We sincerely thank you for your insightful suggestions. In the revision, we clarify the design rationale of EEG covariance alignment and dynamics modelling. We add experiments of supervised contrastive learning, batch construction methods (temporal alignment vs. random) and batch-size analyses. Besides, we show DEAP to other datasets transfer obtained comparable performance with SEED to others, supporting cross-rating-paradigm generalization capability of our model.

Q1. The motivation behind the model designs.

Our work aims to develop a unified framework for multi-dataset EEG training and cross-dataset generalization. All components of our design are centered around this objective. Specifically, we introduce two complementary alignment losses: the Covariance Distribution Alignment (CDA) loss and the Inter-Subject Alignment (ISA) loss. The CDA loss aligns global second-order statistics (i.e., inter-channel covariance structures) across subjects and datasets. This captures inter-region coupling (or functional connectivity) patterns in EEG—features that have been increasingly emphasized in recent neuroscience and EEG decoding literature [1-3]. While MMD with a Gaussian kernel can capture higher-order statistics, there is currently limited evidence in the EEG literature suggesting the functional relevance of such statistics. Moreover, MMD introduces sensitivity to kernel bandwidth and other hyperparameters. Therefore, we prioritize CDA loss as a simple, interpretable, and stable alignment objective in our framework.

For the model architecture, we aim to design a physiologically plausible and computationally efficient EEG encoder. We were not fully satisfied with standard Transformer architectures due to their limited interpretability and relatively large parameter counts. As such, we propose a tailored architecture consisting of an MLLA-based Channel Encoder and a Spatiotemporal Dynamics Module. The former captures rich temporal dynamics from each EEG channel, while the latter integrates information across channels and time, enabling the model to learn coordinated patterns of brain activity.

The MLLA component is particularly suited for modeling long-range dependencies in EEG due to its combination of an input gate, linear attention, and a forget gate. We apply overlapping strided patches (see Fig. 2D), where each patch is first linearly projected—functionally similar to strided convolution—to extract local features. These are then processed by linear attention for global dependency modeling, followed by a forget gate to filter salient patterns. In contrast, standard Transformers use the QKᵀV mechanism to assign weights to input tokens by computing dot products between queries and keys, which, in our view, lacks physiological priors to effectively capture EEG-specific characteristics. The Spatiotemporal Dynamics Module further captures inter-channel dynamics and uses local attention to emphasize important spatial-temporal patterns.

While model architecture is not the main contribution of this work, we designed our encoder based on domain-specific understanding of EEG data and validated its effectiveness via lesion studies. We believe this principled design—though not exhaustively benchmarked against all alternatives—should not undermine the merit of our proposed framework. In the revised version, we will further explain the motivation of our framework design and discuss its relashionship with other methods.

[1] Zhang Y, Chen Z S. Harnessing electroencephalography connectomes for cognitive and clinical neuroscience[J]. Nature Biomedical Engineering, 2025: 1-16.
[2] Li C, Tang T, Pan Y, et al. An efficient graph learning system for emotion recognition inspired by the cognitive prior graph of eeg brain network[J]. IEEE Transactions on Neural Networks and Learning Systems, 2024, 36(4): 7130-7144.
[3] Wang Y, Qiu S, Ma X, et al. A prototype-based SPD matrix network for domain adaptation EEG emotion recognition[J]. Pattern Recognition, 2021, 110: 107626.

Q2. The comparison with other self-supervised frameworks using trial information.

To demonstrate the effectiveness of our self-supervised strategy over others using trial information, we compared it with Supervised contrastive learning (supervised CL), which regards samples with the same emotion labels as positive pairs and those with different emotion labels as negative pairs in each dataset. It directly optimizes based on emotion labels (but agnostic to specific emotion categories). We use SupCon loss [4] and keep all other settings the same. Our strategy outperforms supervised contrastive learning by a large margin, indicating the effectiveness of our ISA loss for EEG alignment.

[4] Khosla P, Teterwak P, Wang C, et al. Supervised contrastive learning[J]. Advances in neural information processing systems, 2020, 33: 18661-18673.

Q3. The effects of training batch selection in pre-training.

Indeed, the trianing batch selection is carefully designed to learn subject-invariant representations. Sampling two temporally aligned samples as the positive pair is critical. In the preliminary experiment, we found that when the two positive samples are randomly selected from the same trial without temporal alignment, the model has a near random top1 accuracy in contrastive learning. This is possibly due to that temporally aligned samples carry more shared information, which facilitates the model to focus on meaningful shared representations across subjects and discard subject-specific noise. We will add these comparison experiments in the revised version.

