CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding

Dear Reviewer 6YFP,

Thank you for your encouraging feedback and constructive suggestions! We are greatly encouraged by your recognition that the paper tackles a timely and significant problem in EEG, paving the way for a generalized framework for EEG representation, and is considerably valuable for future directions in EEG research. We are also pleased that you found our proposed method thoroughly justified and technically sound. Due to time and page limitations, we have made every effort to address your questions and suggestions through the clarifications provided. We hope these responses resolve most of your concerns and will be considered in your final assessment.

W1. About the TUEx Datasets

We clarify that our pretraining task, masked signal reconstruction, is purely unsupervised, involving no downstream task labels. Thus, there is no label leakage. The aim is to learn general spatiotemporal EEG representations. In prior foundation model work, such as [NeurIPS2024] BIOT and [ICLR2025] Cbramod, TUEx datasets were also used for pretraining, and they reported fine-tuned results on TUEV and TUAB datasets. To further enhance transparency, we will explicitly mark the TUEx results and compute an additional macro-average excluding TUEx datasets in Table 2. On this new macro-average, CSBrain achieves 0.6829, outperforming CBraMod (0.6537) and LaBraM (0.6430), consistently surpassing competitors by 3+%.

W2&Q3. Spatial Assignment into Brain Regions

Our fundamental goal is to develop a generalizable EEG foundation model applicable to diverse EEG tasks. To ensure broader applicability, the spatial region assignment adheres to the standard 10-20 system compatible with all used datasets. In this system, electrode labels serve as proxies for anatomical brain regions, and electrodes are mapped to five primary regions: Frontal (F), Central (C), Parietal (P), Temporal (T), and Occipital (O). This information was briefly introduced in Sec D.1 (L 727-728). To further enhance readability, we will expand this explanation into a dedicated subsection (e.g., C.5) in the supp. mat., detailing electrode assignments for each dataset. We will also release the code and model to support community use of our work.

W3. Classic Work on Multi-Scale Modeling in CV

We appreciate your insightful suggestion regarding Szegedy et al. as a key early work that introduced multi-scale convolutions in the CV domain. We will gladly include this classic work as a reference in the main paper's Introduction (L76-80) to further elaborate on the significance of multi-scale modeling.

W4. Differences to Prior Works

As shown in Fig. 1 of the main paper, CSBrain fundamentally differs from prior EEG foundation models by explicitly addressing and leveraging the intrinsic cross-scale spatiotemporal structure of neural activity, a property often overlooked by previous scale-agnostic dense modeling paradigms.

• Difference in Tokenization: Prior models typically rely on one-time, fixed-scale tokenization strategy, yielding token representations that struggle with diverse neural patterns across tasks. In contrast, our proposed CST aggregates multi-scale features within localized temporal windows and anatomical brain regions via various convolution kernels. This produces compact, scale-aware neural tokens, enabling adaptable alignment with diverse spatiotemporal requirements of EEG decoding tasks.
• Difference in Spatio-Temporal Modeling: Previous works commonly adopt structure-agnostic attention (e.g., dense attention or Criss-Cross Attention) for spatio-temporal modeling, which can introduce redundancies and computational overhead. In contrast, our proposed SSA models cross-window and cross-region dependencies. This structure-aware design enhances scale diversity, reduces spurious dependencies, and lowers computational overhead, yielding more discriminative and efficient representations.
• Hierarchical Architecture: Unlike prior works, CSBrain establishes a hierarchical architecture by alternately stacking CST and SSA, enabling synergistic enhancement and robust EEG representations that capture complex neural patterns. This leads to strong generalization across diverse EEG decoding tasks, as evidenced by significant performance gains on 16 datasets.

We greatly appreciate your suggestion and will revise the Introduction by adding a separate paragraph after Line 88 to more clearly emphasize the differences between our approach and prior state-of-the-art EEG foundation models.

Q1. Comparison of Number of Kernels

• Computational Cost Comparison: For kernel configurations K=1, K=2, and K=3, parameters are 4.88M, 4.92M, 4.94M respectively, and FLOPs are 246.22M, 265.42M, 275.02M. The parameter count increases only slightly, and FLOPs show a modest increase from K=1 to K=3. This is because we keep the total feature embedding dimension (=200) constant via an exponentially decaying scheme for individual kernel dimensions, balancing representational capacity and efficiency (main paper L168-172). Despite these constrained increases, our model demonstrates significantly improved performance at K=3 across multiple tasks (main paper Fig. 4, L248-253). This outcome strongly supports the effectiveness of cross-scale spatiotemporal modeling, proving performance gains stem from respecting EEG's intrinsic cross-scale properties, not solely from capacity increase.

