

    Source: §4.1 · p. 7

Abstract

from arXiv · show

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.

1 INTRODUCTION

Companion physiological signals are increasingly recorded with EEG, but their availability varies across datasets and labels are often unavailable for target adaptation. ZeroMAG addresses this by generating target-specific adapters from unlabeled recordings while keeping the EFM fixed.

1 INTRODUCTION

  • EEG recordings increasingly include companion signals such as eye movements, muscle activity, and cardiac dynamics, while signal availability varies widely across datasets.At least 124 deep-learning studies combined EEG with EOG, EMG, ECG, or GSR between 2016 and 2024.Source: §1 · p. 1
  • Existing multimodal foundation models often require a new backbone with a fixed modality set, while adapting an EFM is difficult when target recordings lack labels.Label-free adaptation can still require target-side optimization.Source: §1 · pp. 1–2
  • A unified adapter must accommodate heterogeneous companion-signal configurations and infer target-specific adaptation from unlabeled recordings and task context.The appropriate adaptation can differ across datasets and tasks even when they use the same companion signals.Source: §1 · p. 2
  • ZeroMAG generates a target-specific adapter around a fixed EEG anchor, using modality–subject–task conditioning and function-constrained generation from source adapters.The generator is trained to match source adapters in predictions as well as parameters because weight similarity alone does not ensure functional consistency.Source: §1 · p. 2
  • Across six held-out targets and three EFM backbones, ZeroMAG improves balanced accuracy by 7.22 points over EEG-only inference and 4.89 points over direct weight regression, averaging 0.50 points below supervised multimodal adaptation.It also remains competitive with multimodal foundation models that require multimodal pretraining.Source: §1 · p. 3

2 PRELIMINARIES

The preliminaries define the frozen EEG model and adapter interface, then formulate multimodal extension as generating adapter parameters from unlabeled target data and task context. The generator is trained to match source references in both parameters and predictions.

2.1 EEG FOUNDATION MODEL

  • An EEG foundation model comprises an encoder and a prediction head, both kept frozen during source adapter fitting, generator training, and target inference.The encoder produces the EEG representation, and the compatible head produces task logits.Source: §2.1 · p. 3

2.2 ADAPTER-BASED MULTIMODAL INTEGRATION

  • The multimodal adapter is inserted between the frozen encoder and prediction head to incorporate the available companion signals without modifying the EFM.Its parameters determine the multimodal representation and the logits returned by the frozen head.Source: §2.2 · p. 3

2.3 PROBLEM FORMULATION

  • The adapter uses a shared organization of modality-specific and shared parameters, activating only branches for modalities available in a given configuration.The generator maps unlabeled target samples and task context to adapter parameters without target labels or dataset identifiers.Source: §2.3 · p. 3
  • Source triples pair unlabeled recordings and a task with a reference adapter fitted on labeled source data; generation is trained to match references in parameters and predictions.At deployment, the generator produces the target adapter in one forward pass without target-side optimization.Source: §2.3 · p. 3

3 METHODOLOGY

ZeroMAG generates a plug-in adapter by forming a target condition from unlabeled signals and task context, then generating adapter parameters in a learned latent space. Its architecture, conditioning, and two-stage training are designed to support heterogeneous modalities while preserving predictive behavior.

3.1 FRAMEWORK OVERVIEW

  • Unlabeled target samples and a supported task form a recording-specific condition that guides latent adapter generation, after which a frozen decoder maps the latent to adapter parameters.The condition extractor, conditional sampler, and decoder are fixed at deployment.Source: §3.1 · p. 4

3.2 MULTIMODAL ADAPTER ARCHITECTURE

  • A fixed EEG representation anchors a 1+N adapter, with one branch per available companion modality and a shared coordinator integrating projected features.Each branch uses temporal–spectral encoding and feature projection; only available branches are evaluated while the block schema remains fixed.Source: §3.2 · pp. 4–5

3.3 UNIFIED CONDITION CONSTRUCTION

  • The target condition combines modality, subject, and task factors because recordings with the same modalities can differ in physiology and task.Modality features are pooled as a set, sample-level features are attention-pooled for the subject factor, and the task factor is a learned embedding.Source: §3.3 · p. 5

3.4 ADAPTER GENERATOR TRAINING

  • Fitted source adapters motivate a compact block-aligned latent space, while functional discrepancy measures prediction differences through the same frozen EFM.The discrepancy is label-free and evaluated on a source subset held out from reference-adapter fitting.Source: §3.4 · p. 6

