acceptodds
Under review as a conference paper at ICLR 2027

One-Step Latent-free EEG Generation with Mean-Field Flow

Abstract

EEG datasets remain limited in size and diversity, leaving substantial waveform variation across subjects, sessions, and recording conditions underrepresented in training. Generative modeling offers a way to address this limitation, but the standard approaches for image or language generation may not transfer directly to EEG. Latent compression may discard amplitude structure, phase relationships, and transient events. Standard self-attention provides no explicit prior for shared cortical sources underlying cross-electrode correlations. Multi-step samplers require repeated inference, increasing the generation cost. In this work, we propose Mean-Field Flow (MFF), a one-step, latent-free method for raw EEG generation. MFF operates directly on raw EEG and uses mean-field token coupling to model channel and temporal interactions through separate shared low-dimensional fields. We train MFF with an average-velocity objective via regression, reducing sampling to a single forward pass. Across 14 public EEG benchmark datasets, MFF achieves the lowest average TS-FID, reducing it by 27.8% relative to the strongest baseline. It also improves downstream brain-computer interface (BCI) classification accuracy by 2.80 percentage points on average, and its generated samples more closely match real EEG in Hjorth parameters, skewness, and scalp topographies. We hope that this study will motivate broader exploration of generative modeling for electrophysiological signals.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_AFKKOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce Mean-Field Flow (MFF), a class-conditional EEG generator trained separately on each of 14 datasets. It combines the idea of one-step flow generation with mean-field token coupling. In the experiments, MFF showed the best quality of synthetic EEG data and could act as a data augmentation tool that improves decoding performance.

Critical strengths
  1. The proposed latent-free, one-step generative framework operates directly on raw EEG. The reported inference efficiency is particularly impressive, which could be useful when large quantities of synthetic EEG are required.
  2. The authors tested quite a lot of datasets and models, thereby giving convincing empirical results.
Critical weaknesses
  1. Limited depth of scientific findings and neurophysiological interpretation that are useful for the broad community.

I acknowledge the extensive evaluation across diverse EEG datasets. I feel that the work could have established substantially more meaningful scientific findings. The experiments primarily demonstrate improvements in statistical similarity and downstream classification performance, but provide limited insight into what physiologically meaningful information the generator actually learns. Although the authors evaluate spectral statistics, phase locking, and ERD-related spatial patterns, these analyses remain relatively superficial. It is unclear whether the generated signals preserve task-specific neural characteristics, distinguish meaningful neural activity from recording artifacts, or offer new insights into EEG physiology. Given the potential of generative modeling, I believe this represents a significant missed opportunity.

  1. Unclear practical utility and conditions under which synthetic EEG is beneficial.

The motivation of the paper is closely related to EEG data scarcity, yet the experiments do not establish when synthetic EEG generation is practically useful or how much real data it can replace. In particular, EEG exhibits substantial variability across subjects, sessions, and recording conditions. The benefit of synthetic augmentation may depend strongly on the available real-data quantity and quality, task characteristics, and whether the generated signals represent the target subject or a broader population. While the authors provide useful augmentation controls and investigate synthetic-data ratios, these experiments do not establish whether synthetic EEG is preferable to collecting additional independent real recordings, or under which circumstances the proposed approach offers meaningful data-efficiency improvements. For the field, I think people are more interested in knowing in which cases one should simply record more data from the subject/patient or when such a generative model could be used.

  1. Limited generalizability across datasets and decoding approaches.

Although the paper evaluates 14 diverse EEG datasets, the generative model is trained independently on each dataset. Consequently, the proposed framework does not exploit shared knowledge across subjects, tasks, or recording configurations. Its reliance on dataset-specific channel ordering further limits its applicability across electrode montages and devices. Additionally, downstream task performance gain is primarily demonstrated using neural-network classifiers, particularly pretrained CBraMod. It remains unclear whether the augmentation benefits extend to classical EEG decoding approaches or stronger foundation models. These limitations make it difficult to assess the broader applicability of the proposed framework beyond the specific datasets and evaluation configurations considered.

Questions for authors
  1. Beyond matching statistical distributions, can the authors provide more detailed analyses of whether generated EEG preserves established task-specific neurophysiological markers? For example, does the model learn channel-specific comparisons of class-specific dynamics, spectral differences between subjects/ tasks, hemispheric lateralization, or clinically meaningful waveform characteristics between abnormal and normal EEG?

  2. How do the authors evaluate the quality and physiological informativeness of individual generated samples? Since the training datasets may contain recording artifacts and noisy signals, how can one distinguish realistic neural activity from realistic artifacts? Could the proposed framework also help identify high-quality EEG or reveal meaningful differences between normal and abnormal conditions?

  3. How do generation quality and downstream effectiveness vary with the quantity and quality of real training data? The evaluated datasets differ substantially in size and task characteristics, but this does not establish how much real data is necessary to train a useful generator. Could the authors provide controlled data-scaling experiments and investigate whether the benefits depend on task complexity, subject diversity, and data quality?