We also experiment with larger batch sizes in pre-training. As previous work in contrastive learning, such as SimCLR, demonstrated that contrastive learning benefits from a larger batch size, we increase our batch size to double: Obtain two samples from each trial in one batch (one sample from each trial was obtained in the previous version). Samples that are not temporally aligned across subjects are regarded as negative samples.

We found doubling the batch size increases the leave-one-dataset-out few-subject fune-tuning accuracy on the SEED-V dataset by 2.08% and by 2.27%. This demonstrates that larger batch size also benefit EEG contrastive learning for learning effective representations. However, doubling the batch size also increases the occupation of GPU memory from approximately 20G to 40G in joint pre-training, which poses a higher demand of the hardware. Future work is needed to more comprehensively investigate the effect of batch size on downstream performance.

Q4.1 Transfer from the DEAP dataset to SEED datasets.

We conducted experiments to train the model on the DEAP dataset and transfer it to other datasets. In the table, we compare the performance of transfer from DEAP with transfer from SEED.

The performance of transfer from DEAP is slightly lower than transfer from SEED to other SEED-series datasets, but the performance gap is within 2.3%. Besides, the accuracy of transfer from DEAP to FACED is higher than transfer from SEED to FACED by 3.9%. These results demonstrate the effectiveness of our framework in transferring between datasets with discrepancies in emotion label paradigms. It aligns with our expectation that the model can learn generalizable emotion representations through contrastive learning, which do not rely on emotion label information.

Q4.2 The issue of individual differences in valence/arousal experience on the DEAP dataset.

We agree that inter-subject variability is a fundamental challenge in affective EEG analysis. Such variability arises from two sources:
(1) Differences in emotional perception (i.e., subjects experiencing different emotions in response to the same stimulus)
(2) Differences in neural signal properties that are unrelated to emotion (e.g., due to montage, scalp geometry, or baseline spectral patterns)

In this work, we focus on mitigating the second type—signal-level inter-subject variability, which poses a major obstacle to generalization across subjects and datasets. Modeling both types of variability simultaneously would significantly complicate the training process and potentially reduce robustness.

To address this, mdJPT leverages temporally aligned EEG segments as a weak prior for capturing shared neural patterns. Although such alignment does not ensure identical emotional states across subjects, it increases the likelihood of emotion similarity compared to unaligned segments. This provides a practical and effective anchor for extracting robust emotion-related representations.

We acknowledge the importance of accounting for subjective emotional differences, and we plan to incorporate more individualized modeling strategies (e.g., soft contrastive labels based on individualized emotion ratings) in future work.

We further add the formal comparison results with positive samples from the same trial but not strictly temporally aligned, as denoted in the response to Q3.

Without temporal alignment, the model performance drops significantly. Under the not aligned setting, the model achieves only 55.47% accuracy and 80.97% AUROC. This comparison validates the importance of using temporally aligned positive pairs in contrastive learning for EEG data.

Thanks for your effort, not bad results, but this is not what I interest.

We sincerely thank you for your suggestions to make the results more convincing. In revision, we (i) add zero‑shot evaluations on unseen datasets, where mdJPT outperforms all foundation‑model baselines and broaden evidence with DEAP arousal, dominance and VA‑quadrant prediction. Interpretability results, loss curves and hyperparameter evaluations are further added. The related work and the discussion about the results are revised. These updates more convincingly, and more comprehensively, validate the model’s generalization across datasets and label regimes.

Q1. The issue of model complexity due to multiple components.

Although our model integrates multiple components, we have paid careful attention to efficiency during design. The total number of parameters remains relatively small (1.0M) compared to existing EEG foundation models (e.g., LaBram: 5.8M, EEGPT: 4.7M), as shown in Table S2. In addition, our architecture achieves faster inference speed, owing to the parameter-efficient designs of the MLLA encoder and spatiotemporal dynamics module. We believe this strikes a favorable balance between architectural expressiveness and practical usability.

Q2. The issue of fine-tuning on the target dataset and the evaluation in zero-shot settings.

To evaluate our method on real zero-shot generalization tasks, we further conducted an experiment to directly apply the pre-trained model to a new dataset without fine-tuning. We compute cosine similarity across samples and define zero‑shot accuracy by whether the top‑similar pair shares the same label. Samples from the same trial are excluded to avoid leakage. The results with SEED-V as the target dataset are shown below.

