acceptodds
Under review as a conference paper at ICLR 2027

GRFBrain: Graph-Structured Rectified Flows for EEG Dynamic Modeling

Abstract

Forecasting time-varying functional connectivity from electroencephalography (EEG) requires modeling both history-dependent trends and structured variability across channels. Conditional flow matching provides a framework for distributional forecasting, yet it remains unclear whether graph-informed source distributions offer practical advantages over isotropic noise and strong deterministic predictors. We introduce a graph-structured residual flow framework that separates conditional mean prediction from stochastic residual transport. A history-only predictor estimates the future connectivity graph, while a graph Gaussian source encodes dependencies derived from past connectivity through a Laplacian-based covariance. A conditional velocity field transports source samples to future graph residuals, with transport time explicitly distinguished from physical EEG time. Our study identifies the conditions and controls needed to distinguish useful residual transport from improvements attributable to deterministic prediction, learned representations, and sampling effects.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, Flow matching, Brain network, continuous dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_PdqHOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates how to use EEG graph structures, then construct effective representation learning tasks for predicting future node features and functional connectivity. GRFBrain uses the historical node mean and output of frozen graph predictor as reference states. Constructs a Gaussian source using the historical graph Laplacian. And then employs coupled conditional flow matching and joint node–edge transport. The method is evaluated on abnormal EEG classification, using the TUSZ and TUAB datasets and compared with conventional EEG models, discrete graph dynamics methods, continuous-time models, and general-purpose flow-based predictors. The main findings are based on a joint node–edge velocity network, joint supervision of nodes and edges, and a short-term prediction objective with a 1-second prediction horizon.

Critical strengths

S1. The paper designs the prediction task using historical inter-channel relationships. It constructs a sample-dependent noise correlation structure from the historical EEG graph and controls the total noise energy. The conditional graph source achieves an F1 score of 0.523, higher than the IID source and the global graph source. This design addresses the inter-channel dependencies ignored by generic Gaussian sources. S2. Node–edge interaction updates channel activity and functional connectivity through bidirectional interactions within a single network. The full model achieves F1 scores of 0.523 and 0.488 with 12 s and 60 s inputs, respectively. Both exceed the corresponding scores of the strongest unidirectional variant. The consistent gains across both input lengths support the role of bidirectional interaction in learning representations for classification.

Critical weaknesses

W1. Key evaluation procedures are not adequately described, making the main results difficult to verify. Page 6 does not specify which dataset was used to select the threshold that maximizes F1. These omissions make the performance comparisons in Table 3 difficult to assess. The paper should report the numbers of subjects, and recordings in each split, and specify the data used for pretraining.

Questions for authors

Were TUSZ and TUAB first divided into subject-disjoint training, validation, and test sets? Were any validation or test recordings used during pretraining or training with the predictor frozen? Which dataset was used to determine the optimal threshold mentioned on page 6? Was this threshold then fixed for testing? Did all methods follow the same rule?

1 comment

We appreciate your time and the recognition of the strength of our work. Below we address each weakness and question with additional appendix that we will incorporate into the revised version.

Regarding W1 & Q1, insufficient description of the data partitioning and evaluation protocol has affected the verification of results.

Sorry for confusing with insufficient description of the data partitioning and evaluation protocol,now we have added more details in appendix.

Specifically, TUSZ uses subject-disjoint train/dev/test sets with 527/61/45 subjects. EEG is aligned to a fixed 19-channel order, resampled to 200 Hz, and segmented into non-overlapping 12s and 60s clips. Each one-second window is represented by 100-dimensional log-amplitude FFT features, normalized using fixed training-set statistics. Thresholds are selected on the dev dataset and fixed for testing, without re-optimization. For TUAB, we follow the official data splitting. The EEG data preprocessing refers to existing works of BIOT (NeurIPS 23), EvoBrain (NeurIPS 25), and ODEBrain (ICLR26) for fair comparisons.

For TUSZ and TUAB, the downstream threshold was selected on the development split, separately for each model and random seed. For each development-selected checkpoint, we collected the predicted probabilities and ground-truth labels for all development clips, constructed the precision–recall curve, and selected the candidate threshold with the highest F1 score. All methods evaluated through our unified TUSZ pipeline followed this same selection rule.