3.4.1 STAGE 1: FUNCTION-PRESERVING REPRESENTATION LEARNING

  • Stage 1 encodes adapter blocks into contextualized latent tokens and decodes them with shared experts and block-specific heads, training reconstruction alongside predictive-behavior preservation.After training, the encoder and decoder are frozen, and contextualized posterior means provide the deterministic target latent for Stage 2.Source: §3.4.1 · p. 6

3.4.2 STAGE 2: FUNCTION-CONSTRAINED CONDITIONAL GENERATION

  • Stage 2 uses conditional latent diffusion to predict noise from the target condition, then decodes the estimated clean latent into adapter weights under latent-, weight-, and function-level alignment.Those alignment terms apply only at low-noise timesteps, while denoising supervision applies at all timesteps.Source: §3.4.2 · pp. 6–7

4 EXPERIMENTS

The experiments evaluate ZeroMAG on held-out datasets with frozen EEG backbones and unlabeled target calibration. It leads target-label-free baselines, approaches supervised adaptation, and ablations and generalization tests examine the roles and scope of its components.

4.1 EXPERIMENTAL SETUP

  • The benchmark covers six held-out targets across five task types and six companion-signal types, with targets excluded from training and model selection.For each target subject, ZeroMAG uses unlabeled calibration signals and a supported task to generate an adapter, then predicts on a disjoint query subset with model parameters fixed.Source: §4.1 · p. 7

4.2 PERFORMANCE COMPARISON WITH BASELINES

  • Across CBraMod, CSBrain, and CodeBrain, gains over EEG-only inference are 7.12, 7.30, and 7.24 points; with CodeBrain, ZeroMAG exceeds SleepFM by 1.85 points on average.It trails PhysioOmni by 0.31 points and Brant-X by 0.80 points on average.Source: §4.2 · p. 7
  • ZeroMAG ranks first among target-label-free methods in all 18 backbone–dataset evaluations, improving average B-ACC by 7.22 points over EEG-only inference and 4.89 over Direct MLP.It remains 0.50 points below supervised multimodal tuning on average.Source: §4.2 · p. 7, Table 1 · p. 8

4.3 ABLATION STUDY

  • Removing subject information lowers B-ACC by 11.09 points, while shuffling subject conditions lowers it by 9.24 points, indicating the condition must align subject information with the target recording.Source: §4.3 · p. 8
  • Direct weight regression lowers B-ACC by 11.07 points, while adding the modular latent space, DiT, and MoE decoder raises it through successive ablation variants.The reported B-ACC values rise from 62.34 to 65.23, 68.34, and 73.41.Source: §4.3 · p. 8
  • Sharing the modality encoder, removing the spectral branch, and replacing learned coordination with mean fusion reduce B-ACC by 4.16, 8.12, and 6.58 points, respectively.Source: §4.3 · p. 8
  • Removing functional supervision from Stage 1, Stage 2, or both lowers B-ACC by 1.15, 1.58, and 2.87 points, respectively.Without the Stage 1 term, reconstruction error falls by 6% while prediction KL to the reference adapter rises by 31%.Source: §4.3 · pp. 8–9

4.4 GENERALIZATION AND EFFICIENCY

  • ZeroMAG recovers 89.7% and 82.6% of supervised multimodal gain on unseen combinations of known modalities and unseen modality types, respectively.Performance saturates at about 64 unlabeled calibration windows; generation takes 2 minutes versus 15 minutes for supervised target adaptation under the same protocol.Source: §4.4 · p. 9

5 RELATED WORK

Prior work either builds multimodal foundation models or adapts existing models through methods that still require target-side optimization. ZeroMAG instead generates a modular adapter for a frozen EFM from unlabeled target information.

5 RELATED WORK

  • Prior multimodal physiological models integrate signals through multimodal pretraining or EEG-to-physiology alignment, while ZeroMAG adds a plug-in adapter to an existing frozen EFM.This contrasts with approaches that train a new multimodal backbone and retain its predetermined modality set.Source: §5 · p. 9
  • Parameter-efficient adaptation still needs target-side optimization, whereas ZeroMAG generates a modular physiological adapter from unlabeled target information in a structured latent space.Unlike test-time adaptation, deployment requires no pseudo-labels, gradients, or iterative optimization.Source: §5 · p. 9

6 CONCLUSION

ZeroMAG extends frozen EEG foundation models to heterogeneous multimodal recordings through generated plug-in adapters. The appendix describes its model interfaces, adapter and condition design, and evaluation comparisons.