• Ablation with Additional Kernel Variants: To address your concerns further, we added experiments with the kernel configurations [1,1,1], [3,3,3], and [5,5,5], as well as a baseline with [3,3,3]+Dense Attention. Due to time and computational constraints, these experiments were performed using a subset of the TUEG dataset (2000 hours, excluding TUEV) and evaluated on two downstream tasks: |Kernel Config|TUEV B-Acc|TUEV Kappa|TUEV F1-W|BCIC2020-3 B-Acc|BCIC2020-3 Kappa|BCIC2020-3 F1-W| |-|-|-|-|-|-|-| |[1,3,5]|0.6414±0.0187|0.6236±0.0198|0.7971±0.0183|0.5443±0.0232|0.4254±0.0301|0.5527±0.0297| |[1,1,1]|0.6122±0.0193|0.5596±0.0212|0.7713±0.0205|0.5041±0.0208|0.3821±0.0223|0.5019±0.0202| |[3,3,3]|0.6338±0.0184|0.6137±0.0179|0.7870±0.0188|0.5298±0.0247|0.4354±0.0298|0.5399±0.0280| |[5,5,5]|0.6188±0.0190|0.5765±0.0209|0.7732±0.0195|0.4998±0.0237|0.3809±0.0242|0.4931±0.0279| |[3,3,3]+Dense Attention|0.5901±0.0178|0.5380±0.0199|0.7531±0.0212|0.5076±0.0221|0.3844±0.0269|0.5023±0.0222|

These results further validate the effectiveness of CST for cross-scale modeling, where the [1,3,5] configuration achieves the best performance, while [3,3,3] provides the second-best results. This demonstrates that explicitly modeling diverse scales is more effective than simply adding more filters of the same scale. This comprehensive analysis reinforces that CSBrain's performance gains are indeed attributable to its principled cross-scale spatiotemporal modeling. We will update Supp. Sec. E to reflect these insights.

Q2. CST Procedure

Q2-1. Replacement of intra-region/window attention:

• We considered replacing CST's convolutions with intra-region/window attention modules during the early stages of our project. However, preliminary experiments showed negligible performance improvements (~1% on key metrics for Seed-V, TUEV) while substantially increasing parameters due to doubled attention computations.
• We will gladly cite and discuss the classic work by Wang et al. in the Introduction and Related work.
• We also explored Swin-style models early in our project. However, we found that their shifted window partitioning for efficient computation (as shown in Fig. 4 of their paper) would likely disrupt the temporal and spatial order inherent in EEG signals. While highly effective for images, transference of such mechanisms to EEG data poses fundamental challenges, as preserving sequential and local spatial dependencies is paramount. We will update the manuscript to reflect these considerations more explicitly.

Q2-2. Inter-region Attention:

• To clarify, as stated in L193, tokens are sequentially sampled from each region, not randomly. In Supp.E.3, we compare four strategies for constructing Regional Descriptors, including random sampling. Results show that sequential sampling with the average feature gives the best performance, ensuring each token represents its region’s spatial context.
• In inter-region attention, token count is not reduced. As in Equation 5, each position $x_{i,j}$ is assigned to a spatial group for attention, ensuring that no information is lost during sampling. This design preserves the spatial and temporal relationships between regions and guarantees that tokens flow consistently throughout the model. We will rewrite L191-198 for further clarity.

Q2-3. Zero-padding: Yes. For regions with fewer electrodes than the kernel size, we apply zero-padding during the convolution operation.

Q2-4. Inter-window Attention:

• In inter-window attention, each token is enriched with information from other tokens in the same window via CST, ensuring tokens share a common spatiotemporal context. This allows relative indices to be grouped together, as windowed tokens have a consistent temporal structure.
• In contrast, for inter-region attention, varying electrode counts prevent tokens from fully representing their regions. Therefore, we use token descriptors, adding the average feature of the region to each token.
• Our preliminary experiments showed that adding a token descriptor to windowed tokens did not significantly impact performance (~<1% difference).

Due to rebuttal space constraints, some brief experimental results will be included in revised versions of the paper. We will expand L191-198 in the main paper to clarify the specific implementation details and design motivations for SSA. Furthermore, we will open-source the relevant code to facilitate community use of our work.