Reviewer_agsbOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, GRFBrain is put forward, and EEG dynamic modeling is a framework for historical condition prediction and residual transmission of graphic structure. Firstly, the history-only predictor is used to get the definitive reference of the future connection graph, then the future state is decomposed into reference and residual, and then the graph Gaussian source constructed by the Laplacian covariance of the historical function graph is used to transmit the source samples to the future node-edge residual through the joint node-edge velocity field. The transmission time is clearly distinguished as artificial transport time, not physical EEG time.

Critical strengths
  1. Problem decomposition and ablation design are targeted.
  2. It is better than strong baseline in some settings, especially TUAB multi-index and TUSZ F1.
Critical weaknesses

The gain from the strong deterministic baseline is limited and the indicators are inconsistent, and the statistical significance is insufficient.

Questions for authors
  1. Please provide reference predictor only, deterministic zero-residual with the same architecture, and complete random transmission readout for controlled ablation.
  2. Is the gain of Conditional GGRF relative to IID/global GGRF statistically significant? Please report the results of paired test, confidence interval and multi-seed/multi-partition.
  3. AUROC does not exceed ODEBrain on TUSZ. How does the author explain the practical advantages of "distributional forecasting"?
  4. The two GRFBrain/ODEBrain lines in Table 3 correspond to multi-step and single-step respectively, please clearly mark them in the table, and supplement the details of data set division, preprocessing, class imbalance treatment and multi-step objective.
3 comments

We appreciate your time and the recognition of the strength of our work. Below we address each weakness and question with additional analysis and clarifications that we will incorporate into the revised version.

Regarding W1, the gain from the strong deterministic baseline is limited and the indicators are inconsistent, and the statistical significance is insufficient.

Thank you for your comments. The main gain of our work comes from stochastic modeling to model continuous dynamics over graph structures. By formulating EEG dynamics as a conditional distribution learning task, the model can effectively handle substantial noise, randomness, and fluctuations. Our improvements arise from two key aspects: (i) obtaining a spectrally structured source distribution, and (ii) formulating a context-rich and stable conditional velocity field. The conditional modulator adaptively adjusts the path distribution and learned transport trajectories, mitigating the classical path-transport problem in Flow Matching and improving long-term stability.

We have also added new log-spectral amplitude distributions and a new table (Figure 6 and Table 5 in the revised manuscript) to show the Jensen–Shannon divergence (JSD) between real and generated log-spectral amplitude distributions on TUSZ. We compared with JET [1], which focuses on using the class-label conditioning velocity network for the flow matching-based EEG modeling.

Table 5 Comparision on real and generated log-spectral amplitude distributions on TUSZ

ModelTypeSeizure ↓Non-Seizure ↓
JETClass-guided0.1180.171
GRFBraindeterministic0.0710.083
GRFBrainVanilla IID0.0450.048
GRFBrainConditional GGRF0.0130.018

[1] Wang, Yifan, et al. "Let eeg models learn eeg." ICML 2026.

Regarding Q1, please provide reference predictor only, deterministic zero-residual with the same architecture, and complete random transmission readout for controlled ablation.

We have added two new tables: one (Table 4 in the revised manuscript) is for full model comparison with supervised-only and deterministic-forecast training, and one (Table 6 in the appendix) is for the ablation of backbone architectures.

Table 4 full model comparison with supervised-only and deterministic-forecast training.

TypeModelAUROCF1
DeterministicGraph-SDE0.818±0.0030.457±0.018
ODEBrain0.881±0.0060.496±0.017
GRFBrain¹0.834±0.0060.475±0.020
GRFBrain²0.857±0.0080.493±0.0016
GRFBrain³0.864±0.0090.509±0.012
StochasticGRFBrain0.877±0.0030.523±0.014

¹ Supervised-only;² Deterministic forecasting;³ Neural ODE framework

Regarding Q2, Is the gain of Conditional GGRF relative to IID/global GGRF statistically significant?

Thank you for your comments. We have added a new table (Table 7 in the appendix) to further show the contribution of the stochastic source geometry within the same conditioned velocity network setting.

Table 7 Comparison of the stochastic source geometry with 12s EEG signals on TUSZ data.

