acceptodds
Under review as a conference paper at ICLR 2027

Latent-Space Predictive Learning for EEG Foundation Models via Geometric Modeling and Masking

14 views

Abstract

EEG foundation models aim to learn general-purpose neural representations that can transfer across heterogeneous downstream tasks. However, existing models often rely heavily on end-to-end fine-tuning, while their frozen representations remain substantially less effective, indicating limited representation transferability. We identify three key limitations in current pretraining paradigms: a focus on low-level waveform reconstruction rather than transferable contextual representation learning, insufficient integration of electrode geometry into inter-channel modeling, and local information leakage under random masking. To address these limitations, we propose **EEG-GPRL**, a **G**eometry-aware **P**redictive **R**epresentation **L**earning framework that learns transferable contextual EEG representations by predicting teacher representations in latent space. EEG-GPRL incorporates Geometry-aware Attention (GeoAttn) to couple channel interactions with relative electrode geometry. We propose Spatiotemporal Gaussian Field Masking (ST-GFM) to construct correlated masked contexts to reduce local information leakage. VISReg is used to prevent representational collapse. Across nine datasets spanning six EEG decoding paradigms, EEG-GPRL achieves the highest balanced accuracy on eight datasets under both KNN and frozen probing, demonstrating strong transferability across heterogeneous tasks. Ablation studies further validate the complementary contributions of its individual components.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG Foundation Model, Latent-Space Predictive, Geometry-Aware Model, Gaussian Field Mask

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9Fd9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EEG-GPRL combines absolute electrode-coordinate features, three-dimensional rotary attention, spatially and temporally correlated masking, and EMA-teacher latent prediction regularized with VISReg. The model is pretrained on approximately 18,546 retained hours and evaluated on nine EEG datasets using frozen representations with k-nearest-neighbor and MLP readouts. It achieves the highest reported balanced accuracy on eight of nine datasets under each protocol, although the exceptions differ between protocols. Ablations vary the prediction target, geometric encoding, VISReg, and masking strategy. The principal unresolved issues are how much the ranking depends on readout conventions and whether the reported dataset partitions establish the claimed unit of generalization.

Critical strengths
  1. Frozen-representation evaluation across nine datasets with two readouts. The nine datasets cover six task families, and both kNN and trained MLP readouts are considered. Agreement across these different readouts is more informative than relying exclusively on full fine-tuning. The model does not dominate every dataset, and the paper reports those exceptions.

  2. Useful ablations of the main design choices. Table 3 compares latent and waveform targets, removes geometric encoding and VISReg, and evaluates random, block, channel, and temporal masking alternatives. For the five-task aggregate, the full model reaches 52.9 kNN balanced accuracy versus 42.9 with random masking and 49.7 without geometric encoding; the MLP comparisons also favor the full recipe. These are meaningful controls, and the proposed mask is computationally well specified through its spatial and temporal covariance construction.

Critical weaknesses
  1. The readout protocol leaves a consequential ambiguity in the representation-quality comparison. Appendix C applies Euclidean kNN without feature standardization or L2 normalization, so distances reflect each model's coordinate scaling and anisotropy as well as label information. The MLP protocol flattens model-specific features and sets hidden dimensions relative to their size, which does not hold classifier capacity constant across encoders. This does not make the reported pipeline scores incorrect, but weakens an interpretation as a comparison of intrinsic representation quality. Table 2 illustrates why a sensitivity check matters: on Mumtaz, CBraMod has 53.34% balanced accuracy despite 97.37% AUROC and 97.01% AUPR. That discrepancy calls for checking thresholding, calibration, and optimization before interpreting the balanced-accuracy gap as poor features. Please provide a compact, validation-selected normalization/readout sensitivity analysis, including a genuine linear probe or a matched-capacity classifier. The aim is to establish that the main ranking survives reasonable readout choices, not to optimize every baseline exhaustively.

  2. The dataset partitions and uncertainty reporting do not yet make the generalization claims auditable. Table 5 lists Mumtaz training/validation/test counts of 42 MDD + 40 controls, 10 + 8, and 10 + 10, totaling 120 units, without identifying whether these are people or recordings. The linked public dataset includes multiple recording conditions, and its BIDS conversion lists 64 participants. These facts can be consistent if the table counts recordings, but then person-level grouping must be made explicit. Please provide a file-to-participant-to-split manifest establishing that recordings from one person remain together. Separately, SEED-V and SEED-VII use trial splits within sessions, whereas other tasks use subject splits; the results should not be summarized as uniformly cross-subject. Finally, kNN standard deviations over K choices describe hyperparameter sensitivity, not uncertainty over subjects or training runs. These distinctions matter especially for datasets with only two held-out test participants. A participant-level analysis and correctly labeled variability would make the evidence much more interpretable.

Questions for authors
  1. What exactly do the Mumtaz split counts in Table 5 count? Can a manifest demonstrate participant-disjoint assignment across eyes-open, eyes-closed, and any other included recordings? If a different dataset version or subset is used, please identify it precisely.

  2. Does the ranking persist with train-fitted feature normalization and a common linear or capacity-controlled readout? What explains the near-chance Mumtaz balanced accuracy but high AUROC/AUPR of CBraMod?

