acceptodds
Under review as a conference paper at ICLR 2027

SiDe: Long-Sequence Electrophysiological Generation with Raw-Signal Detail Recovery

Abstract

Electrophysiological recordings often become long sequences due to long temporal spans or high sampling rates. Yet most biosignal generation studies remain in short-window settings, and scaling them to long sequences is not a simple extrapolation. Without compression, modeling a large number of raw time points in the full signal space quickly becomes computationally expensive; with front-end latent compression, high-frequency statistics and local signal details may be suppressed before generation begins. We study where compression should occur: long-range structure can be modeled in compressed tokens, but the final raw-space velocity should not be forced to recover local detail only through these tokens. We propose SiDe, a Signal-space Detailer that keeps compression in a conditional Transformer backbone while providing an output-side raw-signal path for velocity prediction. On pose-conditioned EMG generation, SiDe reduces rFID from 7.363 to 2.414 compared with a JiT-style baseline, and better preserves frequency profiles and common EMG statistical feature distributions. The generated EMG also follows pose-speed manipulations and matches real-EMG degradation trends under controlled perturbations. Additional experiments on Sleep-EDF EEG and PTB-XL ECG further characterize where raw-signal detail recovery helps across modalities.

Primary area:
generative models
Keywords:
Electrophysiological Signal Generation, Biosignal Generation, Long-Sequence Modeling, Flow Matching, Generative Modeling, Surface Electromyography (sEMG), Signal Compression, Diffusion Transformers

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8jSuOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on how to efficiently generate electrophysiological signals with a large number of sampling points while preserving local waveform details and spectral characteristics. The author proposes SiDe (Signal-space Detailer), which divides the computational process into two parts: the compressed Transformer backbone is responsible for the context and conditional information of the entire signal; the local one-dimensional convolution detailer simultaneously reads the features of the backbone and the noisy original signal, and predicts the update direction in the original signal space. The core idea is: compress the calculation in the backbone while retaining the direct path from the original signal to the output end, reducing the limitation of compressed representation on detail restoration. Main contributions:

  1. Proposes a local detail restoration structure suitable for electrophysiological signals.
  2. Systematically analyzes the problems caused by compression.
  3. Establishes a comprehensive verification combination. The most fully validated aspect is the quality of local signals and conditional responses; the true long-term dependencies, as well as whether the synthetic signals can be reliably used for the robustness evaluation of the real decoder, still require more direct experiments.
Critical strengths
  1. The performance improvement on the main task is significant and is supported by the strengthened baseline and various evaluation methods. This is the primary basis for my consideration: The paper has made clear progress in tasks with high sampling rates and multi-channel generation, and the main benefits have consistency across metrics and evaluators.
  2. The effect on the original signal pathway was subjected to a discernible controlled verification, forming a design basis that can be adopted by subsequent research. This point influenced my overall assessment because the paper provided operational experience on how the electrophysiological generation model arranges compression and detail restoration, increasing its contribution value to the community.
Critical weaknesses
  1. The core claim of long-term dependency lacks experiments that can directly test it. The main signal statistics of EMG are calculated on 0.5-second segments; although Sleep-EDF generates 5 minutes and 10 epochs at once, they are split into individual 30-second epochs for evaluation. Therefore, the existing results sufficiently prove the quality of local signals, but cannot rule out the possibility that "short segments are independently generated and then concatenated to achieve similar indicators". This gap directly limits the contribution strength of the paper in the "long sequence modeling" aspect, and is the most significant negative consideration.
  2. Compared with the existing original space generation methods, the contribution of the newly designed mechanism has not been fully separated. The comparison between SiDe and JiT involves changes in the output structure and the parameterization of predictions; while the bottleneck ablation mainly proves the importance of the original signal pathway, it cannot fully explain why SiDe is superior to JiT which already has an explicit original pathway. This enables me to recognize the empirical effect of the final architecture, but I still have doubts about its incremental independent method and the benefits proposed in the paper, and thus is not sufficient to give a stronger acceptance recommendation.
Questions for authors
  1. Among the current results, which pieces of evidence directly support "learning long-term dependencies or epoch-crossing relationships"? The main signal statistics of EMG are calculated on 0.5-second segments, while Sleep-EDF evaluates on a single 30-second epoch. In the existing analysis, is there any evidence that can distinguish "joint generation of the entire segment" from "concatenation of independent short segments"? If not, should the "long-range/cross-epoch dependencies" mentioned in the text be understood as the modeling capability of the architecture rather than a conclusion that has been experimentally verified?
  2. Can the existing comparison in Table 2 confirm the independent benefits of the specialized output structure of SiDe? Considering that the rFID gap between SiDe and ResNet-3/4 is small, does the existing evidence support a stable improvement in spectral recovery or a broader improvement in generation quality?
2 comments