Dear Reviewer 6YFP,

Thank you for your thoughtful feedback. We agree that the term "covering partition" more accurately and conceptually describes our procedure. We will update the manuscript, particularly the expression in Line 193, to enhance conceptual clarity and facilitate future readers' understanding.

Furthermore, we are very grateful for your active and insightful feedback throughout the entire review process, and we believe it has significantly improved the quality of our work. We hope that, through the collective efforts of the entire community, we can advance EEG modeling toward a more generalizable direction and facilitate its application in the real world (including clinical diagnostics, cognitive augmentation, and brain-computer interfaces) to improve human life.

Dear Reviewer yshg,

Thank you for your encouraging feedback and constructive suggestions! We are greatly encouraged by your recognition of our very comprehensive downstream evaluation and the use of a large pre-training dataset. Due to time and page limitations, we have made every effort to address your questions and suggestions through the clarifications provided. We hope these responses resolve most of your concerns and will be considered in your final assessment.

W1-3. Clarification of Tokenization

• In the filed of EEG foundation model, which commonly employ Transformer architectures, tokenization is not just a convolutional operation but the essential process of converting continuous spatiotemporal signals into discrete, Transformer‑friendly units (i.e., EEG tokens). This terminology is well established in prior works (e.g., [ICLR 2024] LaBraM; [NeurIPS 2024] BIOT), and since each EEG token serves as the minimal modeling unit, effective tokenization is a fundamental prerequisite for any EEG foundation model.
• In CSBrain, our Cross-Scale Spatiotemporal Tokenization (CST) module aggregates multi-scale features across both temporal and spatial dimensions into compact, structured token representations. This design is guided by the intrinsic property of neural activity — the cross-scale spatiotemporal structure — and aligns with the underlying structure of EEG signals. The key distinction of our design is that tokenization in CSBrain is not simply a stack of convolutions, but a hierarchical process that aggregates information across scales and spatial regions. This allows the model to capture richer, more granular representations of EEG signals. The resulting tokens, which encapsulate both local and global spatiotemporal information, are passed to the attention layers where the model learns to capture dependencies across different time windows and brain regions.

We will rewrite L131-140 of the main paper to more explicitly clarify the role of tokenization and its contribution to the CSBrain architecture.

W4. About MSE loss

For masked pretraining, MSE is a widely used and intuitive loss function in signal and image reconstruction tasks, as seen in prior works like MAE and Cbramod.. While MSE does prioritize low-frequency features due to its focus on minimizing larger amplitude errors, this characteristic aligns well with EEG signal properties, where energy is predominantly concentrated in lower frequency bands.

Importantly, in CSBrain, input EEG segments undergo preliminary encoding to extract both temporal and spectral (FFT) features (Sec.2.1 L124-129). During pretraining, the model reconstructs the raw EEG based on these spectral-rich features, compelling it to utilize high-frequency FFT components for accurate reconstruction. Additionally, CSBrain's CST module extracts features across multiple temporal scales, capturing information from a range of EEG frequencies. Therefore, our masked reconstruction task could enforce the model to learn beyond low-frequency features, enabling a more comprehensive understanding of EEG signals.

Our extensive experiments across 11 EEG tasks and 16 datasets demonstrate CSBrain’s strong performance, including tasks sensitive to high-frequency or transient features, such as Motor Imagery Classification and Event Type Classification. To enhance clarity, we will integrate this discussion into Sec. 2.4 (L205-215). Furthermore, inspired by your insightful comment, we agree that augmenting pretraining with multiple objectives could capture more complexity, and we will incorporate this into our “Future Work” section (Sec. G.3) to encourage broader community efforts.

W5-7&Q4. About Pre-trained Model Evaluation

1. Pretraining Data Amount.

• Fairness of evaluation protocols: In the EEG community, comparing pretrained models in downstream task performance under varying pretraining data is a standard practice adopted by prior works such as [ICLR2024]LaBraM, and [ICLR2025]Cbramod. Similarly, in broader AI, foundation models like CLIP, BLIP, and GPT are rarely pretrained on identical datasets, as uniform pretraining with large-scale data is impractical. Thus, our evaluation protocol is NOT unfair.