  3. What evidence shows that ST-GFM reduces the local predictive shortcuts motivating it, beyond producing spatially correlated masks? A matched local-predictability diagnostic or shuffled-coordinate mask control would be informative; alternatively, the explanation can be stated as a design hypothesis supported by downstream ablations.

9 comments
AuthorsOpenReview ↗

Response to W1 & Q2

W1.1 & Q2.1. Feature Standardization or L2 Normalization in KNN

We appreciate the reviewer's concern that feature scaling may affect KNN comparisons, particularly since VISReg regularizes the feature distribution. We therefore re-evaluated all models under two configurations:

  • (1) Train-fitted feature standardization followed by Euclidean distance.
  • (2) L2 normalization followed by cosine similarity.

All models use identical downstream splits and KNN settings (K=3,5,7,9,11). Detailed results are provided in Tables 1–2 below.

Feature normalization indeed affects several baselines, potentially due to differences in feature scales. Nevertheless, EEG-GPRL retains the highest mean balanced accuracy across nine datasets under both configurations, outperforming the other baselines. These results further support the robustness of EEG-GPRL's representation transferability across different normalization settings.

We appreciate your suggestion, which helps us establish a fairer and more reliable comparison across baseline. We have update the paper with the feature-normalized results.

W1.2 & Q2.2. Common Linear Probing

We agree that the backbone-dependent hidden dimensions of the original MLP readout may introduce differences in classifier capacity. To address this concern, we additionally evaluated all models using a common linear-probing protocol, freezing the pretrained encoders and replacing the nonlinear MLP heads with single-layer linear readout. The mean Bal. Acc. summarized below, with detailed results provided in Table 3 below.

Mean linear-probing balanced accuracy across nine datasets (%).

MethodBal. Acc.
LaBraM44.73
CBraMod49.45
CSBrain54.92
REVE57.26
CodeBrain57.56
EEG-GPRL (ours)61.44

EEG-GPRL achieves the highest mean balanced accuracy (61.44%), outperforming the other baselines. As shown in Table 3, EEG-GPRL achieves the highest or tied-highest performance on eight datasets and ranks second on PhysioNet-MI.

These results further demonstrate that EEG-GPRL's advantage persists with a simple linear readout, rather than depending on nonlinear MLP classifier capacity.

Q2.3. Results of CBraMod in Mumtaz2016

Thanks for pointing this out. We checked CBraMod's confusion matrices across five seeds (0 = HC, 1 = MDD. rows = true labels, columns = predictions):

[[607, 6],  [[608, 5],  [[608, 5],  [[612, 1],  [[612, 1],
 [538,60]]   [545,53]]   [551,47]]   [559,39]]   [580,18]]

CBraMod predicts most samples as Class 0 (HC), resulting in nearly 100% recall for Class 0 but very low recall for Class 1 (MDD). Since balanced accuracy averages the recall of both classes, Bal.Acc.=Recall0+Recall12,\mathrm{Bal.Acc.} = \frac{\mathrm{Recall}_0 + \mathrm{Recall}_1}{2}, its Bal. Acc. remains close to 50%. However, the high AUROC (97.37%) and AUPR (97.01%) suggest that cbramod can distinguish the two classes well, but the default threshold (0.5) is not suitable for CBraMod.

We therefore re-evaluated all models using thresholds selected on the validation set:

Bal. Acc.LaBraMCBraModCSBrainREVECodeBrainEEG-GPRL
Default (0.5)84.54±1.2253.34±1.0689.04±0.2293.81±1.5989.78±0.3995.55±2.66
Validset selected84.58±2.4680.86±1.2288.54±0.2193.81±0.7889.15±0.8995.51±1.07

With the validation-selected threshold, CBraMod's Bal. Acc. increases from 53.34% to 80.86%, indicating that threshold selection accounts for much of the original gap. EEG-GPRL maintains the highest balanced accuracy (95.51%) under the same protocol. We will update the binary-classification results using validation-selected thresholds and clarified the evaluation protocol in the revised manuscript.

AuthorsOpenReview ↗

Response to W2 & Q1

W2.1 & Q1. The dataset split for Mumtaz2016

Thank you for pointing out this ambiguity. The counts in Table 5 refer to EEG recordings, not unique participants. We use 120 recordings from 63 participants (30 HC and 33 MDD), excluding MDD S8 whose recordings are unavailable. Among these participants, seven recordings are missing (HC: S12 EC, S18 EC, S25 EO. MDD: S4 EC, S7 EO, S12 EC, S16 EC), while HC S15 has two EO recordings, resulting in 120 recordings in total.

We verified the participant-disjoint splits:

  • Train (21 HC + 22 MDD): HC S1–S2, S10–S28; MDD S1–S2, S10–S29.
  • Val (4 HC + 5 MDD): HC S3–S4, S29–S30; MDD S3, S30–S33.
  • Test (5 HC + 6 MDD): HC S5–S9; MDD S4–S7, S9, S34.

HC and MDD IDs refer to distinct participants. All available EO/EC recordings and their derived segments from each participant are assigned to the same split, with no participant overlap. We have added a detailed description of the Mumtaz2016 splitting protocol in the appendix.