Thank you for the author's addition of targeted experiments. Regarding cross-epoch dependencies, the newly added joint generation and stage-matching independent concatenation control, as well as the classification analysis near the stage transition, substantially address my main concerns. These results provide direct evidence that was lacking in the original manuscript, supporting the claim that joint generation can retain the cross-epoch statistical correlations and transition-related features that independent concatenation cannot. I also agree with the author's report on the smoothing of time. The revised manuscript should retain this limitation and confine the conclusion to the statistical dependencies actually supported by the experiments.

Regarding the contribution of the specific detailer, the author clearly positions its main benefit as spectral restoration and acknowledges that it has a relatively small difference from rFID of ResNet-3/4. This explanation is consistent with the results in Table 2, resolving my doubts about the scope of the contribution. The new results of SiDe-64 also demonstrate a valuable trade-off between computational cost and generation quality, further enhancing the practicality of the method.

Based on the original manuscript and this response, I maintain the recommendation of weak acceptance. The author has clearly alleviated the most significant evidence gap and more accurately defined the benefits of the specialized output structure, so my support for acceptance is more sufficient than in the initial review. Please incorporate the new experiments, necessary implementation details, and identified limitations into the revised manuscript.

Thank you very much for your careful follow-up and for taking the time to examine our additional experiments.

We are glad that the new controlled experiments addressed your concerns regarding cross-epoch dependencies and clarified the specific contributions of the Detailer. We also appreciate your recognition of the quality-efficiency trade-off demonstrated by SiDe-64.

We will incorporate the additional experiments, implementation details, and the identified limitation regarding temporal over-smoothing into the revised manuscript, while keeping our claims aligned with the supporting evidence.

Thank you again for your constructive feedback and thoughtful evaluation. We sincerely appreciate your support.

Reviewer_4kXQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the trade-off between computational efficiency and raw-signal fidelity in long-sequence electrophysiological signal generation. The authors argue that although compressed token representations reduce computational costs, they may restrict the flow of high-dimensional local information to the final velocity prediction, resulting in a so-called velocity-path bottleneck.

To address this issue, the authors propose SiDe (Signal-space Detailer), which employs a compressed conditional Transformer backbone to model long-range temporal dependencies and a local detailer that directly accesses the noisy raw signal to predict flow-matching velocities in the original signal space. This design aims to improve the recovery of local waveforms and high-frequency signal statistics while retaining the computational benefits of a compressed backbone.

The authors investigate the proposed mechanism through synthetic experiments, bottleneck analyses on real EMG signals, and VAE spectral diagnostics. On EMG2Pose, SiDe reduces rFID from 7.363 with the JiT-style baseline to 2.414. Additional experiments on Sleep-EDF EEG and PTB-XL ECG explore the method's applicability across electrophysiological modalities.

Critical strengths

S1. Clear motivation and a well-defined, testable research problem.

The paper addresses information bottlenecks in high-dimensional raw-signal generation and investigates the problem from three perspectives: spectral characteristics, VAE reconstruction distortions, and velocity prediction pathways. These analyses are closely connected to the proposed architecture, providing a coherent motivation for the method.

S2. Systematic experimental design with convincing results on the primary task.

The authors investigate the proposed mechanism through synthetic experiments, patch-size and hidden-width sweeps, raw-path bottleneck tests, and detailer ablations. SiDe achieves substantial improvements across multiple generation-quality and signal-level metrics on EMG2Pose. Additional analyses of random-seed variability and sensitivity to frozen evaluation models further strengthen the empirical evidence.

