acceptodds
Under review as a conference paper at ICLR 2027

Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series

Abstract

Generative models have changed how machine learning represents complex data distributions, especially in language and vision, yet many real-world systems are observed instead as continuous, high-dimensional, and noisy *sensor time series*. Existing generative modeling of sensor data, however, remains fragmented across modalities, datasets, and task formulations, limiting a systematic understanding of *when, how, and why* generative models succeed or fail in real-world settings. To address this gap, we introduce SensorGen, a large-scale study of sensor-signal generation spanning 14 settings across 4 domains, 7 datasets, and 12 signal modalities. Leveraging SensorGen, we systematically evaluate generative models from five major families and uncover three key findings: (1) flow-matching models provide strong overall performance across most settings; (2) signal properties matter, with demographic covariates improving longitudinal generation and time-frequency modeling improving high-frequency signal generation; and (3) generated signals have practical utility beyond visual realism, with scaling improving generation quality and synthetic data improving downstream performance. Together, SensorGen establishes a broader understanding of design choices, evaluation protocols, and failure modes in real-world sensor data generation.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Generative model, Sensor signal, Foundation model

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_HWGBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper provides benchmarking across diverse model families (Diffusion, Flow Matching, Autoregressive, GANs), offering useful empirical insights regarding model convergence, sampling speed, and structural signal fidelity under non-stationary noise.

Critical strengths

The paper offers valuable practical design principles for sensor time-series generation, such as demonstrating that flow-matching models (SiT) serve as strong baselines, fixed-range min-max normalization outperforms z-score normalization for generative targets, and spectrogram time-frequency conditioning enhances high-frequency signal generation (e.g., EEG).

Critical weaknesses

As shown in Figure 3, the dataset is heavily skewed towards one specific task category. This disparity creates potential bias when reporting aggregated metrics across categories.

Paired metrics are computed "in the normalized signal space used during training". Finding 2 compares z-score and min-max by MSE/MAE in each scheme's own space. Those scales differ, so the numbers are not comparable. Per-window min-max also discards absolute amplitude that the metrics then cannot see, and it is itself fragile to artifact spikes.

Questions for authors

MAR and autoregressive baselines often depend heavily on the continuous-to-discrete tokenization or latent patchification strategy. How discrete/patch tokens were constructed for MAR on continuous multi-channel signals, and whether continuous AR formulations were considered?

4 comments
AuthorsOpenReview ↗

Response to Reviewer HWGB (1/2)

We thank the reviewer for the thoughtful feedback and for recognizing the practical value of our findings and design principles for sensor signals generation. We hope the clarifications and additional results below further address your concerns and reinforce your positive assessment of the paper.

 

Q1: "As shown in Figure 3, the dataset is heavily skewed towards one specific task category …… creates potential bias when reporting aggregated metrics across categories."

We thank the reviewer for raising this concern. We agree that the distinction between dataset composition and evaluation weighting could be made clearer, and clarify our aggregation procedure below.

First, we would like to clarify that the data proportions shown in Figure 3 do not represent the weights used to aggregate task performance. Figure 3 provides a comprehensive breakdown of the collected data based on the number of windows available for each generation setting. These window counts do not determine how much each setting contributes to the reported performance averages.

In fact, as stated in Section 4.1 (lines 301–302), we aggregate results within each task category (e.g., Interpolation & Extrapolation) and report category-level averages. The results for different categories are reported separately in Tables 2–6. Therefore, the larger volume of sleep-imputation data does not influence the reported Semantic-to-Signal, Translation, or Editing results.

Second, we further prevent larger datasets from dominating the category-level averages by assigning equal weight to each task setting within the category. Each setting is evaluated on its corresponding dataset, and its contribution to the category-level average does not depend on the number of data windows in that dataset. This allows the results to better reflect model performance across different dataset properties and signal configurations.

Finally, we report the individual results for all 14 settings in Appendix C.1. For the reviewer’s convenience, we map the category-level averages to the corresponding setting-level results below:

Task CategoryCategory-level ResultsTask SettingsSetting-level Results
Semantic-to-SignalTable 43Table 15
Interpolation & ExtrapolationTable 33Tables 11–12
TranslationTable 24Tables 10 and 14
EditingTables 5–64Tables 13 and 16

Thus, the dataset-volume imbalance shown in Figure 3 does not give larger datasets disproportionate weight in the reported averages.