6 CONCLUSION

  • ZeroMAG extends frozen EFMs to heterogeneous multimodal recordings with generated plug-in adapters, without target labels or target-side optimization for the added adapter.It relies on a supplied fine-tuned EFM and adequate source-bank coverage, and cannot incorporate modality types outside M without retraining.Source: §6 · p. 10

A.1 EEG REPRESENTATIONS

  • EEG foundation models differ in tokenization, objectives, and recording domains, differences that matter when defining an adapter’s input and output interface.The appendix discusses LaBraM, EEGPT, BIOT, and Brant as examples of differing representation approaches.Source: §A.1 · p. 18

A.2 FROZEN MODELS AND SEPARATE GENERATORS

  • Each backbone uses a separately trained generator with its compatible frozen encoder and prediction head; results do not test cross-backbone checkpoint transfer.Daphnet supervision trains the compatible freezing-of-gait head, not reference adapters or the adapter generator.Source: §A.2 · p. 18

A.3 SIGNALS AND ADAPTER PARAMETERS; A.4 LATENTS AND CONDITIONING

  • Notation distinguishes EEG and companion inputs, fitted and generated adapters, unlabeled condition samples, and the deterministic latent U0 from its signal features.The latent is obtained by substituting posterior means before the cross-block transformer; it is not a new encoder architecture.Source: §A.3 · pp. 18–19, §A.4 · p. 19

A.5 PARAMETER OWNERSHIP

  • A frozen decoder remains differentiable for an output loss, while generated adapters are system outputs rather than target-fitted optimizer states.Generation may repeat denoiser evaluations without target-side gradient updates.Source: §A.5 · pp. 19–20, Table A.3 · p. 20

B.1 VARIABLE-LENGTH BLOCK LAYOUT; B.2 TEMPORAL AND SPECTRAL BRANCHES; B.3 SHARED COORDINATION AND MASKING

  • Each available modality contributes three functional blocks beside one shared block; temporal and spectral streams feed coordination, which adds an auxiliary residual while EEG stays frozen.Unavailable modalities have no parameter blocks or latent tokens and do not contribute to the prediction path.Source: §B.1 · p. 20, §B.2 · p. 20, §B.3 · p. 21

B.4 THREE CONDITION FACTORS; B.5 ORDER AND IDENTITY; B.6 SUPPORT OF THE CONDITION

  • Conditioning separates signal identity, recording-derived subject physiology, and supported task identity; order permutation invariance does not make signal identities interchangeable.FoG uses an unspecified-task embedding, and the supported block schema does not create an encoder for signal types outside M.Source: §B.4 · p. 21, §B.5 · p. 21, §B.6 · p. 21

C.1 EVALUATION TARGETS; C.2 CONTROLLED COMPARISONS

  • The six-target display spans four task families, while supervised fine-tuning is a separate labeled reference, not ZeroMAG training or a mathematical upper bound.Cross-paradigm scores are evaluations reported in this paper for models with native backbones and different multimodal procedures.Source: §C.1 · p. 22, §C.2 · p. 23

C.3 RESULTS BY TASK FAMILY

Across task families, ZeroMAG improves over EEG-only inference, with gains that are similar across three backbones and aggregate results close to supervised tuning. The reported comparisons are based on means and do not establish statistical significance or equivalence.

C.3 RESULTS BY TASK FAMILY

  • Sleep gains over EEG-only were 7.48–8.01 points, SEED-V gains 4.21–4.36, EEGMAT gains 7.61–7.89, and FoG gains 7.86–8.64.For SEED-V, recovery of supervised gain ranged from 87.34% to 88.71%; the source of this task-dependent gap is not identified.Source: §C.3 · pp. 23–24, Table C.3 · p. 23, Table C.4 · p. 24

C.4 GAINS ACROSS BACKBONES

  • All 18 backbone–dataset settings show positive mean gains over EEG-only and Direct MLP; the Direct MLP advantage ranges from 2.97 to 6.86 points.Figure C.1 shows both comparisons on a common scale; its differences are based on means, without paired uncertainty or significance inference.Source: §C.4 · p. 24, Figure C.1 · p. 25