File-to-participant-to-split manifest:

SplitHC subjectsMDD subjects
TrainS1–2,S10–28S1–2,S10–29
ValS3–4,S29–30S3,S30–33
TestS5–9S4–7,S9,S34

Each listed subject has EC.edf and EO.edf, except:

  • Missing: HC (S12-EC,S18-EC,S25-EO) and MDD (S4-EC,S7-EO,S12-EC,S16-EC).
  • For HC S15 has 6921143_H S15 EO.edf, 6921959_H S15 EO.edf and H S15 EC.edf.

W2.2. Trial Splits within Sessions for SEED-V / SEED-VII

We agree with the reviewer that SEED-V and SEED-VII use trial-level splits within sessions, whereas the other datasets primarily use subject-disjoint splits. These protocols evaluate different generalization settings and should not be interpreted uniformly as cross-subject generalization.

Our primary goal is to assess representation transferability on downstream tasks, rather than exclusively cross-subject generalization. All pretrained encoders remain frozen during downstream evaluation, and identical splits are used for all compared models within each dataset.

To avoid potential ambiguity, we will provide a more detailed description of the dataset-specific splitting protocols in the revised paper and explicitly distinguish trial-level and subject-level evaluation settings when interpreting generalization results.

W2.3. KNN Variability and Participant-Level Uncertainty

We appreciate the reviewer's suggestion regarding uncertainty reporting. The original KNN standard deviations quantify sensitivity across K choices, rather than inter-participant or run-to-run uncertainty. We conducted a participant-level KNN analysis on six subject-disjoint datasets using L2-normalized features and cosine similarity. We report the mean and standard deviation of balanced accuracy across test participants.

Participant-level KNN balanced accuracy (%).

DatasetParticipantsBal. Acc.Individual Scores / Range
BCICIV-2a233.91±2.7331.98, 35.83
CHB-MIT279.37±12.7788.40, 70.34
MentalArithmetic465.21±14.7548.06, 60.00, 70.00, 82.78
ISRUC1060.00±8.9241.96–70.24
PhysioNet-MI2030.53±5.3421.63–40.98
HMC2647.69±10.3628.60–70.67

These results reveal varying degrees of inter-participant variability. We acknowledge that datasets with few test participants cannot provide reliable estimates of population-level uncertainty. Nevertheless, this analysis makes participant-level performance differences explicit rather than relying solely on aggregate metrics.

We will include these results, along with participant-level analyses for the remaining K values, in the revised appendix and clarify the interpretation of the reported standard deviations in the Experimental Setup.

AuthorsOpenReview ↗

Additional Experimental Results

Table 1. KNN-probing with feature standardization and Euclidean distance (%).

BCICIV-2aSEED-VMentalArithmetic
MethodBal. Acc.kappaW-F1Bal. Acc.kappaW-F1Bal. Acc.AUC-PRAUROC
LaBraM30.07±0.766.76±1.0129.70±0.7022.60±0.073.41±0.0822.57±0.4355.71±2.9244.25±5.4757.34±4.26
CBraMod32.12±1.619.49±2.1431.34±1.3721.52±0.142.12±0.2121.69±0.1754.10±1.2435.45±5.2956.52±2.42
CSBrain27.67±1.433.56±1.9126.76±1.2120.75±0.161.03±0.2419.72±0.4452.36±1.4632.70±1.4353.79±1.79
REVE33.65±0.5911.53±0.7926.68±1.3421.15±0.091.67±0.1320.60±0.5450.62±1.8125.93±1.4753.06±2.80
CodeBrain28.94±0.735.25±0.9722.06±1.3126.02±0.507.75±0.6726.65±0.4154.38±1.7339.67±1.1058.39±3.23
EEG-GPRL (ours)35.75±2.3314.33±3.1035.06±2.2429.31±0.6311.66±0.8529.75±0.6459.24±0.6347.71±4.7461.86±3.22
Physio-MIMumtaz2016SEED-VII
MethodBal. Acc.kappaW-F1Bal. Acc.AUC-PRAUROCBal. Acc.kappaW-F1
LaBraM28.77±1.225.02±1.6228.73±1.2480.36±0.2688.47±2.0889.33±2.4620.27±0.327.20±0.4119.50±0.11
CBraMod32.74±1.4510.32±1.9331.10±1.4363.27±0.5465.30±3.4165.32±0.8217.55±0.243.94±0.3016.63±0.18
CSBrain28.58±0.454.77±0.6027.97±0.5376.50±0.4574.56±1.8880.24±1.0517.05±0.083.37±0.1015.88±0.38
REVE29.59±0.626.13±0.8329.20±0.6066.32±0.4578.64±0.9572.84±2.6523.50±0.4811.17±0.3123.77±0.28
CodeBrain30.63±1.437.51±1.9130.39±1.4380.79±1.0684.98±1.5286.10±1.3324.19±0.6712.31±0.5424.02±0.53
EEG-GPRL (ours)30.75±1.127.67±1.5030.72±1.2384.36±0.2589.68±0.3890.69±0.9825.57±0.2713.41±0.3625.91±0.29
ISRUCCHB-MITHMC
MethodBal. Acc.kappaW-F1Bal. Acc.AUC-PRAUROCBal. Acc.kappaW-F1
LaBraM26.20±0.976.48±1.3017.82±1.7561.77±0.846.30±2.5966.52±0.6940.62±1.9724.26±0.9935.27±0.63
CBraMod32.47±0.1418.30±0.2932.92±0.2770.58±1.1030.11±2.5880.94±1.7225.98±0.396.35±0.5620.40±1.01
CSBrain22.63±0.754.06±0.8619.88±1.4871.39±1.2226.88±2.5676.52±1.3543.50±0.2225.41±0.1435.33±0.60
REVE52.64±1.1437.94±1.3452.26±0.9550.55±0.0922.76±3.2351.24±0.3743.68±0.7028.43±0.9643.96±0.76
CodeBrain54.34±0.5842.29±1.7353.41±0.8173.64±0.5121.17±6.1178.63±0.7543.34±0.6930.10±1.0445.74±0.55
EEG-GPRL (ours)56.37±1.4346.76±2.5959.85±1.8574.80±0.6333.89±0.5984.12±2.1145.89±0.7830.68±1.2446.67±1.10
Reviewer_zsXQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper identifies three main limitations of current EEG pretraining paradigms: low-level waveform reconstruction, insufficient integration of electrode geometry, and information leakage caused by current masking strategies. To address these challenges, the authors propose EEG-GPRL, a novel framework. To enhance geometric representation, EEG-GPRL applies Rotary Positional Embedding (RoPE) to 3D coordinates to create relative positional encodings. Furthermore, it introduces a masking strategy that masks highly correlated segments to prevent local information leakage, and employs latent-space predictive pretraining with VISReg instead of raw masked reconstruction. Experimental results demonstrate that EEG-GPRL significantly outperforms five other methods across almost all datasets using both KNN and linear classifiers. The experimental result is promising, and my evaluation may change depending on the authors’ response during the rebuttal.