We have revised the evaluation description to state the equal-weight aggregation protocol explicitly in the revised paper(Sec. 4.1, lines 301–303).

 

Q2: "Paired metrics are computed "in the normalized signal space used during training" …… and it is itself fragile to artifact spikes (one outlier compresses the rest of the window)."

We thank the reviewer for the insightful observations. We agree that comparing normalization strategies requires a common evaluation scale, and would like to clarify a possible misunderstanding with our evaluation procedures.

The normalization comparisons reported in our manuscript already use a common evaluation scale. Predicted time series from both normalization methods are mapped to the same [-1,1] scale defined by the corresponding ground-truth signal before computing MSE, MAE, and SMSE. The sentence cited by the reviewer was intended to describe the default evaluation setting for our main comparisons, which consistently use min–max normalization. While in Appendix F.2, Evaluation settings for analysis studies seeks to discuss analysis study-specific evaluation settings, it did not explicitly describe the shared-scale procedure for the normalization experiment. We apologize for this ambiguity and have fully clarified the procedure in the revised manuscript (Section F.2).

Furthermore, we extend the normalization comparison to glucose signal translation, using the same shared-scale evaluation protocol, to further substantiate Finding 2.

ModelNormalizationMSE ↓MAE ↓SMSE ↓
DiTMin-max0.09740.22300.0320
DiTZ-score0.1110.2430.042
SiTMin-max0.09100.21560.0302
SiTZ-score0.1180.2460.039

Min–max normalization achieves lower errors across all three metrics for both models, reducing MSE by 12.3% for DiT and 22.9% for SiT relative to z-score normalization. These results demonstrate that the advantage observed in our original analysis extends to an additional task, providing further support for Finding 2. In summary, the additional experiment strengthens Finding 2 by demonstrating the benefit of min–max normalization on another task under the common evaluation scale.

We have made the shared-scale evaluation procedure explicit (Appendix F.2, line 1959), and incorporated the additional normalization comparison on glucose signal translation (Sec. 5, lines 467–472; Appendix D.2, line 1324; Table 24).

AuthorsOpenReview ↗

Response to Reviewer HWGB (2/2)

Q3: "MAR and autoregressive baselines often depend heavily on the continuous-to-discrete tokenization or latent patchification strategy. Could the authors clarify how discrete/patch tokens were constructed for MAR on continuous multi-channel signals, and whether continuous AR formulations were considered?"

We appreciate this question and would like to clarify how MAR handles continuous sensor signals in our implementation. Our MAR baseline already operates in a continuous-valued autoregressive formulation. Following Li et al. (2024), Autoregressive Image Generation without Vector Quantization, we model the conditional distribution of each token using a lightweight diffusion head. As described in Appendix G.3 (lines 2062–2070), signal values are not converted into discrete codebook indices or amplitude categories. Continuous autoregressive modeling is therefore already represented in our comparison. To adapt this formulation to sensor data, we operate directly on normalized raw signals, treating sensor channels analogously to RGB channels in images. We use a one-dimensional convolution to construct continuous embeddings from non-overlapping temporal patches, with its kernel size and stride both equal to the patch length. MAR performs masked autoregressive modeling over these patches, while its diffusion head directly generates the continuous waveform values across all channels within each patch. Our implementation therefore does not require a VAE latent representation or a VAE decoder after generation. This is analogous to the pixel-space MAR formulation demonstrated in Appendix D.1 of Li et al. (2024) [1], where the model directly generates RGB pixel patches without an image tokenizer. We have expanded the MAR description to make this continuous, tokenizer-free formulation explicit (Appendix G.3, lines 2070–2073).