C.5 AGGREGATE COMPARISONS

  • 7.22 points over EEG-only and 4.89 over Direct MLP, with a 0.50-point supervised gap, summarize the 18-setting equal-weight means.The corresponding advantages over mean and nearest source weights are 6.7772 and 5.9989 points.Source: §C.5.1 · p. 25, Table C.5 · p. 25

C.5.1 AGGREGATE GAIN RECOVERY

  • The ratio of aggregate gains is 93.50%, compared with 93.12% when row-wise recovery ratios are averaged.All displayed supervised gains are positive.Source: §C.5.1 · p. 26

C.5.2 BACKBONE COMPARISON

  • EEG-only gains are 7.12, 7.30, and 7.24 points for CBraMod, CSBrain, and CodeBrain; ZeroMAG means exceed label-free alternatives in every setting.Across the 18 rows, the supervised-minus-ZeroMAG margin ranges from 0.38 to 0.61 points.Source: §C.5.2 · p. 26

C.6 REPORTED VARIATION AND THE SUPERVISED GAP

  • Reported seed-to-seed standard deviations are not pairwise confidence intervals; 0.38–0.61-point supervised gaps across settings do not establish equivalence.The mean supervised-reference gap is 0.5022 points.Source: §C.6 · p. 26

C.7 CROSS-PARADIGM COMPARISONS

  • With CodeBrain, ZeroMAG exceeds SleepFM by 1.8533 points on average and trails PhysioOmni and Brant-X by 0.3050 and 0.8033, respectively.These comparisons do not isolate backbone architecture, pretraining exposure, or target adaptation as a single controlled factor.Source: §C.7 · p. 26

D.1 TWO TRAINING STAGES

ZeroMAG separates offline learning into adapter representation learning and conditional generation, then deploys the trained components using unlabeled calibration data with fixed parameters. Its training objectives connect adapter weights, latent representations, and predictions through the frozen EFM.

D.1 TWO TRAINING STAGES

  • Stage 1 fits reference adapters on labeled source data and learns an encoder and decoder; Stage 2 learns conditional generation while retaining the Stage 1 representation.Reference-adapter parameter targets and predictions on a separate source subset provide parameter and functional supervision, respectively.Source: §D.1 · pp. 26–27, Figure D.1 · p. 27, §D.2 · p. 28

D.3 STAGE 1: REPRESENTATION LEARNING

  • Stage 1 combines block reconstruction, direction alignment, functional agreement through the frozen EFM, posterior regularization, and expert balancing.The objectives constrain different properties: parameters, direction, predictions, the Gaussian posterior, and expert use.Source: §D.3 · p. 28

D.4 STAGE 2: CONDITIONAL GENERATION

  • Stage 2 conditions a diffusion noise predictor on unlabeled recording information and task context, with latent, weight, and functional auxiliary losses active only at low-noise timesteps.The diffusion loss applies at every timestep; the auxiliary losses use the gate ᾱt > 0.5.Source: §D.4 · pp. 28–29, §E.7 · p. 31

D.5 DEPLOYMENT

  • At deployment, unlabeled calibration data and the supported task produce an adapter in one forward pass, with fixed model parameters and no target-side optimization.Query samples are excluded from condition construction, and sampled initial noise makes generation stochastic.Source: §D.5 · p. 29

E.1 CANONICAL LATENT COORDINATES

  • The canonical Stage 2 target applies cross-block contextualization to posterior means, rather than estimating the mean of the nonlinear contextualized posterior.In general, contextualizing the expected pre-contextual tokens differs from taking the expectation after contextualization.Source: §E.1 · p. 29

E.5 PREDICTION AGREEMENT

  • Functional matching aligns generated predictions with a source reference, not necessarily with true labels or target-query behavior.The reference predictions inherit both useful behavior and errors from an adapter fitted using source labels.Source: §E.4 · p. 30, §E.5 · p. 31

E.9 GRADIENTS THROUGH A FROZEN DECODER AND EFM

  • Functional supervision trains the generator through the frozen decoder and EFM because freezing their stored parameters does not remove their input derivatives.The resulting first-order prediction perturbation depends on both decoder and EFM sensitivity, so equal latent errors can yield different logit errors.Source: §E.9 · p. 32, §E.10 · p. 32

F.1 MEASURED TARGET-SIDE TIME

  • About 2 minutes for condition construction and adapter generation versus 15 minutes for supervised target adaptation in the measured procedures.The timed procedures exclude source-bank construction and offline generator training.Source: §F.1 · p. 32

F.2 COST DECOMPOSITION

The cost discussion distinguishes measured target-side procedures from offline training and query prediction. It also describes how deployment volume, latent computation, and adapter output costs affect the comparison.