Table: Comparison of different models under zero-shot settings.

mdJPT outperforms DE baseline in the zero‑shot setting. It improves by 9.6% over the DE baseline on the SEED-V dataset. On the FACED and DEAP datasets, mdJPT also outperforms the DE baseline by 8.5% and 17.9%, respectively. We will add the results of other comparison models in the discussion session.

To our knowledge, this is the first zero‑shot cross‑dataset EEG emotion experiment. Prior work typically fine‑tunes per subject. These results highlight the robustness and generalizability of our framework.

Q3. Present concrete case studies or examples of how the model performs across varied datasets

We identify the most critical features for emotion classification on multiple datasets and visualize its related temporal patterns and spatial transition patterns through the integrated gradient method. The feature importances for an emotion category are correlated between different datasets (as shown in Fig. S3), indicating a shared feature space is learned across datasets. We will further include the visualization of temporal and spatial patterns of critical features in the revised version.

Q4. Introduce important EEG foundation models in Related Work.

To present a more comprehensive review of the current foundation models, we add the introduction of BrainBERT, Brant, Brant-X, Brant-2, and CBraMod in the Related Work.

"BrainBERT implemented a masked auto-encoder on SEEG time frequency representations and improved the performance of speech decoding."

"Brant-X aligns EEG with EXG (EOG/ECG/EMG) signals through contrastive learning. It pulls the representations of simultaneous EEG and EXG patches together and pushes apart those of different time segments. Brant-2 combines EEG and SEEG data from more than 15k subjects, uses temporal and frequency features for mask-prediction and forecasting in pre-training, and demonstrates its strength in downstream tasks of seizure detection/prediction, sleep stage classification, motor movement/imagery, and emotion recognition. CBraMod designed a criss-cross attention mechanism to process temporal and spatial information separately. It used more than 27,000 hours of EEG data for pre-training and was evaluated on 10 downstream tasks. These methods demonstrate the current trend of merging larger datasets for pre-training and combining the information from multiple modalities."

Q5. The prediction of dimensions other than valence on the DEAP dataset.

In the previous submission, we prioritized a compact cross‑dataset analysis and therefore focused on valence, the dimension most commonly evaluated in previous EEG emotion studies. To evaluate the model more comprehensively, we now add the binary prediction of high/low arousal and high/low dominance, and a four-category classification of the valence-arousal quadrants. Our model outperforms the baseline model by a large margin, demonstrating the generalization ability of the pre-trained model to new emotion rating paradigms (emotion dimensions rather than discrete categories) and its adaptability across different affective dimensions.

Q6. Whether training for 20 epochs leads to overfitting? Can the loss curves be added?

Specifically, we pre-train the model for 20 epochs. As the data size in pre-training is relatively large, this does not lead to overfitting. The training and validation loss generally converges after 20 epochs. In fine-tuning, the classifier is trained with early stopping. The loss curves generally converge in a few epochs and exhibit small fluctuations since then. We will also add the loss curves for pre-training and fine-tuning in the revised version.

Q7. How to explain the performance drop on the SEED-IV dataset?

To better understand the relatively lower performance on the SEED-IV dataset, we conducted a t-SNE visualization of the learned feature representations and observed that SEED-IV samples are more distant from those of other datasets in the latent space, suggesting a larger domain gap. During preprocessing, we noticed that SEED-IV recordings tend to contain more signal fluctuations and potential artifacts, such as higher variance in channel recordings and frequent noise spikes.

These differences may partially account for the performance drop on this dataset. Compared to methods such as MMM, which first extract and smooth features before model training, mdJPT directly learns from EEG time series. This design prioritizes the preservation of fine-grained temporal information but may also make the model more exposed to noise in certain cases. Nevertheless, on datasets other than SEED-IV, mdJPT generally achieves stronger results than MMM, highlighting the benefits of learning directly from EEG time series. Furthermore, mdJPT consistently outperforms other foundation models that also operate on EEG time series (e.g., LaBram, EEGPT) on SEED-IV, demonstrating a degree of robustness even under challenging recording conditions.

Q8. Evaluation of other model hyperparameters.