• Comparison under the same pretrained data amount: CSBrain uses more pretraining data than LaBraM but less than BIOT, and uses the SAME pretraining data as the previous SOTA model Cbramod (using the TUEG dataset). Across 11 representative tasks and 16 datasets, CSBrain consistently outperforms Cbramod, achieving a macro-average score of 0.7095, surpassing Cbramod’s score of 0.6760. This demonstrates that even with the same amount of pretraining data, our architecture holds a clear advantage. We will add a comparison of the pretraining data amount in L263 of the main paper for greater clarity and transparency.

Your insightful comments inspire us to consider future work in building sufficiently large-scale, high-quality open-source EEG databases for unified pretraining. We will incorporate this direction into Sec. G.3 of our future work to encourage community efforts toward these advancements.

2. Regarding Why We Focus on Downstream Performance For Comparisons.

Evaluation on downstream performance aligns with the core goal of EEG foundation models, which is to achieve performance gains and generalization on downstream tasks through large-scale unsupervised pretraining and a unified model architecture. Since pretraining data and objectives vary across models, prior foundational models such as LaBraM, BIOT, and Cbramod have followed similar evaluation pipelines, focusing on downstream performance after pretraining, without directly evaluating the pretraining task itself. In line with these models, we prioritize the evaluation on downstream tasks.

3. More Validation on Pretraining.

In Supp. Mat., we have provided further validation of the quality of our pre-trained model's representations through several approaches:

• Pre-training Effectiveness: In Supp. Mat. E.1, we show significant performance improvements on downstream tasks by comparing models with and without pre-training, validating the effectiveness of our pre-training strategy.
• Pre-training Scale Analysis: Pre-training Scale Analysis: In Supp. Mat. E.2, our experiments demonstrate that increasing pre-training data leads to better performance, proving our compliance with the Neural Scaling Laws for foundation models and good scalability.
• Pre-training Process Transparency: In Supp. Mat. B.2, we provide the Training Loss Curve and hyperparameters, demonstrating effective convergence under our pretraining objective.

Q1. FFT for Spectral Information

The use of Fourier spectrum information has been widely adopted in previous EEG foundational models ([ICLR2024]LaBraM, [NIPS2024]BIOT, and [ICLR2025]CBraMod). These models consistently show that integrating frequency-domain information with temporal features significantly enhances decoding performance. Following their practice, our model integrates FFT‑based spectral extraction with 1D convolutions, harnessing their complementary strengths. and the physiological relevance of EEG frequency‑domain information. While 1D convolutions excel at learning local temporal dependencies, FFT provides an explicit representation of the signal's frequency components, enabling the model to capture essential frequency-based patterns that 1D convolutions might miss. By combining both approaches, our model effectively leverages both temporal and spectral information, providing a comprehensive representation of the EEG signal.

Q2. The difference of SSA and spatiotemporal Attention

We compare SSA with Temporal and Spatial Attention (TSA) from both architectural and experimental perspectives.

• Architectural Comparison: The key difference between SSA and TSA lies in how they model long-range spatiotemporal dependencies. TSA typically focuses on independently modeling dependencies along the time and space axes, without distinguishing between intra- and inter-region, or short-term and long-term cross-scale structures. In contrast, SSA respects the intrinsic property of neural activity, accounting for cross-scale interactions across brain regions and time windows. It works in close conjunction with CST to integrate these spatiotemporal dependencies in a more structured and efficient manner. By leveraging the inherent cross-scale structure of EEG signals, our model effectively captures both local and global patterns.
• Experimental Comparison: We compare SSA with Cbramod [ICLR2025] that explores TSA by a Criss-Cross Transformer for modeling dependencies along the time and space axes. As shown in Section 3.3 (L258-277) and Table 2 of our paper, our model achieves a 0.7095 macro-average across 11 representative EEG tasks and 16 datasets, significantly outperforming Cbramod (0.6760 macro-average). Additionally, in L293-305 and Figure 5, our proposed SSA demonstrates superior efficiency compared to TSA (Criss-Cross Attention).

We will open-source the relevant code and models to facilitate community use and further advancements in this area.

Q3. Kernel Stacking Comparison

As discussed in L278-292, our "Tokenization Comparison" ablation study clearly demonstrates that full cross-scale tokenization (K=3) significantly outperforms both single-scale (K=1) and dual-scale (K=2) variants. Specifically, our SSA (K=3, stacked) consistently achieves the best performance across four representative tasks. This suggests that explicitly modeling multiple temporal and spatial scales contributes to learning richer EEG representations, further validating the effectiveness of our core insight—cross-scale spatiotemporal modeling.