Critical strengths
  • The proposed masking strategy is novel; by masking highly correlated segments, EEG-GPRL minimizes local information leakage.
  • The proposed method demonstrates promising downstream performance, outperforming other methods.
  • All core components are ablated in detail to verify their advantages in terms of downstream performance.
Critical weaknesses
  • The geometry-aware module and latent-space prediction are not new to the EEG field, making the proposed method somewhat incremental.
  • The proposed method is validated solely in terms of performance; there is a lack of further evidence explaining why the proposed components work. For example, while Table 3 shows that applying Geo-Attn increases performance, it would be beneficial to see if the correlation or attention scores of specific channels increase after its application. This would provide a much stronger explanation for its effectiveness.
  • The three main components proposed by the authors appear to be independent of each other. However, the ablation study (Table 3) shows that removing just one component causes performance to drop significantly across both classifiers, often falling below the baseline. Because of this, it is unclear which component actually drives the improvements. Why are all three blocks strictly necessary, and why does addressing only two out of the three problems fail to have any performance gain?
  • Minor: The authors mention VISReg in the Introduction and the first part of the methodology, but the citation for it does not appear until Section 3.3.
Questions for authors

See the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to W1 & W2

W1. Clarification of Methodological Novelty

We agree that latent prediction and geometric modeling have been explored individually. However, our contribution lies in a unified geometry-aware design for latent predictive learning. EEG-GPRL incorporates electrode geometry into two fundamental aspects of latent predictive learning: how contextual information is selected and how inter-channel information is integrated.

(1) Geometry-aware feature interactions. Unlike approaches that primarily encode electrode geometry through input embeddings, GeoAttn introduces continuous 3D relative electrode displacement directly into q-k interactions, explicitly modulating content-dependent channel attention with relative geometry.

(2) Geometry-aware predictive context. Unlike predefined block or topology-based masking patterns, ST-GFM samples a continuous spatiotemporal Gaussian field whose covariance depends on electrode distances and temporal proximity. This generates correlated masks that reduce local information leakage and encourage prediction from broader neural context.

(3) Complementary design for EEG predictive learning. ST-GFM determines which contextual information is hidden, while GeoAttn governs how the remaining channel information is integrated for latent prediction. Together, they incorporate EEG's spatial structure into both predictive context construction and representation learning.

Importantly, our additional analyses provide evidence beyond performance gains: GeoAttn increases attention to the eight nearest electrodes from 16.61% to 28.86%, while coordinate shuffling reduces mean KNN balanced accuracy by 8.43 percentage points. These results support the importance of meaningful electrode geometry rather than merely introducing coordinate-dependent transformations.

W2. Explaining why the Proposed Components Work

We thank the reviewer for this suggestion, which helps us better explain how GeoAttn affects EEG channel interactions. We have added a dedicated analysis section (Sec. 4.5) with attention visualizations (Fig. 4), quantitative analysis (Appendix G, Table 10), and a coordinate-shuffling experiment.

(1) Attention Visualization (Fig. 4). We compare attention with and without coordinate-aware rotation while keeping query/key features and learned projections fixed, using 64 segments from each of 21 Trujillo2017 participants across 12 layers and 8 heads. Figure 4 visualizes four representative electrodes (F5, F6, P5, P6). The difference maps show increased attention toward spatially neighboring electrodes, with the preferred regions shifting according to the query electrode.

(2) Quantitative Analysis (Appendix G, Table 10).