11 seconds past length, 1 second forecasting

MethodAccF1AUROC
Vanilla IID0.875±0.0160.507±0.0140.871±0.007
Global GGRF0.871±0.0020.501±0.0120.868±0.006
Permuted GGRF0.868±0.0060.498±0.0230.871±0.004
Sensor Geometry0.869±0.0120.490±0.0030.864±0.002
Geometry-matched0.872±0.0090.500±0.0180.869±0.008
GRFBrain0.878±0.0010.523±0.0140.877±0.003

3 seconds past length, 19 second forecasting

MethodAccF1AUROC
Vanilla IID0.878±0.0180.486±0.0150.869±0.002
Global GGRF0.874±0.0010.480±0.0110.865±0.004
Permuted GGRF0.872±0.0030.482±0.0210.871±0.005
Sensor Geometry0.877±0.0070.476±0.0070.861±0.003
Geometry-matched0.879±0.0040.489±0.0270.860±0.007
GRFBrain0.883±0.0040.503±0.0120.876±0.001

Regarding Q3, AUROC does not exceed ODEBrain on TUSZ. How does the author explain the practical advantages of "distributional forecasting"?

Thank you for your comments. We selected the downstream checkpoint with the highest development-set average precision score to prioritize positive-class precision–recall performance under a class-imbalanced EEG dataset. For the selected model, we then choose the classification threshold that maximizes the development-set F1 score and keep it fixed for testing. In contrast, ODEBrain retains the checkpoint with the highest development-set AUROC. We confirmed the results of GRFBrain with the same checkpoint as ODEBrain, as shown in the following table.

Table AUROC-selected result with single-step forecasting.

MethodAccF1AUROC
ODEBrain0.877 ± 0.0040.496±0.0170.881±0.006
GRFBrain0.879±0.0070.521±0.0240.886±0.001

Regarding Q4, the two GRFBrain/ODEBrain lines in Table 3 correspond to multi-step and single-step, respectively, please clearly mark them in the table and supplement the details of data set division, preprocessing, class imbalance treatment, and multi-step objective.

Sorry for confusing, we have marked the settings of multi-step and single-step in Table 3. We also provided more details in the appendix. We have added a new table (Table 10 in the appendix) to further show the results with different multi-step objective settings.

Table 10 TUSZ classification results under different pretraining history lengths and forecast horizons over 3 seeds.

SettingAcc ↑AUROC ↑AUPRC ↑F1 ↑Recall ↑
C11–T10.87697 ± 0.0110.87787 ± 0.0030.51589 ± 0.0030.52268 ± 0.0140.61485 ± 0.023
C9–T30.87901 ± 0.0160.87348 ± 0.0050.50662 ± 0.0130.51364 ± 0.0120.58223 ± 0.053
C6–T60.88329 ± 0.0040.87633 ± 0.0010.50275 ± 0.0120.51298 ± 0.0070.56222 ± 0.018
C3–T90.87851 ± 0.0070.86983 ± 0.0090.51768 ± 0.0300.50674 ± 0.0250.57355 ± 0.074

C denotes the pretraining past length in seconds; T denotes the maximum forecast horizon, with one horizon sampled per training example.

Reviewer_VeoKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GRFBrain learns EEG representations through conditional flow matching in a joint node-edge residual space. Future states are centered using historical node features and a frozen edge forecaster. A graph-Laplacian kernel defines an energy-normalized, history-conditioned Gaussian source, and a bidirectional node-edge Transformer predicts transport velocities. The supplied implementation classifies deterministic node features at zero residual and transport time 1, after supervised fine-tuning with an auxiliary flow loss. Experiments on TUSZ and TUAB report gains over several graph/dynamics baselines, with ablations of source geometry, node-edge interactions, transport targets, and prediction horizon.

Critical strengths
  1. A coherent conditional graph construction

The residual formulation separates a history-based reference from future variation, and trace normalization explicitly addresses the confound between source covariance and total noise energy (Eqs. 3–5, pp. 4–5). The paper also correctly distinguishes artificial transport time from physiological time and states that its functional graph is not a causal network. These choices make the proposed representation-learning mechanism interpretable at the modeling level.

  1. Useful downstream and interaction evidence

The single-step model improves TUSZ F1 from 0.496 for ODEBrain to 0.523, and the multi-step model achieves TUAB AUROC 0.878 and F1 0.786 (Table 3, p. 7). Removing bidirectional interaction causes substantial losses (Table 2). The code specifies training-derived normalization, validation-based checkpoint/threshold selection, and a single-evaluation downstream readout at transport time 1. These implementation details resolve important ambiguities in the main text and support a computationally simple downstream use case.