  4. Can synthetic EEG meaningfully substitute for additional real recordings? For example, how does training with 80% real data plus 20% synthetic data compare with using 100% real data, when the generator is trained only on the initial 80%? Given inter-subject variability, it would also be informative to compare additional real or synthetic data originating from the target subject versus other subjects, ideally across different calibration budgets.

  5. Why is the model trained independently on each dataset rather than learning a shared EEG representation across datasets? Could the authors investigate whether pretraining on larger EEG corpora enables transfer or few-shot adaptation to smaller datasets, potentially with electrode-aware representations that generalize across montages? I think a general model would be more useful to people working on different downstream tasks.

Reviewer_jqEpOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

One-Step Latent-free EEG Generation with Mean-Field Flow proposes a flow matching model for generation of synthetic EEG data using a mean-field approximation of the shared latent across both space and time. The resulting model shows substantial improvements in terms of FID with real data and also provides a notable uplift in downstream task accuracy.

Critical strengths
  1. The paper is clear and well-written. I particularly appreciate the use of appropriate colors and highlight blocks.
  2. The contributions are clearly presented and are novel in both conception and execution.
  3. The mean-field assumption is clever and well-justified, albeit its execution as concatenation is somewhat simplistic. Nonetheless, it is effective.
  4. The experiments are well-made and extensive, and the results clearly corroborate the effectiveness of the MFF approach.
  5. The use of both reconstruction metrics and downstream accuracy paints a complete picture, and the ablations also serve to reinforce the message.
Critical weaknesses
  1. The L1 loss is unreasonably effective, and the ablations with Huber pinpoint the effect but do not explain it; I would like to see more emphasis placed on this aspect
Questions for authors
  1. I would like to see a stronger explanation or practical ablation for the effectiveness of the L1 loss
3 comments

Dear Reviewer jqEp,

Thank you again for your constructive feedback and positive assessment. We have incorporated the explanation of L1’s effectiveness and the additional loss comparisons into Section 3.4 and Appendix C, with changes marked in blue. We hope these additions fully address your question and reinforce your positive assessment. We greatly appreciate your time and support.

Best,

The Authors

My concerns have been addressed, and I have independently verified the effectiveness of L1. I will raise my score accordingly.

We sincerely thank you for the thoughtful review and support for our paper! Your constructive feedback has helped us strengthen the analysis and improve the clarity and quality of our paper. We greatly appreciate your time and positive assessment.

Reviewer_PdeEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose Mean-Field Flow, a class-conditional generator of raw EEG that samples in a single forward pass. To this end, self-attention along channels and time is replaced by a coupling layer in which tokens interact through a shared low-dimensional summary, and the model is trained with an L1 average-velocity objective adapted from mean flows. One model is trained per dataset on 14 datasets and evaluated with TS-FID, downstream augmentation, and signal statistics.

Critical strengths

First, the evaluation of the generated EEG is broad, including downstream augmentation on real test data with several classifiers, signal-level statistics, and controls showing that the gains disappear with duplicated data and reverse with permuted labels. Hence, the claim of downstream utility is considered credible.

Second, the one-step generator is practical since it samples faster and the comparison with flow matching and a consistency model supports the value of the formulation for augmentation.

Critical weaknesses

First, I am not convinced that the objective actually learns the mean velocity. The mean velocity should depend on both endpoints, yet the network is not conditioned on the end time r, which is only used as an additional query. The target also approximates the time derivative at a fixed point and leaves out the transport term. The same applies to the guidance scale and interval, which are not given to the network (Appendix A.2), even though Table 5 reports values for them at inference. It is therefore unclear whether the one-step sample really transports noise to data and it may simply act as a one-step regressor, which Section 3.4 does not justify. In addition, the MeanFlow version in Appendix C only changes the loss, so the paper does not compare against a proper MeanFlow objective.

Second, the architecture seems to reuse an existing block, and the main evaluation metric is not properly validated. Eqs. 10–12 closely follow the CoTAR block of Yu et al. (2026), yet the paper cites this work only for the dual-axis design, and Theorem 1 is essentially a definition. TS-FID uses a hand-crafted magnitude spectrum that ignores phase and fine spatial structure. Unlike the band-power and PLV analysis in Appendix E.1, it has no real-versus-real reference, and without repeated generator trainings, values ranging from about 80 to over 1,100 are hard to interpret.

Questions for authors

How does the coupling layer differ from CoTAR? It would help to clarify what the objective converges to when the network is conditioned on neither r nor the guidance scale, and to compare it with a proper mean-flow objective conditioned on both time points.

For evaluation, report a real-versus-real TS-FID for each dataset, confidence intervals over repeated generator trainings, and possibly a nearest-neighbor memorization analysis. It is also not clear whether downstream accuracy is plain or balanced, which matters given the class imbalance of several datasets. Finally, a comparison with recent EEG generators that the paper cites, such as JET, and exact parameter and compute counts for the comparison with self-attention would be informative.