S3. Discussion of the method's applicability and limitations.

The paper includes supplementary experiments on EEG and ECG and acknowledges that SiDe underperforms JiT on certain ECG morphology metrics. This provides useful insight into how the benefits of local detail recovery depend on signal characteristics and evaluation objectives.

Critical weaknesses

W1. Limited Architectural Novelty and Insufficient Differentiation from Prior Work

The core architecture of SiDe closely resembles existing pixel-space generative approaches. DiP already combines a compressed Transformer backbone with a patch detailer for local detail recovery, while DeCo similarly employs a global–local decomposition to recover details in the original signal space. In comparison, SiDe primarily adapts this design to one-dimensional electrophysiological signals using a dilated ResNet and conditional modulation for local velocity prediction. Its independent methodological novelty therefore remains unclear.

Although Table 2 compares SiDe with one-dimensional adaptations of DiP, DeCo, and PixelDiT, the paper does not sufficiently clarify whether these detailers are comparable in terms of parameter count, computational budget, receptive field, and adaptation strategy. Consequently, the current results do not clearly distinguish the benefits of a genuinely new mechanism from those of one-dimensional convolutional inductive biases and implementation choices. The paper also does not fully establish what new mechanistic insights it provides beyond prior work.

W2. Incomplete Theoretical Characterization and Insufficient Isolation of the Proposed Mechanism

The paper provides controlled experiments and high-dimensional velocity-target analyses to support the existence of a velocity-path bottleneck, but its theoretical characterization remains incomplete. A raw patch having more dimensions than its corresponding token does not necessarily imply that the optimal conditional velocity field requires an equally high-dimensional internal representation. In flow matching, the per-sample velocity target is also distinct from the conditional expected velocity field learned by the model. Furthermore, the velocity definition in Section 3.2 and the x-prediction conversion formula in Section 3.3 require a consistent time parameterization to clarify the velocity direction and the claimed error amplification near the high-noise endpoint.

The existing ablations demonstrate the importance of a raw-space information path but do not sufficiently isolate the benefits of bypassing the compression bottleneck from those of using a more complex learnable detailer. JiT already preserves raw-input information through a fixed residual path, while SiDe's additional improvements may arise from local inductive biases, parameterization choices, or increased computational cost. Controlled comparisons with simpler raw shortcuts or learnable residual gating mechanisms would therefore help establish the independent necessity of the proposed dilated ResNet detailer.

W3. Insufficient Evaluation of Computational Scalability and Quality–Efficiency Trade-offs

Computational efficiency is a central motivation of the paper, yet SiDe introduces substantial additional computational overhead. According to Table 9 in Appendix E, SiDe requires 2.92 TFLOPs per forward pass and 17.47 GiB of memory, compared with 0.817 TFLOPs and 7.81 GiB for JiT. Although the authors include JiT-768 as an approximately compute-matched baseline, the paper still lacks a systematic quality–efficiency analysis covering actual inference latency, throughput, and performance across different computational budgets.

Moreover, a large number of raw samples does not necessarily demonstrate the ability to model long-range temporal dependencies. The primary experiment uses six-second EMG windows, with only 150 target tokens processed by the Transformer, while standard attention still contains a quadratic computational component with respect to token count. The paper does not systematically evaluate generation quality and computational scaling across increasing sequence lengths. The Sleep-EDF evaluation also focuses mainly on epoch-level metrics, providing limited evidence of cross-epoch temporal consistency. Therefore, the current results do not sufficiently substantiate broader claims regarding long-context scalability and long-range temporal modeling.

W4. Limited Validation of Generalization and Physiological Fidelity

Although the main EMG2Pose experiments provide relatively comprehensive test-set results, the supplementary Sleep-EDF and PTB-XL experiments primarily report validation-set performance, without independent test-set evaluation and stability analyses comparable to those of the main task. Furthermore, while SiDe improves PSD fidelity on ECG, it underperforms JiT on several heartbeat detection and morphology metrics, indicating that its benefits are task-dependent. Similarly, the VAE spectral diagnostics in Section 2 are based on the evaluated compressors and EMG setting, and their conclusions cannot necessarily be generalized to all latent-space generative approaches.