Metricw/o GeoAttnw/ GeoAttn
Attention allocated to N-nearest electrodes (N=8)0.16610.2886
Spearman ρ (distance vs. attention)-0.1858-0.4497
  • Attention to nearest 8 measures the proportion of non-self attention allocated to the eight nearest electrodes, which increases by 12.25 percentage points, indicating stronger local attention.
  • Spearman ρ measures the rank correlation between electrode distance and attention weight. Its stronger negative correlation indicates that GeoAttn assigns relatively less attention to distant electrodes, demonstrating more geometry-dependent channel interactions.

(3) Coordinate-Shuffling Experiment. We further disrupt the correspondence between electrodes and their coordinates to assess the importance of meaningful geometry.

KNN-Probe (Bal. Acc. %)BCIC-IV2aSEED-VMentalArith.MumtazISRUCAvg.
GeoAttn coord.-shuffled30.73±1.5423.93±0.3153.72±1.5176.13±1.3048.09±1.3546.52
Ours33.91±1.4430.11±0.6965.21±2.0685.04±0.3260.49±1.5754.95

Correct electrode coordinates improve mean balanced accuracy by 8.43 percentage points, confirming that meaningful geometric information contributes to the learned representations.

We also examined how ST-GFM affects local information leakage: the additional local prediction error is 0.2010, versus 0.0351 with shuffled coordinates and 0.0076 with random masking (Sec. 4.5; Fig. 5; Appendix H, Table 11).

Together, these analyses show how geometry shapes GeoAttn's channel interactions and ST-GFM's masked context. These findings complement the component ablations in Table 3 and help explain the roles of both designs in EEG-GPRL.

AuthorsOpenReview ↗

Response to W3 & W4

W3. Component Contributions

We thank the reviewer for raising this important question. The proposed components address distinct limitations but are complementary rather than strictly independent.

(1) Clarification of Table 3. We identified and corrected a labeling error in the second row: replacing latent prediction with raw EEG reconstruction using a linear decoder also removes VISReg. The correct configuration is (Lat.Pred. ×, GeoAttn ✓, VISReg ×, ST-GFM ✓). VISReg regularizes latent representations to stabilize predictive learning and was therefore omitted from the waveform-reconstruction variant. This row evaluates the combined change in objective and regularization, rather than the independent contribution of latent prediction.

(2) Roles of each components. Latent prediction encourages contextual representation learning rather than low-level waveform reconstruction, while VISReg regularizes the latent feature distribution to mitigate representation collapse. GeoAttn further models relative electrode geometry in channel interactions, while ST-GFM reduces local information leakage during masked prediction. These designs respectively address what to predict, how to integrate channel information, and how to construct predictive contexts.

(3) Ablation interpretation. Removing one component does not necessarily eliminate the benefits of the remaining components. Replacing GeoAttn with vanilla attention still achieves 49.70%/59.97% mean balanced accuracy under KNN/frozen probing, compared with 54.95%/67.06% for the full model. Similarly, replacing ST-GFM with block masking yields 46.50%/60.78%. These results show that the remaining components still learn transferable representations, while their combination achieves stronger performance. The purpose of these ablations is to isolate each component's contribution within EEG-GPRL, rather than to establish that every partial configuration must outperform independently developed foundation models.

Thus, we do not claim that all components are strictly necessary or that addressing only two limitations yields no benefit. Rather, Table 3 demonstrates their complementary contributions: latent prediction and VISReg establish the representation-learning framework, while GeoAttn and ST-GFM improve geometric interactions and predictive context construction.

W4. Minor

Thank you for pointing this out. We have added the VISReg citation at its first mention in the Introduction and ensure consistent attribution throughout the manuscript.

Reviewer_bXYEOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EEG-GPRL, which learns transferable EEG representations by predicting latent representations from an EMA teacher, combining GeoAttn, ST-GFM, and VISReg. The model is pretrained on approximately 18,500 hours of EEG, and its frozen representations are evaluated across six EEG task categories.

Critical strengths

s1. The method targets the quality of frozen EEG foundation model representations through a well-motivated combination of latent prediction, electrode geometry modeling, and spatiotemporally correlated masking. s2. The model achieves the highest balanced accuracy on 8/9 datasets under each probing protocol, while component ablations, comparisons of masking strategies, and hyperparameter sensitivity analyses provide reasonably comprehensive empirical evidence.

Critical weaknesses

w1. Although the paper discusses EEG-JEPA and related work, it lacks direct comparisons with the closest structured latent-prediction methods, leaving the relative benefits of the new designs insufficiently established. w2. The sensitivity of the evaluation results to feature normalization and the split definitions for some datasets require further clarification.

Questions for authors

q1. EEG-JEPA, cited in Section 2, also uses latent prediction and structured electrode–time masking. Could the authors clarify the specific contributions of GeoAttn and ST-GFM beyond this approach? q2. In Appendix C, KNN uses Euclidean distance without feature standardization or normalization, while VISReg constrains the feature distribution. Could the authors provide results with standardized features or cosine similarity to assess whether the model rankings remain stable? q3. The train/val/test counts for Mumtaz2016 in Table 5 total 120; do these count unique subjects or EEG recordings? Are eyes-open and eyes-closed recordings from the same subject, along with all their segments, always assigned to the same split?