Thank you for the detailed responses. I now understand better the details of the model. I understand that tokenization is a hierarchical process, but still it doesn't produce discrete vectors, which I would call tokens.

For a lot of your arguments you are citing Labram and CBramod, two papers which came out in the last year. I would appreciate a more grounded view of EEG modeling literature. Just because one recent paper does something does not mean that it is well established.

Dear Reviewer yshg,

Thank you for your further comments. We are glad our response has clarified the model details.

In EEG modeling, tokenization refers to converting continuous spatiotemporal EEG signals into discrete token sequences that serve as fundamental units for Transformer-based modeling. Each token itself is a compact, continuous vector representation. This is consistent with established practices in the broader Vision Transformer literature [1],where images are transformed into sequences of tokens, each representing a continuous vector of image patches. Within the EEG community, numerous studies have explored tokenizing EEG signals for Transformer-based models [2-7].

[1] Dosovitskiy, Alexey, et al. "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale." International Conference on Learning Representations.

[2] Song, Yonghao, et al. "EEG conformer: Convolutional transformer for EEG decoding and visualization." IEEE Transactions on Neural Systems and Rehabilitation Engineering 31 (2022): 710-719.

[3] Shen, Qi, et al. "LGSleepNet: an automatic sleep staging model based on local and global representation learning." IEEE Transactions on Instrumentation and Measurement 72 (2023): 1-14.

[4] Mentzelopoulos, Georgios, et al. "Neural decoding from stereotactic EEG: accounting for electrode variability across subjects." Advances in Neural Information Processing Systems 37 (2024): 108600-108624.

[5] Khadka, Rabindra, et al. "Inducing inductive bias in vision transformer for EEG classification." ICASSP IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024.

[6] Kim, Sung-Jin, et al. "Toward domain-free transformer for generalized EEG pre-training." IEEE Transactions on Neural Systems and Rehabilitation Engineering 32 (2024): 482-492.

[7] Cui, Wenhui, et al. "Neuro-gpt: Towards a foundation model for eeg." 2024 IEEE International Symposium on Biomedical Imaging (ISBI). IEEE, 2024.

LaBraM and Cbramod are cutting-edge models in EEG foundation research, which is why we have compared and referenced them extensively. We agree that a grounded view of EEG modeling literature is important. In fact, we cite and discuss 74 and 81 papers in the main paper and supplementary materials, including many classic works in EEG modeling. We hope that, building on these prior community efforts, our work will contribute to advancing EEG modeling toward more generalizable solutions, facilitating real-world applications such as clinical diagnostics, cognitive augmentation, and brain-computer interfaces.

Thank you again for your review. If you have any further questions, please feel free to contact us.

Dear Reviewer Yw1n,

Thank you for your encouraging feedback and constructive suggestions! We are greatly encouraged by your recognition of our comprehensive evaluation and that the paper is well-written, with a clear structure and logical flow. Due to time and page limitations, we have made every effort to address your questions and suggestions through the clarifications provided. We hope these responses resolve most of your concerns and will be considered in your final assessment.

W1&Q1. Feature Fusion Strategy

Thanks for the suggestion. In EEG analysis, both concatenation and addition are commonly used to fuse temporal and spectral information. We opted for concatenation as it preserves the integrity and independence of each feature type, thereby providing a richer input space. The choice of concatenation-based feature fusion is also widely supported by prior validated studies [1-2].

Notably, the core contribution of our work lies in the CSBrain architecture itself, particularly the Cross-Scale Spatiotemporal Tokenization (CST) and Structured Sparse Attention (SSA) modules. These components build cross-scale dependencies upon these fused initial features, capturing intricate spatiotemporal patterns. The effectiveness of our work has been validated across 11 representative tasks and 16 datasets.

We agree that examining their impact on EEG representation learning is valuable and will consider this in future work, given the limited rebuttal timeframe. Your insightful comment will be included in Sec. G.3 (Future Work) to guide further research.

[1] Li, Yang, et al. "A temporal-spectral-based squeeze-and-excitation feature fusion network for motor imagery EEG decoding." IEEE T-NSRE2021, Cited by 142

[2] Jia, Ziyu, et al. "Sst-emotionnet: Spatial-spectral-temporal based attention 3d dense network for eeg emotion recognition." ACM MM2020, Cited by 170

W2/3&Q2/4. Design Motivation of SSA

• Motivation: SSA is designed to model long-range spatiotemporal dependencies across time windows and brain regions via efficient sparse attention. It complements CST that captures structured, cross-scale representations within local temporal and spatial contexts. This synergy enables CSBrain to effectively model brain activity from local to global scales, aligning with the cross-scale nature of neural dynamics.