Hyperparameters were primarily determined through established heuristics, with targeted searches conducted for several pivotal parameters. Beyond CDA weight, we found several important hyperparameters in preliminary experiments, including MLLA patch stride (which cannot be too large to enable a better extraction of local temporal features) and MLLA output dimension (needs to be at least 32 to achieve a satisfactory performance). We will report the effects of varying these hyperparameters on model performance.

We sincerely thank you for the thoughtful review and for recognizing our performance gains. In the revision, we add true zero‑shot evaluations on unseen datasets, where mdJPT outperforms all foundation‑model baselines. We also assess higher‑order alignment via MK‑MMD (Gaussian, multi‑kernel) and find CDA remains competitive while being simpler to optimize. We further include t‑SNE and clustering‑coefficient analyses to demonstrate the pre-training effect. We appreciate your suggestions—they substantially improved the clarity and completeness of the work.

Q1. The limitation of fine-tuning on the target dataset.

To evaluate our method on real zero-shot generalization tasks, we further conducted an experiment to directly apply the pre-trained model to a new dataset without fine-tuning. We compute cosine similarity across samples and define zero‑shot accuracy by whether the top‑similar pair shares the same label. Samples from the same trial are excluded to avoid leakage. The results with SEED-V as the target dataset are shown below.

Table: Comparison of different models under zero-shot settings.

mdJPT outperforms DE baseline in the zero‑shot setting. It improves by 9.6% over the DE baseline on the SEED-V dataset. On the FACED and DEAP datasets, mdJPT also outperforms the DE baseline by 8.5% and 17.9%, respectively. We will add the results of other comparison models in the discussion session.

We clarify the original evaluation protocol as "few-subject fine-tuning", and the new protocol as "zero-shot generalization".

To our knowledge, this is the first zero‑shot cross‑dataset EEG emotion experiment. Prior work typically fine‑tunes per subject. These results highlight the robustness and generalizability of our framework.

Q2.1 Residual feature distribution differences persist despite the application of CDA. The reliance of CDA solely on covariance may overlook higher-order distribution shifts.

We understand the reviewer's concern about the possibility of higher-order distribution shifts. We agree that relying only on covariance may miss higher‑order distribution differences. Indeed, CDA loss aligns global second-order statistics (i.e., inter-channel covariance structures) across subjects and datasets. This captures inter-region coupling (or functional connectivity) patterns in EEG—features that have been increasingly emphasized in recent neuroscience and EEG decoding literature [1-3].

Although CDA can align EEG with the second-order relations, residual distribution shifts may persist, especially under extreme heterogeneity. To test the impact of higher-order distribution, we add an experiment with higher-order alignment using Multiple Kernel Maximum Mean Discrepancy loss (MK-MMD, Gaussian kernel). The MMD loss with a Gaussian kernel includes all moments of the distribution. It implicitly maps inputs into an infinite-dimensional reproducing kernel Hilbert space (RKHS), where the distance between distributions captures differences across all orders of moments. Multiple kernels (five here) are introduced to reduce the reliance on the Gaussian kernel length. All other settings are kept the same as our models. The results with SEED-V as the target dataset are shown below:

Table: Comparison with MK-MMD loss.

CDA slightly outperforms MK-MMD in our setting, suggesting that aligning second-order statistics already captures much of the cross-dataset shift in EEG (e.g., montage/linear mixing). Higher-order criteria like MK-MMD can, in principle, model richer differences, but in our experiments they did not yield clear gains. The residual discrepancies likely arise from shifts not easily addressed by distribution matching alone (e.g., elicitation protocol mismatch or pronounced nonstationarity), so even higher-order matching may not fully close the gap. Besides, kernel-based higher-order alignment adds tuning complexity and can be harder to optimize stably. We will explore complementary solutions in future work.

Q2.2 The effectiveness of feature distribution alignment should be visualized.

In Appendix D.2, we visualized the latent representations on each target dataset. Although emotion labels were not used in pre-training, the features extracted by pre-trained mdJPT are visually more separable than the original DE features, especially on datasets with fewer emotions, like SEED. This demonstrates the effectiveness of pre-training in facilitating downstream emotion recognition tasks. In the revised version, we will provide the t-sne visualization of all datasets before and after pre-training, to investigate whether pre-training can effectively mitigate the data discrepancy across datasets.

Q3. The tuning of the CDA weight not inadequately explained in Section 3.5.