4 comments
AuthorsOpenReview ↗

Response to W1 (Q1)

W1 (Q1) Differences from EEG-JEPA

We thank the reviewer for highlighting this closely related work. EEG-JEPA and EEG-GPRL share the latent predictive learning principle, but differ in their geometric modeling, masking strategy, and representation regularization.

(1) Geometry-aware attention. EEG-JEPA adopts a CBraMod-based backbone, with electrode topology primarily incorporated through structured target selection. In contrast, our GeoAttn explicitly integrates continuous 3D relative electrode coordinates into channel-wise query-key interactions via coordinate-aware rotary embeddings (Eq. 4–5). Thus, geometry directly modulates attention rather than only determining which tokens are masked.

(2) Spatiotemporal masking. EEG-JEPA's N-MET combines six predefined masking primitives (temporal/channel stripes, local blocks, topology pairs, and channel dropout). In contrast, ST-GFM samples a continuous Gaussian random field with spatial and temporal covariance determined by electrode distances and temporal separation (Eq. 6–7). This generates correlated masks without relying on predefined masking patterns or anatomical groups, naturally adapting to different electrode configurations.

(3) Representation learning. EEG-JEPA emphasizes multi-depth latent supervision with variance-covariance regularization. EEG-GPRL instead combines latent prediction with GeoAttn, ST-GFM, and VISReg, which regularizes feature scale, center, and distributional shape.

Our contributions therefore lie not in introducing JEPA to EEG, but in explicitly modeling relative electrode geometry within attention and using covariance-driven masking to reduce local information leakage. We will expand Section 2 to clarify these distinctions.

W1 (Q1) Comparison with the structured latent-prediction methods (REVE-JEPA)

EEG-JEPA is a concurrent work released on arXiv within 2 months prior to our submission, and its pretrained checkpoints are not open available. Moreover, its benchmarks and evaluation protocols differ from ours, preventing a direct comparison under identical settings. We therefore regret that a direct performance comparison with EEG-JEPA is not feasible at this stage.

Nevertheless, we fully agree that comparisons with structured latent-prediction methods are essential. We therefore additionally evaluated the publicly released JEPA model of Guetschel et al. [1], based on the REVE architecture, hereafter referred to as REVE-JEPA.

Using the same downstream splits and evaluation protocols, EEG-GPRL outperforms REVE-JEPA by 6.02, 5.97, and 3.66 percentage points in mean balanced accuracy across nine datasets under KNN (standardized Euclidean), KNN (cosine), and frozen probing, respectively. Detailed results are provided in Tables 1–3 below.

These additional comparisons strengthen the evidence for EEG-GPRL's representation transferability. We have incorporated the results and a more detailed discussion into the revised manuscript.

[1] Guetschel et al. What Masking Geometry Works Best for EEG Foundation Models? arXiv:2609.33487, 2026.

[2] Li et al. EEG-JEPA: Structured Latent Prediction for EEG Foundation Models. arXiv:2608.00114, 2026.

AuthorsOpenReview ↗

Response to W2 (Q2 & Q3)

W2 (Q2). Feature Standardization and Normalization in KNN

We appreciate the reviewer's concern that VISReg may influence KNN performance through feature distributions. To examine this, we re-evaluated all models under two additional configurations:

  • (1) Train-fitted feature standardization followed by Euclidean distance.
  • (2) L2 normalization followed by cosine similarity.

All models use identical downstream splits and KNN settings (K=3,5,7,9,11). Detailed results are provided in Tables 1–2 below.

Feature normalization indeed affects the performance of several baselines, potentially due to differences in feature scales, whereas EEG-GPRL's feature distribution is regularized by VISReg. Although normalization changes the relative rankings of several baselines, EEG-GPRL retains the highest mean balanced accuracy across nine datasets under both configurations, outperforming the other baselines.

These results further support the robustness of EEG-GPRL's representation transferability across different normalization settings.

Mean KNN-probing balanced accuracy across nine datasets (%).

MethodStd.+EuclideanL2+Cosine
LaBraM40.7146.10
CBraMod38.9342.77
CSBrain40.0544.97
REVE41.3043.79
CodeBrain46.2546.55
REVE-JEPA43.1045.17
EEG-GPRL (ours)49.1251.13

We appreciate your suggestion, which helps us establish a fairer and more reliable comparison across baseline. We have update the manuscript accordingly.

W3 (Q3). Dataset Split for Mumtaz2016

Thank you for pointing out this ambiguity. The counts in Table 5 refer to EEG recordings, not unique participants. We use 120 recordings from 63 participants (30 HC and 33 MDD), excluding MDD S8 whose recordings are unavailable. Among these participants, seven recordings are missing (HC: S12 EC, S18 EC, S25 EO. MDD: S4 EC, S7 EO, S12 EC, S16 EC), while HC S15 has two EO recordings, resulting in 120 recordings in total.

We verified the participant-disjoint splits:

  • Train (21 HC + 22 MDD): HC S1–S2, S10–S28; MDD S1–S2, S10–S29.
  • Val (4 HC + 5 MDD): HC S3–S4, S29–S30; MDD S3, S30–S33.
  • Test (5 HC + 6 MDD): HC S5–S9; MDD S4–S7, S9, S34.