Critical weaknesses
  1. The evidence does not fully isolate the contribution of stochastic source geometry

The conditional source is the distinctive proposal, so I want a comparison that changes its covariance while holding the rest of the model fixed. Equation (6) directly supplies L_C and K_C as edge-token features, and the release implements that path in joint_transformer.py, lines 99–105. K_C therefore affects the network both as a deterministic condition and as the source covariance. Figure 5 needs to state whether the same deterministic features remain available under conditional, global and IID sources. Otherwise that comparison may measure a change in conditioning information as well as source geometry.

The reported conditional-versus-IID F1 difference is 0.523 versus 0.507; the interaction-architecture ablation produces much larger changes. That pattern makes the backbone and fine-tuning controls important. The code fine-tunes the representation with labels and an auxiliary flow loss, so I would compare the same backbone under supervised-only training and deterministic residual forecasting. These tests address whether stochastic source training contributes beyond the architecture and deterministic history predictor. GJI and classification metrics measure useful representations, but do not establish calibrated future node-edge distributions. The manuscript acknowledges that calibration remains untested; the interpretation of its flow visualizations should reflect that scope.

  1. The complete evaluation protocol is still not recoverable from the supplied materials

The release is useful but covers a narrower experiment than the full paper: its reference runner expects 19 channels and a 12-second TUSZ clip. It leaves channel alignment, raw-recording segmentation and labels outside that runner, and supplies an example manifest rather than the retained-patient partitions (docs/data_format.md, lines 11–14). The TUAB, 60-second, multi-step and baseline configurations are also not specified by this default. Those choices can affect small differences in seizure F1 and should be traceable to the reported tables. The code does specify development-set threshold selection and test-time reuse, so that point is resolved.

One remaining correspondence issue is visible in the training path. pretrain.py re-estimates q_X/q_A on training samples and saves them in the checkpoint configuration (lines 53–83). train_downstream.py constructs the model from an external YAML and loads only state["model"] (lines 23–34), while q_X/q_A are ordinary float attributes in system.py, lines 57–59. Consequently those calibrated values are not restored by state_dict unless the external configuration is updated separately. I would like the paper-specific configuration or a corrected restoration path. This matters for the auxiliary loss used during fine-tuning and for a controlled source comparison.

Questions for authors
  1. In Figure 5, do all source variants receive identical L_C/K_C token features and use the same normalization, training budget and fine-tuning schedule? How does the full model compare with supervised-only and deterministic-forecast training on the same node-edge backbone? Please report participant-based paired uncertainty for the conditional-versus-IID difference.

  2. Please provide the EEG appendices and the actual patient partitions, corpus versions, label/segmentation rules and configurations for TUSZ, TUAB, 60-second and multi-step experiments. Do all baselines receive the same observed EEG duration and preprocessing? These records would let me assess the comparative result independently of the reference runner.

  3. Which configuration restores the training-estimated q_X/q_A for downstream fine-tuning? Please reconcile the checkpoint configuration with the external YAML load and give the values used for Table 3 and Figure 5. Does restoring those values change the relative source-geometry result?

2 comments

We appreciate your time and the recognition of the strength of our work. Below we address each weakness and question with additional analysis and clarifications that we will incorporate into the revised version.

Regarding W1, the evidence does not fully isolate the contribution of stochastic source geometry and stochastic source training.

Thank you for your comments. The main gain of our work comes from stochastic modeling to model continuous dynamics over graph structures. By formulating EEG dynamics as a conditional distribution learning task, the model can effectively handle substantial noise, randomness, and fluctuations. Our improvements arise from two key aspects: (i) obtaining a spectrally structured source distribution, and (ii) formulating a context-rich and stable conditional velocity field. The conditional modulator adaptively adjusts the path distribution and learned transport trajectories, mitigating the classical path-transport problem in Flow Matching and improving long-term stability.

We have added a new table (Table 7 in appendix) to further show the contribution of the stochastic source geometry within the same conditioned velocity network setting.

Table 7 Comparison of the stochasitc source geometry with 12s EEG signals on TUSZ data.