In addition, EMG conditional consistency is primarily evaluated using a frozen regression model. A lower angle MAE for generated signals does not necessarily imply greater physiological realism. The highly correlated degradation curves between real and generated signals in Figure 6 mainly demonstrate similar degradation trends, rather than agreement in absolute error levels or the full conditional distributions. Therefore, broader conclusions regarding cross-modal generality, realistic physiological simulation, and the use of generated signals as substitutes for offline robustness evaluation should be interpreted more cautiously.

Questions for authors

Q1. Novelty and Prior Work

What are the key architectural innovations of SiDe compared with DiP and DeCo? Beyond adapting a two-dimensional local recovery module into a one-dimensional dilated ResNet, does the paper introduce a new modeling mechanism or insights specific to electrophysiological signals? Are the one-dimensional baseline adaptations comparable in terms of parameter count, computational cost, and hyperparameter tuning budgets?

Q2. Velocity-path Bottleneck and Detailer Necessity

Could the authors further clarify the relationship between the high effective dimensionality of velocity targets and the representation requirements of the optimal conditional velocity field? Could they also specify the time interpolation and velocity-direction conventions used in flow matching? In addition, could comparisons with simpler raw shortcuts, learnable residual connections, or local gating mechanisms help disentangle the contributions of the raw-input path and the more complex detailer?

Q3. Scalability and Computational Efficiency

How do SiDe's actual inference latency, peak memory usage, and generation quality change with different sequence lengths and patch sizes? Could the authors provide a more comprehensive quality–efficiency comparison and clarify the model's scalability to longer temporal contexts?

Q4. Cross-modal and Physiological Validation

Could the authors provide independent test-set results for the EEG and ECG experiments and further analyze cross-epoch temporal continuity and conditional distribution consistency? For EMG, how do the authors distinguish downstream decodability from physiological realism? Additionally, do the conclusions of the VAE spectral diagnostics hold across different compressor architectures?

2 comments

Thank you for the detailed rebuttal. The additional experiments have addressed several of my concerns, particularly regarding W2 and W4. I acknowledge the authors' clarification that the main contribution is a unified mechanistic interpretation rather than architectural novelty. However, whether this interpretation constitutes sufficient methodological novelty remains unclear. Furthermore, the main SiDe model still shows lower inference throughput than several baselines. Given these remaining concerns, I will maintain my original rating.

Thank you for your follow-up and for acknowledging the additional evidence addressing several of your concerns. We appreciate your careful consideration of our rebuttal and the detailed feedback throughout the review process. We will incorporate the clarifications and additional experimental results into the revised manuscript.

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

SiDe is a Signal Detailer architecture to conditionally compress raw signal space while preserving the task-specific signal characteristics. The aim is to preserve local waveform and spectral details.

Critical strengths
  • SiDe architecture seems to show solid improvements across baselines and metrics in Table 1, supporting the overall design.
Critical weaknesses
  • Evaluation centres only on EMG2Pose and ECG improvements seem mixed. Could the authors extend to more datasets/scenarios to show the generality of their method?

  • Practical efficiency is insufficiently established. (Appendix's Table 9 reports FLOPs and memory, but SiDe costs more than JiT; end-to-end generation latency and throughput are missing and should be a primary evaluation candidates as motivated in Figure 1).

Questions for authors

(See Weaknesses please)

Comments on the paper

1 comment
Area_Chair_gPhtOpenReview ↗

Discussion period ending soon

Dear Reviewers,

Thank you for your reviews of this submission. The authors have posted responses addressing the concerns raised.

With limited time remaining in the discussion period, please take a moment to review the authors' responses and post a brief follow-up indicating whether your concerns have been addressed or what remains outstanding. If your assessment changes as a result, please update your score accordingly.

Active engagement at this stage is essential for a fair and well-substantiated final decision. Thank you for your time and effort in reviewing this submission.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.