• Details: CST aggregates contextual information within local temporal windows and brain regions, endowing each token with window/region-level awareness. Based on this, SSA performs sparse attention computation across windows and regions. Specifically, cross-scale tokens are sampled and grouped based on the same indices, and attention is then computed within each group, effectively having a representative from each window/region interact across regions/windows. Multiple groups are set up to ensure every token is included in at least one group for computation, preventing information loss. Additionally, in inter-region attention, we introduce a regional descriptor, which adds the average feature of the region to the sampled token. Unlike temporal windows (which have fixed lengths), the number of electrodes varies across brain regions. In regions with more electrodes, CST might not fully capture all information. By introducing the regional descriptor, we ensure complete information flow between regions.

As shown in Sec. 2.2 (L293–305), SSA achieves superior performance and efficiency compared to other attention mechanisms. We will integrate this explanation into L185–203 of the main paper for clarity, and release our code and models to support future research.

W4. Details on Foundation Model Baseline

For all compared foundation models, we utilized their officially published weights from their respective GitHub repositories. This is a common and validated practice adopted by prior works including [ICLR24]Labram and [ICLR25]CbraMod for evaluating foundation models, ensuring fair comparison against their intended performance.

The specific evaluation protocols for each task and dataset are outlined in Supp. Sec. C&D, following practices of CbraMod and other well-established studies. To further enhance transparency, we will include an additional note in Sec.C.2 to specify the use of the published weights and provide weight download links.

W5. More Ablations on CST

• As shown in L278-292 and Fig. 4 of the main paper, we have already provided an analysis of token granularity choices in CST, as well as the effect of CST stacking strategy on performance. These results demonstrate CST's role in learning richer, more flexible EEG representations and the importance of token granularity in effectively capturing diverse EEG dynamics across tasks.
• To further address your concerns and provide deeper insights into token granularity and window/region design, we conduct additional experiments to explore the configuration of using multiple kernels of the same scale: [1,1,1], [3,3,3], and [5,5,5], in contrast to our multi-scale [1,3,5] design. Due to time and computational constraints, these experiments were pre-trained on a subset of the TUEG dataset (2000 hours, excluding TUEV) and evaluated on two representative tasks.

These results further validate CST’s effectiveness in cross-scale modeling. Our proposed [1,3,5] multi-scale configuration consistently outperforms single-scale alternatives, showing that modeling diverse scales is more effective than simply increasing uniform kernels. This reinforces the importance of CST’s design for capturing EEG’s complex multi-scale spatiotemporal patterns and enhancing generalization. We will update Supp. Sec. E to reflect these insights.

W6. GNN Baselines

We will reference and discuss the two classic GNN works you mentioned in both the Introduction and Related Work, and include them as baselines in our evaluation. Due to time constraints, we conducted tests on 4 representative tasks and will include the full results in the revised paper. Note that HetEmotionNet was originally designed for heterogeneous multimodal inputs, so when applied to EEG-only datasets, the graph is simplified, which may reduce its performance.

As shown, CSBrain consistently outperforms GNNs, demonstrating its superior capability in capturing spatiotemporal dependencies.

W7/8&Q5. More Visualization Results

We greatly appreciate your valuable suggestions. We fully agree that such analyses would significantly enhance model interpretability and highlight CSBrain's advantages. However, due to NeurIPS conference committee regulations, we are unable to submit new PDFs or external links to showcase additional visualization results during the rebuttal phase.

Nevertheless, we have already provided a rich set of visualizations in the main paper and supplementary materials, designed to illustrate CSBrain's interpretability and the effectiveness of its cross-scale modeling:

• Topography Visualizations: Presented in the main paper (Fig. 6, L337-345) on four representative tasks. Supp. F.1 (L474-501, Figs.6/7/8/9) provides more detailed topography visualizations for various categories within the four different tasks.
• Representation Visualizations (t-SNE): In Supp. F.2 (Fig. 10/11, L502-512), we visualize the feature distributions of raw EEG signals versus CSBrain's learned representations on two tasks. These demonstrate CSBrain's ability to learn highly discriminative and well-separated representations.
• Channel Relationship Visualizations: Supp. F.3 (Fig. 12/13, L513-525) includes visualizations and analysis of learned inter-channel relationships, highlighting our model's capacity to capture both local and global global spatial dynamics.