3 comments

Dear Reviewer PdeE,

Thank you again for your time and careful review. We have incorporated our clarifications and additional results into Sections 3.2–3.4 and Appendix K, with changes marked in blue. We hope these revisions address your methodological and evaluation concerns, and we welcome any further comments. If your concerns have been resolved, we would be grateful if you could reconsider your assessment and score.

Best,

The Authors

Thank you for the detailed response and the additional experiments, which address my main concerns. I will increase my score.

We sincerely thank you for the thoughtful review and support for our paper! Your constructive feedback has helped us strengthen the analysis and improve the clarity and quality of our paper. We greatly appreciate your time and positive assessment.

Comments on the paper

1 comment

Dear Area Chairs,

As the discussion period draws to a close, we sincerely thank you for coordinating the review process. We also appreciate the reviewers' careful evaluation and constructive suggestions, which have helped us strengthen the manuscript.

We are encouraged that the rebuttal substantially addressed the major concerns raised during review, and that all three reviewers increased their overall recommendations after discussion.

We would like to briefly highlight several strengths of the paper that were explicitly recognized by the reviewers:

  • Reviewer PdeE recognized the broad evaluation, including downstream augmentation, multiple classifiers, signal statistics and augmentation controls, considered the downstream utility credible, and highlighted the practical value of one-step generation.
  • Reviewer jqEp described the contributions as novel in both conception and execution, considered the mean-field design effective, and found that the extensive experiments and ablations clearly corroborated MFF's effectiveness.
  • Reviewer AFKK highlighted the impressive inference efficiency of the latent-free, one-step framework and viewed the evaluation across datasets and models as convincing empirical evidence.

We also summarize below the additional clarifications and experiments provided during the rebuttal to address the reviewers' concerns, as reflected in the reviewers' final comments and updated scores:

  • Additional Clarifications:

    • We clarified the fixed-state bootstrap objective, the roles of the sampled endpoint and guidance variables, and the connection between the learned endpoint prediction and one-step generation (Reviewer PdeE).
    • We made the CoTAR operator attribution explicit alongside the equations, detailed MFF's conditional dual-axis fields and dense velocity decoding, and presented the shared-field statement as a structural proposition (Reviewer PdeE).
    • We derived the coordinatewise gradients under detached adaptive weighting to explain how exact L1 and Huber emphasize small residuals differently (Reviewer jqEp).
    • We clarified balanced accuracy, complementary phase/spatial evaluation, nested training subsets and target-subject selection, and developed practical guidance on when synthetic augmentation is useful (Reviewers PdeE, AFKK).
  • Additional Experiments:

    • We added endpoint-displacement diagnostics and comparisons with original MeanFlow using both endpoint inputs and the full material derivative, including its L1 variant, on three datasets (Reviewer PdeE).
    • We compared exact L1 with Huber on TUEV, BCI IV-2a and ISRUC and added a normalized-Huber threshold control, showing consistent improvements in TS-FID and Hjorth-complexity fidelity with L1 (Reviewer jqEp).
    • We provided real–real TS-FID references for all 14 datasets, confidence intervals from five independent generator trainings per dataset on three representative datasets, JET comparisons, and parameter and FLOP counts against a matched self-attention variant (Reviewer PdeE).
    • We added four task-specific physiological contrasts covering epileptiform morphology, motor-imagery lateralization, sleep stages and abnormal–normal EEG, class-specific sample evaluation, and an exploratory analysis linking signal fidelity to augmentation gains across 14 datasets (Reviewer AFKK).
    • We extended data-scaling experiments to 5% and 10% training budgets on BCI IV-2a and TUEV, reporting absolute sample counts and paired-run variability. We also evaluated training-data quality, subject diversity, classical decoding and cross-dataset transfer (Reviewer AFKK).
    • We compared 80% real + 20% synthetic data with 100% real data on three datasets, extended target-subject calibration to all nine BCI IV-2a subjects, and added cross-subject standard deviations to Table 33. Nearest-neighbor memorization analyses cover TUEV and a BCI IV-2a generator trained with 25% real data (Reviewers PdeE, AFKK).

All three reviewers increased their scores after the discussion, expressing satisfaction with our rebuttal and recognition of the paper's contribution:

  1. Reviewer PdeE stated that the detailed response and additional experiments “address my main concerns,” and increased the score from 2 to 3.
  2. Reviewer jqEp wrote, “My concerns have been addressed,” independently verified the effectiveness of L1, and increased the score from 3 to 4.
  3. Reviewer AFKK wrote, “My concerns are addressed,” appreciated our open reporting of the low-data limitations, and increased the score from 2 to 3.

Thank you for considering our revised manuscript and the outcome of the discussion. We appreciate the time and care devoted to this review process.

Best,

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.