HC and MDD IDs refer to distinct participants. All available EO/EC recordings and their derived segments from each participant are assigned to the same split, with no participant overlap. We have added a detailed description of the Mumtaz2016 splitting protocol in the appendix.

AuthorsOpenReview ↗

Additional Experimental Results

Table 1. KNN-probing with feature standardization and Euclidean distance (%).

BCICIV-2aSEED-VMentalArithmetic
MethodBal. Acc.kappaW-F1Bal. Acc.kappaW-F1Bal. Acc.AUC-PRAUROC
LaBraM30.07±0.766.76±1.0129.70±0.7022.60±0.073.41±0.0822.57±0.4355.71±2.9244.25±5.4757.34±4.26
CBraMod32.12±1.619.49±2.1431.34±1.3721.52±0.142.12±0.2121.69±0.1754.10±1.2435.45±5.2956.52±2.42
CSBrain27.67±1.433.56±1.9126.76±1.2120.75±0.161.03±0.2419.72±0.4452.36±1.4632.70±1.4353.79±1.79
REVE33.65±0.5911.53±0.7926.68±1.3421.15±0.091.67±0.1320.60±0.5450.62±1.8125.93±1.4753.06±2.80
CodeBrain28.94±0.735.25±0.9722.06±1.3126.02±0.507.75±0.6726.65±0.4154.38±1.7339.67±1.1058.39±3.23
REVE-JEPA33.93±1.5411.91±2.0534.08±1.4526.99±0.358.96±0.5127.68±0.3154.03±1.3643.54±2.3064.37±3.06
EEG-GPRL (ours)35.75±2.3314.33±3.1035.06±2.2429.31±0.6311.66±0.8529.75±0.6459.24±0.6347.71±4.7461.86±3.22
Physio-MIMumtaz2016SEED-VII
MethodBal. Acc.kappaW-F1Bal. Acc.AUC-PRAUROCBal. Acc.kappaW-F1
LaBraM28.77±1.225.02±1.6228.73±1.2480.36±0.2688.47±2.0889.33±2.4620.27±0.327.20±0.4119.50±0.11
CBraMod32.74±1.4510.32±1.9331.10±1.4363.27±0.5465.30±3.4165.32±0.8217.55±0.243.94±0.3016.63±0.18
CSBrain28.58±0.454.77±0.6027.97±0.5376.50±0.4574.56±1.8880.24±1.0517.05±0.083.37±0.1015.88±0.38
REVE29.59±0.626.13±0.8329.20±0.6066.32±0.4578.64±0.9572.84±2.6523.50±0.4811.17±0.3123.77±0.28
CodeBrain30.63±1.437.51±1.9130.39±1.4380.79±1.0684.98±1.5286.10±1.3324.19±0.6712.31±0.5424.02±0.53
REVE-JEPA29.27±0.575.69±0.7629.26±0.5978.15±0.3483.95±0.1185.35±1.2523.88±0.6212.16±0.5923.93±0.66
EEG-GPRL (ours)30.75±1.127.67±1.5030.72±1.2384.36±0.2589.68±0.3890.69±0.9825.57±0.2713.41±0.3625.91±0.29
ISRUCCHB-MITHMC
MethodBal. Acc.kappaW-F1Bal. Acc.AUC-PRAUROCBal. Acc.kappaW-F1
LaBraM26.20±0.976.48±1.3017.82±1.7561.77±0.846.30±2.5966.52±0.6940.62±1.9724.26±0.9935.27±0.63
CBraMod32.47±0.1418.30±0.2932.92±0.2770.58±1.1030.11±2.5880.94±1.7225.98±0.396.35±0.5620.40±1.01
CSBrain22.63±0.754.06±0.8619.88±1.4871.39±1.2226.88±2.5676.52±1.3543.50±0.2225.41±0.1435.33±0.60
REVE52.64±1.1437.94±1.3452.26±0.9550.55±0.0922.76±3.2351.24±0.3743.68±0.7028.43±0.9643.96±0.76
CodeBrain54.34±0.5842.29±1.7353.41±0.8173.64±0.5121.17±6.1178.63±0.7543.34±0.6930.10±1.0445.74±0.55
REVE-JEPA45.84±0.1532.66±0.2645.84±0.1662.00±0.4514.58±0.4162.84±0.6933.78±0.4615.16±0.7329.57±0.95
EEG-GPRL (ours)56.37±1.4346.76±2.5959.85±1.8574.80±0.6333.89±0.5984.12±2.1145.89±0.7830.68±1.2446.67±1.10
AuthorsOpenReview ↗

Additional Experimental Results

Table 2. KNN-probing with L2-normalized features and cosine similarity (%).