F.2 COST DECOMPOSITION

  • The 7.5× figure compares target-side condition construction plus adapter generation with supervised adaptation, not the entire offline-plus-online pipeline.No component-wise timing breakdown or speedup for subsequent query prediction is established.Source: Figure F.1 · p. 33, §F.2 · p. 33

F.3 AMORTIZATION ACROSS TARGETS

  • The generated-adapter route costs Toff + NTgen across N deployments, and its symbolic breakeven condition is N > Toff/(13 minutes).No numerical offline training cost or measured deployment count is assigned.Source: §F.3 · p. 33

F.4 LATENT AND OUTPUT COMPLEXITY

  • Reducing latent width can reduce denoiser computation, but adapter output storage and installation remain; sparse expert selection alone establishes no measured runtime speedup.Functional supervision adds a prediction path during offline training, not target deployment.Source: §F.4 · p. 33

G.1 COMPONENT ABLATIONS

The ablations compare condition, generator, adapter, and functional-supervision choices. They report declines relative to the full system and distinguish separate interventions rather than treating them as additive effects.

G.1 COMPONENT ABLATIONS

  • The remaining reported decreases are 8.18 points for modular latent MLP, 5.07 for latent DiT without MoE, 4.16 for a shared modality encoder, 8.12 for temporal-only encoding, and 6.58 for mean fusion.These are separate variants, not additive component contributions.Source: §G.1 · p. 34, Figure G.1 · p. 35
  • 11.09 points for task-plus-modality-only conditioning, 9.24 for shuffled subject conditions, and 11.07 for direct weight MLP, relative to Full ZeroMAG at 73.41 B-ACC.The ablation aggregates CBraMod results over nine held-out target datasets; each decrease uses the same full-model reference.Source: §G.1 · pp. 33–34, Table G.1 · p. 34

G.2 INTERPRETING THE INTERVENTIONS

  • The 9.24-point shuffled-subject decrease tests recording alignment, while the 11.09-point task-plus-modality decrease tests removing subject information.Shuffling retains the set of subject-condition vectors but changes their alignment with recordings.Source: §G.2 · p. 34

G.3 FUNCTIONAL SUPERVISION

  • Removing Stage 1 functional supervision lowers reconstruction error by 6% but raises generated-weight error by 5% and prediction KL by 31%; B-ACC falls 1.15 points.Removing Stage 2 functional supervision raises generated-weight error by 8%, prediction KL by 48%, and lowers B-ACC by 1.58 points; removing both losses lowers B-ACC by 2.87 points and raises KL by 82%.Source: §G.3 · p. 35, Table G.2 · p. 36

H.2 SEPARATE BACKBONE INSTANTIATIONS

  • Each of CBraMod, CSBrain, and CodeBrain uses a separately trained generator, while its EFM encoder and compatible prediction head remain fixed.The target task selects the output layer; no target labels or target-side updates are used.Source: §H.2 · p. 36

H.3 DATASET HOLDOUT AND RECORDING SPLITS

The evaluation holds target datasets out of ZeroMAG training and model selection, uses disjoint calibration and query subsets, and reports several aggregation and metric boundaries. Modality-holdout results apply to the supported interface, not arbitrary signal types.

H.3 DATASET HOLDOUT AND RECORDING SPLITS

  • All target datasets are excluded from encoder/head training, reference-adapter fitting, both generator stages, and model selection; target recordings are used for unlabeled conditioning and evaluation.Calibration and query subsets are disjoint for each target subject, and query labels are used for evaluation rather than adaptation.Source: §H.3 · p. 36

H.4 SOURCE ADAPTER BANK

  • The source bank contains 311 eligible reference adapters per backbone after excluding records from subjects present in target datasets, including across related cohorts.Dataset-balanced sampling prevents large source datasets from dominating adapter batches.Source: §H.4 · p. 37

H.5 BASELINE INTERFACES

  • Mean-source transfer averages compatible fitted source vectors, nearest-source transfer reuses a selected source adapter, and Direct MLP maps the condition to adapter weights.Native multimodal models are separate cross-paradigm references.Source: §H.5 · p. 37

H.6 REPORTED METRICS

  • Balanced accuracy weights class recalls equally, and its percentage-point differences do not determine kappa, weighted F1, AUROC, or AUPRC.AUROC and AUPRC require score-level predictions.Source: §H.6 · p. 38