We agree that performance varies with λ and is not strictly monotonic. In Table 5, the drop at λ=0.01 coincides with a larger STD, which could be attributed to occasional training instability when ISA and CDA losses exert partially opposing gradients at this λ. In the revision, we will run a denser sweep over different λs (λ∈{0, 0.001, 0.002, 0.005, 0.0075, 0.010, 0.015, ...}) and repeat each point over multiple random seeds, providing a clear, reproducible guideline for selecting λ.

Global response

We sincerely thank the reviewers for the thoughtful and constructive feedback. We are encouraged by the recognition of our framework's significance and competitive performance. At the same time, we acknowledge the concerns regarding our motivation, methodology, and clarity, and we are committed to addressing all points thoroughly in our revision. We also provide additional experiments and explanations to address reviewers' concerns:

1. The performance of zero-shot generalization to unseen datasets. (Reviewer BNoM, CDLA, and VGNY)

2. The performance on emotion-elicitation protocols other than video-elicitation. (Reviewer BNoM)

3. Evaluation on more affective dimensions on the DEAP dataset and the transfer from DEAP to other datasets. (Reviewer VGNY and YfAh)

4. Comparison with alternative methods, including supervised contrastive learning and MMD loss. (Reviewer CDLA and YfAh)

5. The motivation behind the design of CDA loss and MLLA encoder. (Reviewer YfAh)

Below, we respond to each reviewer in detail.

Response to reviewer BNoM

We sincerely thank you for the thoughtful and constructive review. We are glad you found the problem setting and our multi‑dataset pre-training framework meaningful. Your comments directly helped us improve the paper’s clarity, rigor, and scope. Following the suggestions, we added a zero‑shot evaluation with no fine‑tuning on the target dataset and included results on the EmoEEG‑MC imagery‑induced dataset to test generalization beyond video stimuli. We add analyses on dataset misalignment and the effect of random seeds, clarify method details, and add the discussion of results and limitations.

Q1. Visualization of dataset misalignment.

We visualize t‑SNE embeddings of differential entropy (DE) features extracted from EEG signals across datasets. Data from the same dataset cluster together, while different datasets are separated. This demonstrates large cross‑dataset misalignment in EEG features. We will add the figure in the revised version.

Q2. No dataset from elicitation techniques other than video-evoking is used, like semi-naturalistic tasks.

Most widely used EEG emotion datasets are video‑induced, which emphasizes controlled, video‑elicited designs. To assess generalizability beyond video stimuli, we add experiments on EmoEEG‑MC [1], a multi‑context dataset published in 2025 that includes an imagery‑induced paradigm (guided narratives with active imagination), eliciting more internally driven and sustained emotions.

We evaluate on the imagery task using the same protocol: pretrain on six video‑induced datasets, then fine‑tune on 1/4 of EmoEEG‑MC subjects and test on the remaining.

Table: Comparison of different models on the EmoEEG-MC dataset with imagery-induced emotion.

Our method outperforms other baselines in the imagery-induced setting. Comparison with other models will be provided in the discussion session. Transfer from video to imagery remains relatively low, likely due to scenario discrepancies. Future work needs to pretrain on more diverse and naturalistic tasks.

[1] Xu X, Shen X, Chen X, et al. A Multi‑Context Emotional EEG Dataset for Cross‑Context Emotion Decoding. Scientific Data, 2025, 12(1): 1142.

Q3. Authors should discuss the lower accuracy on datasets like FACED. Why do the results vary so much?

Performance varies for several reasons. First, label space differs: DEAP (2 categories of high/low valence), SEED (3 categories), SEED‑IV (4 categories), SEED‑V (5 categories), SEED‑VII (7 categories), and FACED (9 fine‑grained emotions, with more positive categories). Several positive emotions in FACED (joy, amusement, inspiration, tenderness) may be subtle in neural differences and harder to separate. Second, annotation standards differ: DEAP uses continuous affective ratings, while others use discrete labels, affecting uncertainty and separability.

Sensor setup and signal quality also matter. SEED uses 62 channels; DEAP and FACED use 32, reducing spatial resolution. During preprocessing, SEED‑IV showed more noise/artifacts (higher channel variance, more disruptions). A t‑SNE visualization places SEED‑IV farther from other datasets, indicating a bigger domain gap, which may explain its relatively lower performance.

Q4. The issue of fine-tuning and the evaluation of zero-shot performance.

To evaluate our method on real zero-shot generalization tasks, we further conducted an experiment to directly apply the pre-trained model to a new dataset without fine-tuning. We compute cosine similarity across samples and define zero‑shot accuracy by whether the top‑similar pair shares the same label. Samples from the same trial are excluded to avoid leakage. The results with SEED-V as the target dataset are shown below.

