- Rating
- 2
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes an agent-based framework that adapt EEG foundation model by generating synthetic EEG. The core premise is that high-fidelity synthetic EEG is not necessarily beneficial for model adaptation. To address this, the author introduce the a planner–evaluator loop refines EEG generation strategies for adaptation. The method is evaluated on nine EEG benchmarks across various foundation models.
Critical strengths
Novelty: Introducing the agent-based adaptation framework for EEG foundation model is a novel and interesting idea.
Clarity: The paper is well organized and easy to follow, with intuitive illustrations.
Critical weaknesses
Weak premise: The core premise of this paper is that high-fidelity synthetic data is not necessarily useful for adaptation. However, it is simply expected and has been observed for generative augmentation in other domains, for example a exacted same synthetic data has high-fidelity with real data, but did't carry new information for model training. Moreover, TS-FID measures only distribution-level similarity to train data and cannot verify whether synthetic datas are novel rather than near-copies. And model utility is defined as (). It is measured only within the proposed loop and is not a general measure of the value of synthetic data. In minor suggestion, the evidence of paper's core premise present a last page and it make some difficult follow the paper.
Limited baselines and marginal performance gain: Recent EEG foundation models such as REVE [1] and CSBrain [2] have become strong baselines in this area, but they are not included in the comparison. Since the proposed method is to be applicable to any EEG foundation model, applying it to these recent models would provide stronger evidence. In addition, the reported baseline performance differs from reported performance in the original papers. Finally, the paper compares only with synthetic-data-based approaches for adaptation. Including other adaptation methods for EEG foundation models, such as [3], would better clarify the contribution. The performance gains of the proposed method in Table 2 are small. In Table 3, the improvements over simple variants are also marginal, and no standard deviations or significance tests are reported. And the agent loop incurs a high computational cost, I wonder detailed cost analysis. The authors should discuss whether the method remains worthwhile given this cost.
[1] REVE: A Foundation Model for EEG, NeurIPS, 2025
[2] CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding, NeurIPS, 2025
[3] Riemannian High-Order Pooling for Brain Foundation Models, ICLR, 2026
Questions for authors
-
If fidelity (TS-FID) is measured against the training data, how can it ensure that synthetic samples are close to real EEG (your train data didn't cover the real world data well)?
-
Could the authors define general utility in a loop-independent way and show the fidelity–utility relationship under this definition?
-
Would additional real data be more beneficial than synthetic data? Please compare training with the same amount of synthetic data versus additional real data.
-
Why do the reproduced baseline results (e.g., CBraMod) differ from those reported in the original papers?
-
Are the improvements in Tables 2 and 3 statistically significant? Please show the statistical test results.
-
Applying the method to recent EEG foundation models (e.g., REVE [1], CSBrain[2]) and comparing other EEG foundation model adaptation method [3] would clarify the contribution.
-
What is the computational cost of the full agent loop compared with standard fine-tuning?
-
[Minor] Please check the column name of BCIC-2A, 4-class in Table 5.
[1] REVE: A Foundation Model for EEG, NeurIPS, 2025
[2] CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding, NeurIPS, 2025
[3] Riemannian High-Order Pooling for Brain Foundation Models, ICLR, 2026
13 comments
Response to Reviewer ZjiB (2/8)
Q2: Analysis of Fidelity–Utility Correlation Using Loop-independent Utility.
Response:
We introduce a loop-independent utility: , where is the performance using only real training data, and is that using real and synthetic data from round . Both settings follow identical adaptation protocols and start from the same pretrained checkpoint. We then recompute the Spearman correlation between fidelity and the new utility. As shown in Table R8, the overall correlation is , , showing no significant correlation between generation fidelity and the actual adaptation benefit of synthetic EEG.
Table R8: Fidelity–Utility Correlation under Loop-independent Utility.
| Dataset | Spearman's ρ | p-value |
|---|---|---|
| Overall | 0.074 | 0.490 |
| BCIC-2A | 0.037 | 0.920 |
| BCIC2020-3 | 0.503 | 0.138 |
| FACED | -0.305 | 0.392 |
| SEED-V | 0.031 | 0.933 |
| TUEV | 0.630 | 0.051 |
| Physionet-MI | -0.379 | 0.280 |
| SHU-MI | 0.273 | 0.446 |
| MentalArithmetic | 0.079 | 0.829 |
| Mumtaz2016 | 0.566 | 0.089 |
Response to Reviewer ZjiB (3/8)
Q3: Comparison of Real and Synthetic Data Augmentation.
Response:
We split the original training set into two disjoint subsets, , where is used for generator training and downstream adaptation, while is excluded from generator training and reserved for real-data augmentation. We compare three settings: training on (i) alone, (ii) , and (iii) , where contains the same number of synthetic samples as , with matched class distributions. All models start from the same pretrained checkpoint and follow identical adaptation protocols. Table R11 reports the results across three downstream tasks, showing that synthetic augmentation achieves comparable performance to additional real-data augmentation under the same data budget. These results suggest that synthetic data can reduce reliance on costly real EEG acquisition while maintaining competitive adaptation performance.
Table R11: Comparison of Real and Synthetic Data Augmentation under Equal Data Budgets.
| Methods | FACED (Bal.Acc) ↑ | SHU-MI (Bal.Acc) ↑ | MentalArithmetic (Bal.Acc) ↑ |
|---|---|---|---|
| Real-only | 55.13 ± 0.95 | 59.41 ± 2.43 | 57.47 ± 3.05 |
| + Additional Real Data | 57.09 ± 0.68 | 60.54 ± 1.58 | 59.79 ± 6.01 |
| + Synthetic Data (Ours) | 57.12 ± 0.63 | 60.12 ± 1.64 | 60.61 ± 2.12 |
Response to Reviewer ZjiB (1/8)
We sincerely thank the reviewer for their helpful suggestions.
Q1:
Response:
The TS-FID metric is based on previous work (JET). We clarify that TS-FID measures distributional similarity to the empirical distribution of observed real EEG, rather than guaranteeing coverage of the entire real-world EEG distribution. To further assess this, we conduct two complementary analyses.
(1) Distributional fidelity. We additionally compute TS-FID using an independent held-out test set as the reference. We normalize TS-FID by the maximum value among the three data. As shown in Table R10(a), the normalized Train–Synthetic and Test–Synthetic TS-FID values are 0.5128 and 0.6722 for BCIC-2A, and 0.4499 and 0.6460 for FACED, respectively. This provides additional evidence that synthetic EEG captures spectral characteristics of independently observed test EEG, beyond the training distribution.
(2) Sample-level novelty. We compare the nearest-neighbor distances of synthetic EEG and test EEG samples to the same training set. We normalize the median nearest-neighbor distances by the maximum value among the three data. As shown in Table R10(b), synthetic EEG achieves a normalized median distance of 1.0000 on both BCIC-2A and FACED, compared with 0.9738 and 0.7497 for real test EEG, respectively. This suggests that synthetic samples are not dominated by near-copies of training examples.
Together, these analyses provide complementary evidence of distributional alignment with test EEG and sample-level novelty. We emphasize that these metrics do not guarantee complete coverage of real-world EEG variability. The test set is used exclusively for post-hoc evaluation and is never involved in generation, strategy planning, or model adaptation.
Table R10: Distributional Fidelity and Sample-level Novelty Analysis.
(a) Distributional Fidelity
(b) Sample-level Novelty