These existing visualizations collectively demonstrate CSBrain's strong interpretability and the effectiveness of its cross-scale modeling. We believe this constitutes one of the most comprehensive sets of visualizations in published EEG foundation model papers to date. We sincerely hope you understand our inability to upload more visualization results due to NeurIPS regulations, and we look forward to your support.

Q3. Aggregation Context

To clarify, the aggregation in L155-156 refers specifically to anatomically defined brain regions. We do not explicitly perform functional aggregation as precise functional connectivities are difficult to define. Instead, our spatial tokenization explicitly incorporates anatomical priors by assigning electrodes to five brain regions (e.g., frontal, central, parietal, temporal, occipital lobes) anatomically defined in the standardized 10-20 system. These lobes broadly reflect the distribution patterns of functional regions and provide a physiologically meaningful framework for spatial grouping. As evidenced by Fig. 6 and visualizations in the supp, our model effectively learns functional characteristics across tasks in a coarse-grained manner. by facilitating the modeling of cross-scale neural structures that respects brain anatomy, our model allows effective adaptation to multiple downstream EEG tasks.

Dear Reviewer 5sTW,

Thank you for your encouraging feedback and constructive suggestions! We are greatly encouraged by your recognition of our principled architectural innovation, the thorough and convincing experimental evaluation, and the paper's well-written and comprehensive method descriptions. We are also pleased that you found the motivation for cross-scale modeling well-articulated, and that the CST and SSA components are described with sufficient detail for understanding and reproducibility. Due to time and page limitations, we have made every effort to address your questions through the clarifications provided. We hope these responses resolve most of your concerns and will be considered in your final assessment.

W1&Q1. Handling of Varying Spatiotemporal Settings.

1. Different Sampling Rates&Temporal Scales: As mentioned in Sec.2.1 (L120), all raw EEG signals undergo a standardized preprocessing pipeline, including resampling to a uniform 200 Hz. This ensures that temporal differences across datasets are normalized before entering the CST, aligning with standard practices in foundation models like [ICLR2024] LaBraM, [NIPS2024] BIOT, and [ICLR2025] Cbramod. Subsequently, CST further adapts to different temporal scales through its multi-scale temporal convolutions, using kernels of different sizes to capture temporal patterns at various scales. CST and the subsequent SSA module closely collaborate and complement each other to jointly capture diverse temporal scales from local to global. This combined setup collectively fulfills the varying temporal scale requirements of different datasets.

2. Different Electrode Configurations: In CSBrain, we first partition electrodes based on the standardized 10-20 system into brain regions (Frontal, Central, Parietal, Temporal, Occipital). Then, within the electrodes of each defined brain region, CST applies local multi-scale spatial convolutions. CST enables the model to learn localized yet multi-scale patterns robustly, irrespective of channel count variations within a region. This design ensures that while electrode numbers and spatial distributions may vary across different datasets, the unified 10-20 system-based spatial region partitioning and the application of local multi-scale spatial convolutions allow CST to effectively adapt to diverse electrode configurations, ensuring robust compatibility.

To make our design motivation and process even clearer, we will modify Sec. 2.1 and 2.2. Specifically, Sec. 2.1 will further emphasize the role of data preprocessing in handling diverse datasets. Sec. 2.2 (L141-153) will add clarifications to highlight CST's adaptation to different temporal scales and channel configurations. Additionally, we will open-source our code and models to facilitate further research and community use.

Q2. Pretraining Data Setting.

The cross-scale capability of CSBrain comes from the intricate model architecture, not from the variations in input signal format. Through the synergistic combination of the CST and SSA modules, we effectively capture cross-scale spatiotemporal dependencies, as demonstrated by our experiments across 11 tasks and 16 datasets. Speciffcally:

1. Design Justification: The main objective of pretraining is to learn robust, generalizable EEG representations from a large, unlabeled corpus. To achieve this efficiently, we chose a uniform data setting to simplify the data pipeline, ensuring stability and efficiency during large-scale training. Using varying input settings, such as varying EEG segment lengths or channel counts, would complicate preprocessing and model training, leading to instability. This is consistent with other widely validated models like CBraMod, which also use uniform input for pretraining. Additionally, as shown in Supp. Sec. E.1, we have demonstrated that our pretraining improves performance across datasets with various spatiotemporal settings.