11 seconds past length, 1 second forecasting

MethodAccF1AUROC
Vanilla IID0.875±0.0160.507±0.0140.871±0.007
Global GGRF0.871±0.0020.501±0.0120.868±0.006
Permuted GGRF0.868±0.0060.498±0.0230.871±0.004
Sensor Geometry0.869±0.0120.490±0.0030.864±0.002
Geometry-matched0.872±0.0090.500±0.0180.869±0.008
GRFBrain0.878±0.0010.523±0.0140.877±0.003

3 seconds past length, 19 second forecasting

MethodAccF1AUROC
Vanilla IID0.878±0.0180.486±0.0150.869±0.002
Global GGRF0.874±0.0010.480±0.0110.865±0.004
Permuted GGRF0.872±0.0030.482±0.0210.871±0.005
Sensor Geometry0.877±0.0070.476±0.0070.861±0.003
Geometry-matched0.879±0.0040.489±0.0270.860±0.007
GRFBrain0.883±0.0040.503±0.0120.876±0.001

We have also added a new log-spectral amplitude distributions and a new table (Figure 6 and Table 5 in revised manuscript) to show with Jensen–Shannon divergence (JSD) between real and generated log-spectral amplitude distributions on TUSZ. We compared with JET [1], which focuses on using the class-label conditioning velocity network for the flow matching-based EEG modeling.

Table 5 Comparision on real and generated log-spectral amplitude distributions on TUSZ

ModelTypeSeizure ↓Non-Seizure ↓
JETClass-guided0.1180.171
GRFBraindeterministic0.0710.083
GRFBrainVanilla IID0.0450.048
GRFBrainConditional GGRF0.0130.018

[1] Wang, Yifan, et al. "Let eeg models learn eeg." ICML 2026.

Regarding W2 & Q2 The complete evaluation protocol is still not recoverable from the supplied materials.

Sorry for confusing, we have provided the completed appendix.

TUSZ uses subject-disjoint train/dev/test sets with 527/61/45 subjects. EEG is aligned to a fixed 19-channel order, resampled to 200 Hz, and segmented into non-overlapping 12s and 60s clips. Each one-second window is represented by 100-dimensional log-amplitude FFT features, normalized using fixed training-set statistics. A clip is positive if it contains a seizure. Thresholds are selected on the dev dataset and fixed for testing, without re-optimization. The EEG data preprocessing refers to existing works of BIOT (NeurIPS 24), EvoBrain (NeurIPS 25), and ODEBrain (ICLR 26) for fair comparisons.

Regarding Q1, do all source variants receive identical L_C/K_C token features and use the same normalization, training budget and fine-tuning schedule? How does the full model compare with supervised-only and deterministic-forecast training on the same node-edge backbone?

Thank you for your comments, all source variants receive identical L_C/K_C token features and use the same normalization, training budget and fine-tuning schedule.

And we have added two new tables, one (Table 4 in revised manuscript) is for full model comparision with supervised-only and deterministic-forecast training, one (Table 9 in appendix) is for the ablation of backbone architectures.

Table 4 full model comparision with supervised-only and deterministic-forecast training.

TypeModelAUROCF1
DeterministicGraph-SDE0.818±0.0030.457±0.018
ODEBrain0.881±0.0060.496±0.017
GRFBrain¹0.834±0.0060.475±0.020
GRFBrain²0.857±0.0080.493±0.0016
GRFBrain³0.864±0.0090.509±0.012
StochasticGRFBrain0.877±0.0030.523±0.014

¹ Supervised-only;² Deterministic forecasting;³ Neural ODE framework

Table 9 Ablation of Joint backbone depth.

Training strategyBlocksAccAUROCF1
Stochastic20.879 ± 0.0080.872 ± 0.0040.502 ± 0.007
40.866 ± 0.0130.865 ± 0.0140.499 ± 0.022
60.878 ± 0.0010.877 ± 0.0030.523 ± 0.014
80.875 ± 0.0040.867 ± 0.0040.497 ± 0.006
Supervised-only20.866 ± 0.0140.840 ± 0.0200.466 ± 0.039
40.868 ± 0.0060.834 ± 0.0060.475 ± 0.020
60.871 ± 0.0060.826 ± 0.0070.454 ± 0.015
80.870 ± 0.0030.836 ± 0.0100.451 ± 0.033

Regarding Q3, issue of restoring calibration parameters.