H.7 SEED VARIATION AND PAIRWISE COMPARISONS

  • Reported standard deviations describe seed-to-seed variation, while marginal standard deviations do not determine the covariance needed for paired error bars.A standard deviation is distinct from a standard error or confidence interval.Source: §H.7 · p. 38

H.8 AGGREGATION

  • The six-target averages equally weight 18 backbone–dataset means, whereas the 73.41 Full ZeroMAG ablation average covers nine target datasets.These summaries have different evaluation scopes; the six-target weighting is not by participant or window.Source: §H.8 · p. 38

I.1 MODALITY HOLDOUT

  • ZeroMAG recovers 89.7% of supervised multimodal gain for unseen combinations of known modalities and 82.6% for held-out modality types.The type-holdout recovery is 7.1 percentage points lower, and both results concern the supported modality interface.Source: §I.1 · p. 38, Figure I.1 · p. 39

I.2 CALIBRATION DATA

The calibration and source-bank experiments examine separate resource requirements for ZeroMAG. Performance saturates at about 64 calibration windows in the reported setting, while using half the source adapter bank retains nearly all of its multimodal gain.

I.2 CALIBRATION DATA

  • Performance saturates at approximately 64 unlabeled calibration windows in the reported experiment.These windows form the target condition; they are not a target optimization batch, because there is no target-side fitting step.Source: §I.2 · p. 39
  • The saturation point is specific to the evaluated calibration setting, not a universal duration across tasks.Window duration and signal sampling rates vary across physiological tasks.Source: §I.2 · p. 39

I.3 SOURCE-BANK SIZE

  • Using 50% of the source adapter bank retains 97.2% of the full-bank multimodal gain in the reported data-efficiency experiment.This is a 2.8% reduction in reference gain, not a 2.8-point reduction in B-ACC.Source: §I.3 · p. 39
  • Calibration windows are collected from a new target recording, whereas source adapters are fitted and organized offline; the summaries do not establish a joint optimum.Reducing one resource requirement does not directly reduce the other.Source: §I.3 · p. 39

J COMPARISON WITH RELATED METHODS

ZeroMAG differs from multimodal pretraining and target-time adaptation by generating plug-in adapters for a selected frozen EFM from unlabeled physiological recordings. Its related-method comparison separates adaptation mechanisms and notes that the evaluated setup does not establish transfer to arbitrary architectures or unsupported tasks.

J.1 EXTENDED RELATED WORK

  • Unlike SleepFM and PhysioOmni, which learn multimodal representations through pretraining, and Brant-X, which trains an EXG encoder, ZeroMAG generates a plug-in adapter while preserving the selected EFM's frozen encoder and head at deployment.SleepFM and PhysioOmni learn multimodal physiological representations through pretraining; Brant-X aligns EEG with other physiological signals using a pretrained EFM.Source: §J.1 · p. 40
  • Adapters, LoRA, prompt tuning, and test-time adaptation differ from ZeroMAG because they optimize target parameters, while ZeroMAG uses an offline-trained generator with no target-side gradient step.The cited parameter-efficient methods still optimize on labeled target data; test-time and unsupervised domain adaptation update parameters on target data.Source: §J.1 · p. 40
  • ZeroMAG replaces per-target optimization with generation conditioned on unlabeled physiological recordings, accommodating variable modality configurations and combining parameter targets with functional agreement through the frozen EFM.Unlike text-conditioned parameter-generation approaches, its condition comes from physiological recordings.Source: §J.1 · p. 40

J.2 ADAPTATION MECHANISMS

  • Table J.1 compares adaptation mechanisms by the object adapted and information used, distinguishing target-time optimization from conditional parameter generation.The listed methods are a methodological comparison, not additional experimental baselines.Source: §J.2 · p. 40, Table J.1 · p. 40
  • ZeroMAG represents companion-signal adapters as a variable-length semantic block sequence and evaluates decoded weights through the frozen EEG prediction path; generators are trained separately for each of three backbones.Its target condition comes from the recording and supported task.Source: §J.2 · p. 41
  • The reported setup depends on a supplied frozen EFM, compatible prediction head, source reference-adapter quality, and calibration-data coverage, and does not establish transfer to arbitrary architectures or tasks outside the supported vocabulary.The stated evidence concerns the evaluated interfaces.Source: §J.2 · p. 41
Loading 2610.03546v1…