Table: Comparison of different models under zero-shot settings.

mdJPT outperforms DE baseline in the zero‑shot setting. It improves by 9.6% over the DE baseline on the SEED-V dataset. On the FACED and DEAP datasets, mdJPT also outperforms the DE baseline by 8.5% and 17.9%, respectively. We will add the results of other comparison models in the discussion session.

To our knowledge, this is the first zero‑shot cross‑dataset EEG emotion experiment. Prior work typically fine‑tunes per subject. These results highlight the robustness and generalizability of our framework.

Q5. Exact subject split ratios for fine-tuning vs. test.

In lines 185–187: “We split the target dataset at a 1:3 subject ratio for fine‑tuning and testing. A leave‑one‑dataset‑out cross‑validation is employed. Our method is pre-trained on all datasets except the target.” Concretely, 25% of target subjects are randomly selected for fine‑tuning and 75% for testing. We repeat the random split six times and report mean ± SD. We will clarify this in Section 3.1.

Q6.1 The number of random seeds or runs per dataset.

For fine‑tuning/testing, we use six random splits (as in Q5) and report mean ± SD. During pre-training, the seed (19260832) was randomly chosen without cherry‑picking. We will replicate the full pipeline multiple times with different random seeds and report their consistency in the discussion session.

Q6.2 Whether splits were consistent across models?

Yes. We strictly use the same fine‑tuning/test splits and the same seeds for all models.

Q7. Whether the same EEG preprocessing was used for the baseline?

Yes. All methods share the same standard EEG preprocessing. Details are in Appendix C.

Q8. Why does the baseline not include any task-specific cross-dataset model?

Our baselines match mdJPT’s goals: (i) multi‑dataset joint pretraining, (ii) generalization to unseen categories, and (iii) subject‑independent evaluation (apply to new subjects directly). Most cross‑dataset methods are one‑to‑one adaptations, assuming the same label space and requiring target data—often per‑subject—for training/fine‑tuning (e.g., adversarial alignment). As summarized in Table 1, these pipelines do not simultaneously handle new categories and new subjects, the core challenge addressed by our framework.

For fair comparison, we use foundation‑model baselines, pretrained on multiple large datasets and evaluated on new ones, under the same lightweight classifier protocol. We will clarify this rationale in the revision.

Q9. Limitations of generalizability, negative societal effects, and the real-world applicability.

We will add the following paragraphs to the Discussion:

“Another limitation is the lack of diversity in current EEG emotion datasets—participants are often from limited age groups (generally young people), cultural backgrounds (Probably British for DEAP and Chinese for other datasets), and mental health status (healthy), limiting generalizability to heterogeneous populations. Real‑world deployment also remains constrained by the practicality of current EEG hardware, which is often cumbersome. The ongoing development of lightweight, wearable EEG is expected to enable more practical applications.

We also acknowledge societal and ethical risks, including potential misuse in surveillance, workplace monitoring, and unauthorized emotional profiling. Such uses may lead to overinterpretation or coercion without informed consent. As the field advances, it is essential to establish clear regulatory oversight and develop privacy‑preserving frameworks.”

Q10. Some figures lack detailed captions explaining all subcomponents or labels.

We will revise the figure captions to include more details. For example, in Figure 2, we revise the caption of sub-figure A as: "A) The overall framework. Our framework includes three stages: multi-dataset joint pre-training, classifier fine-tuning, and testing. The EEG encoder contains an MLLA channel encoder and a spatiotemporal dynamics model. During pre-training, the encoder is optimized with a combination of CDA and inter-subject alignment (ISA) losses. During fine-tuning, the encoder is frozen, and a lightweight classifier is trained on a small subset of subjects from the target dataset. Evaluation is then performed on the remaining (held-out) subjects.

Thank you for the rebuttal. I appreciate the authors’ efforts to incorporate additional experiments and provide clarifications, especially the zero-shot evaluation, t-SNE visualization, and inclusion of the EmoEEG-MC imagery-induced dataset. However, I still believe that some results, particularly the low performance on datasets (FACED and EmoEEG-MC) and zero-shot transferability, still need a deeper discussion at the data level, given that the core claim of the work is to improve generalization across datasets. Overall, this is a promising and timely contribution to EEG pertaining.