2. More Ablation Studies: we further conduct ablation studies to evaluate the impact of pretraining diversity and model configurations on downstream task performance. Due to time and resource constraints, the total pretraining data is fixed at 2000 hours.

(1) Multiple Pretraining (MP): We pretrain on diverse data configurations, including 5s, 20s, and 30s windows with the TUEG dataset (19 channels), and datasets like FACED (10s, 32 channels), SEED-VIG (8s, 17 channels), and BCIC-IV-2a (4s, 22 channels), totaling 2000 hours.

(2) Single Pretraining (SP): We pretrain on a subset of the TUEG dataset (30s, 19 channels), totaling 2000 hours.

(3) Single-Scale CSBrain: A single-scale configuration by adjusting CST convolution kernels to [1, 1, 1].

Results show that, under the same pretraining data volume, variations in data format has a minimal impact on downstream performance. The cross-scale model architecture (CST and SSA) is the primary factor in improving performance, further validating the effectiveness of our pretraining setup and the core idea of cross-scale spatiotemporal modeling.

W3&Q3 Pretrained Channels Selection & Handling of Diverse Downstream Channel Settings.

Pretrained Channel Selection. As noted in Sec.3.1 L225, we follow established practices by [ICLR2025]CBraMod using the TUEG dataset for pretraining. The EEG signals in TUEG range from 19 to 23 channels. We select 19 common EEG channels (Fp1, Fp2, F7, F3, Fz, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, O2) that correspond to a subset of the 10-20 international electrode placement system to ensure clean and uniformly formatted pretraining data. The dropped channels include A1, A2, T1, and T2. A1 and A2 are located near the auricular region, which contain less brain information and are typically used for reference electrodes. T1 and T2 are obsolete and no longer utilized.

Handling of Diverse Downstream Channel Settings.

• Architecture. CSBrain's architecture is inherently designed to adapt to the actual channel count of each downstream dataset. During fine-tuning, for any given downstream dataset, we map its electrodes to a set of anatomically predefined brain regions $\mathcal{R}$ based on the 10-20 system. Within each region ($R_r$), CST applies multi-scale spatial convolution kernels. This allows CST to aggregate cross-scale features from the available electrodes within that region, regardless of their exact number. Subsequently, SSA computes attention based on regional descriptors formed via sequentially sampled tokens $x_r^{rep}$ and average features of $R_r$, establishing long-range dependencies between these regions. Thus, despite varying electrode configurations, our brain-region-based assignment allows diverse downstream setups to benefit from large-scale pretraining.
• Implementation. From an implementation perspective, except for the task-specific heads with fixed input-output dimensions, all model weights are inherited directly from pretraining. This is because the underlying convolutional and Transformer components of CSBrain are flexible enough to handle varying input sizes and dynamically grouped region information.
• Experiment Validation. Furthermore, as stated in supp. mat. Sec. E.1 (L418-435), our work, consistent with prior foundation models like Cbramod and BIOT, clearly demonstrates the effectiveness of pretraining. These pre-trained models significantly improve performance on downstream tasks, even those with different channel settings.

To further enhance transparency, new Sec. B.3 and C.4 will be added to the Supp., detailing the channel settings and region assignments for the pretraining data and each downstream task dataset, respectively. Our code and models for pretraining and downstream fine-tuning will be open-sourced to facilitate community use.

Q4. Complexity Analysis.

We compare the runtime, computational complexity, and performance of SSA versus Dense Attention (DA) on TUEV (C=16, T=5s), BCIC2020-3 (C=64, T=3s), and FACED (C=32, T=10s) datasets. Runtimes are averaged over 500 runs on a single A100 GPU with a batch size of 64. For DA, complexity is $O(N^2)$, where $N = T \times C$. For SSA, complexity is $O(N \cdot k)$, where $k$ = max(electrodes within a brain region) + window length.

We observe that SSA consistently exhibits lower runtime and complexity across various time and channel settings. It should be noted that while SSA offers clear efficiency gains, the absolute runtime difference between SSA and DA is not as dramatically pronounced as the complexity analysis suggests, owing to PyTorch's highly optimized parallel computations for attention mechanisms. Crucially, SSA consistently outperforms DA across all performance metrics. This advantage stems from SSA's respect for the intrinsic cross-scale spatiotemporal structure of neural activity, effectively capturing meaningful dependencies without indiscriminately associating all tokens. We will briefly add this analysis to L293-305 of the main paper and provide the full details in Supp. Sec. E.