BCICIV-2aSEED-VMentalArithmetic
MethodBal. Acc.kappaW-F1Bal. Acc.kappaW-F1Bal. Acc.AUC-PRAUROC
LaBraM30.49±0.337.31±0.4430.43±0.2923.71±0.184.85±0.2323.96±0.2960.90±1.2442.49±4.1266.03±1.89
CBraMod32.40±1.739.86±2.3131.69±1.6522.20±0.173.04±0.2522.58±0.2255.21±1.8041.12±1.7957.66±1.85
CSBrain29.32±1.025.76±1.3628.80±0.7821.15±0.171.55±0.2520.75±0.3954.86±1.3932.66±3.2856.54±2.78
REVE31.81±0.689.07±0.9123.79±1.3321.33±0.151.91±0.2120.88±0.5056.53±6.2428.26±0.6757.55±2.58
CodeBrain27.22±1.322.96±1.7521.27±1.7626.23±0.538.00±0.7326.86±0.4557.61±3.3538.65±0.8069.28±4.30
REVE-JEPA33.42±1.2511.23±1.6730.53±0.6328.48±0.4310.83±0.6229.22±0.4361.67±2.9044.46±3.7164.38±3.67
EEG-GPRL (ours)33.91±1.4411.88±1.9232.13±0.8830.11±0.6912.76±0.9330.74±0.7165.21±2.0655.94±5.1670.81±2.29
Physio-MIMumtaz2016SEED-VII
MethodBal. Acc.kappaW-F1Bal. Acc.AUC-PRAUROCBal. Acc.kappaW-F1
LaBraM30.11±0.746.81±0.9830.12±0.7380.02±0.1788.84±1.8889.74±2.4523.83±0.6211.56±0.7623.48±0.35
CBraMod34.26±1.0712.35±1.4234.28±1.0778.30±0.8380.60±0.4283.24±1.3218.96±0.155.62±0.2018.68±0.12
CSBrain29.61±0.956.13±1.2729.51±0.9777.72±0.7375.98±1.5782.40±1.0818.43±0.145.07±0.2017.77±0.20
REVE30.12±0.546.83±0.7230.01±0.5368.86±0.4582.20±1.5879.18±2.9726.01±0.7214.54±0.5626.24±0.63
CodeBrain30.72±1.367.62±1.8130.51±1.3382.59±1.9086.77±1.0888.49±1.1025.94±0.7814.09±0.9426.19±0.66
REVE-JEPA30.37±0.697.17±0.9230.25±0.6977.60±0.5284.03±0.0885.19±1.0326.48±0.5414.58±0.6626.80±0.51
EEG-GPRL (ours)30.80±0.977.73±1.2931.44±0.8585.04±0.3290.73±0.5091.35±1.0627.49±0.6115.72±0.7428.00±0.64
ISRUCCHB-MITHMC
MethodBal. Acc.kappaW-F1Bal. Acc.AUC-PRAUROCBal. Acc.kappaW-F1
LaBraM56.95±1.7549.31±3.0459.52±2.0463.43±0.417.45±4.0165.53±0.5245.45±2.0130.47±3.0045.38±2.75
CBraMod38.73±0.1027.91±0.5742.55±0.0872.35±0.6635.70±1.2684.10±1.2132.56±0.0814.76±0.1529.75±0.43
CSBrain53.96±1.3642.60±1.9654.45±1.4375.38±0.6434.79±0.9583.51±1.6044.29±1.4729.96±2.0243.43±1.28
REVE57.15±1.4244.96±1.9658.11±1.5057.18±1.6929.84±1.9857.92±1.9545.10±0.6830.28±0.9445.31±0.65
CodeBrain53.65±0.5745.17±1.6754.22±0.6569.25±0.3117.08±6.6575.37±0.4945.76±0.5131.74±0.9845.57±0.52
REVE-JEPA50.07±0.5438.00±0.9650.14±0.6962.44±0.8716.46±0.3362.41±1.1235.97±0.3817.98±0.6232.50±0.82
EEG-GPRL (ours)60.49±1.5751.52±2.6762.29±2.0078.81±0.5836.12±0.3685.96±1.1248.34±1.0334.07±1.5849.10±1.22

Table 3. Frozen-probing comparison with REVE-JEPA (%).

BCICIV-2aSEED-VMentalArithmetic
MethodBal.Acc.KappaW-F1Bal.Acc.KappaW-F1Bal.Acc.AUC-PRAUROC
REVE-JEPA51.52±0.9535.36±1.2750.22±1.2133.44±0.4015.85±0.6632.07±0.7859.03±3.0363.35±2.2878.85±1.79
EEG-GPRL52.00±1.2736.00±1.6950.49±1.7836.73±0.3520.65±0.5836.43±0.6375.69±2.8178.40±3.0188.74±2.26
Physio-MIMumtaz2016SEED-VII
MethodBal.Acc.KappaW-F1Bal.Acc.AUC-PRAUROCBal.Acc.KappaW-F1
REVE-JEPA53.97±0.7438.62±0.9153.55±0.6990.64±1.1297.38±0.1497.01±0.1027.63±0.5815.71±0.8327.21±0.62
EEG-GPRL54.69±0.7539.58±1.0054.57±0.7195.51±1.0799.28±0.0799.32±0.0831.05±0.6419.71±0.6431.19±0.73
ISRUCCHB-MITHMC
MethodBal.Acc.KappaW-F1Bal.Acc.AUC-PRAUROCBal.Acc.KappaW-F1
REVE-JEPA73.42±0.6066.97±1.1174.16±0.7277.71±1.9928.64±3.0189.38±1.2167.90±0.8157.50±1.3366.29±1.29
EEG-GPRL75.35±0.1568.99±0.7075.27±0.4980.40±1.4541.49±6.9590.54±0.8166.75±0.7758.81±0.2166.37±0.12

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.