(qX/qAq_X/q_A) are training-set estimates of the second moments of the target velocities, used to normalize the node and edge velocity regression losses. They are neither probability calibration parameters nor source noise scales. The downstream training code reads both values directly from the checkpoint. The default YAML in the public reference implementation contains the same values and therefore captures this historical configuration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global response to all reviewers

Dear Reviewers,

We thank you for your thoughtful feedback and constructive suggestions. We appreciate your recognition of:

  • The dynamic graph-conditioned Gaussian source field and its explicit control of total noise energy for flow matching-based continuous brain dynamics modeling (Reviewers PdqH and VeoK).
  • The bidirectional node–edge velocity network design and its contribution to downstream EEG representations (Reviewers PdqH and VeoK).
  • The targeted problem decomposition and experiment design with the improvements observed in several TUSZ and TUAB settings (Reviewer agsb).

In response to the main concerns, we conducted additional experiments and analyses and revised the manuscript, with changes highlighted in red. The core edits are summarized below:

  • Source-geometry comparisons — Reviewers VeoK and agsb. We expanded the source ablations to include IID, global GGRF, permuted GGRF, and static electrode-geometry sources. We clarified that the spectral-structural source design can improve conditional flow matching to efficiently learn the joint distribution through continuous transport.

  • Conditioned velocity network design and comparison between deterministic and stochastic continuous brain dynamic modeling— Reviewers VeoK and agsb. We added supervised-only, deterministic-forecasting, and Neural ODE controls based on the proposed velocity network. We also evaluated the capacity of depths of 2, 4, 6, and 8 blocks under supervised-only and full-framework training. These experiments examine the contributions of stochastic continuous brain dynamic modeling beyond comparisons with existing deterministic approaches.

  • Stochastic dynamics modeling evaluation beyond downstream task— Reviewers VeoK and agsb. We added new visualizations and Jensen–Shannon divergence measurements comparing real and generated log-spectral amplitude distributions, including deterministic forecasting, IID-source Flow Matching, conditional GGRF, and the existing class-label-conditioned stochastic dynamics modeling method.

  • Questions about data separation and evaluation reproducibility — Reviewers PdqH, VeoK, and agsb. We clarified that GRFBrain selects checkpoints by development-set average precision, then selects the classification threshold from the development-set precision–recall curve by maximizing F1. The threshold is fixed for testing. Development data support model selection, and test data does not participate in parameter updates or per-run checkpoint and threshold selection.

  • Temporal length and forecasting horizon — Reviewer agsb. We clarified the single-step and multi-step settings and added a new table result with different temporal settings. These results further show that the proposed graph-conditioned source distribution can improve the robustness of stochastic dynamics modelling.

  • Data processing and evaluation details — Reviewers PdqH and VeoK. We provided more detailed information on the datasets, channel alignment, sampling rate, clip segmentation, log-amplitude FFT representation, and training-statistics-based normalization. We also clarified development-only threshold selection.

  • The role of loss-normalization parameters — Reviewer VeoK.

We welcome further discussion and thank the reviewers again for their time and constructive feedback.

Best regards,

The Authors

AuthorsOpenReview ↗

Final Response

Dear Reviewers and Area Chair,

We sincerely thank the reviewers for their thoughtful feedback and recognition of the importance of stochastic continuous brain dynamics modeling through learning the conditional joint distribution of future node spectrum and functional connectivity. We appreciate the reviewers’ recognition of the graph-conditioned Gaussian source design and the bidirectional node–edge velocity network. We are encouraged that the reviewers consistently viewed the soundness concerns as minor or fixable, and we have conducted additional analysis and revisions carefully following the reviewers’ suggestions.

In response to the reviews, we added experiments that examine what the stochastic dynamics contribute. On seizure detection, GRFBrain achieved the best scores, compared with deterministic forecasting and the existing deterministic continuous dynamics modeling work (Table 4). When the proposed velocity network had access to the same setting, replacing energy-matched isotropic noise with the dynamic graph-conditioned source improved mean F1 scores (Table 7). We also evaluated generated samples directly and found the performance of log-spectral amplitude distributions closer to the observed data, while reducing Jensen–Shannon divergence (Table 5). The revised manuscript also reports sensitivity to network depth and forecast horizon and clarifies how the development set was used to select checkpoints and classification thresholds. These revisions are incorporated into the updated manuscript and highlighted in red.

We understand that reviewers are balancing many commitments and may have limited time for further discussion at this stage. As the rebuttal period concludes, we hope the additional evidence and clarifications provide a helpful basis for considering whether the concerns have been addressed. We would be grateful for their consideration in the final assessment. Thank you all again for your time and for engaging with the review process.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.