 


References:

  1. Autoregressive Image Generation without Vector Quantization. NeurIPS, 2024.

 


 

We sincerely thank the reviewer once again for their time and constructive feedback. To support reproducibility and future research on sensor-signal generation, we are committed to releasing the following resources:

  • Benchmark. The SensorGen benchmark construction pipeline, including preprocessing, task configurations, and evaluation protocols, subject to the data use agreements of the source datasets.
  • Pretrained model. Checkpoints of the generative models trained for SensorGen’s task settings, where permitted by the applicable data use agreements, together with example code for inference.
  • Code. We will open-source the complete code to facilitate reuse by the broader research community.

We hope our detailed responses and the additional experiments have addressed all of your concerns, and we would be happy to answer any remaining questions during the discussion period. We would be very grateful if you might consider raising your score in light of these updates.

Thanks a lot for your comment, the authors answered all my concerns. I ll keep my score.

AuthorsOpenReview ↗

Thank you for your encouraging remarks!

Thank you for your encouraging feedback and confirmation that our rebuttal has fully addressed your concerns! Your constructive comments, especially regarding the presentation of methodology and study design, was highly valuable and will be incorporated into the revised manuscript. Please feel free to let us know if there are other clarifications or experiments we can offer.

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

This paper establishes a systematic benchmarking of the generative models, forming taxonomy for the models, tasks, and paradigms to understand the performance. The motivation is clear, given that sensing applications are fragmented and requires a unification way to analyze the performance across datasets and model families. The work has non-trivial contributions to the community for health-related sensing.

Critical strengths
  • The paper provides a very systematic evaluation benchmarking the generation models for health-related sensing applications.
  • The evaluation is already comprehensive.
Critical weaknesses
  • The benchmark largely focuses on health-related applications, yet sensing in general covers more diverse applications such as environment sensing with vibration sensors or RF-based sensing. These applications could have more diverse real-world noise characteristics compared to the applications evaluated in the paper. Those noise could contain more irregularity, uncertainty, and degradation of SNR. Therefore, there is still a gap between the work and the comprehensive sensing benchmarks.
  • Sensing applications contain spatial and modal information that are strongly correlated. Different located sensor signals or modalities can complement each other and can improve the downstream task or generation performance. The paper is less clear on the existing modeling techniques that explicitly consider such modality interaction or spatial interaction.
Questions for authors
  1. STFT Generation has been studied for sensing signals [1]. The paper discusses using STFT as the input for raw EEG generation. STFT can be more similar to image format, and have you considered using STFT as another alternative qualitative analysis task?
  2. In sensing applications, real-world noise often tend to severely degrade signal to noise ratio and more diverse environmental signals tend to have more irregular patterns due to the dynamic context. The evaluation covers additive gaussian noises, but how might structured or non-stationary noise affect the performance?
  3. Can authors provide some insights on subject-specific variability of the signals generated? Can the generation still capture subject-specific information, which is critical for potential privacy leakage or personalized subject-related sensing tasks.

1: Fine-grained Control of Generative Data Augmentation in IoT Sensing, Advances in Neural Information Processing Systems 37 (2024)

7 comments
AuthorsOpenReview ↗

Response to Reviewer Fnjs (1/3)

We thank the reviewer for their constructive feedback and positive evaluation of our work. We sincerely appreciate your recognition of the breadth of our study and the thoroughness of our evaluation. We hope the detailed responses below address your concerns.

 

Q1: "The benchmark largely focuses on health-related applications, ……. Therefore, there is still a gap between the work and the comprehensive sensing benchmarks."

Thank you for the valuable suggestion. We would like to provide additional results on the two mentioned settings.

Following the reviewer's suggestion, we extend SensorGen to vibration-based and RF-based sensing under our unified experimental protocol, introducing two additional generation settings:

  1. Vibration sensing: Following [1], we use the MOD dataset to construct a semantic-to-signal generation task conditioned on vehicle type.
  2. RF sensing: Following [2], we use the RadioML2016.10a dataset to construct a similar task, generating I/Q signals conditioned on modulation type.

As shown above, generative models successfully learn environmental sensing signals, and our key findings remain consistent across both new sensing domains.

Dataset (modality)ModelFID ↓Precision ↑Recall ↑
vibrationSiT0.5120.9670.804
vibrationDiT1.0770.9750.679
RFSiT0.7690.9200.814
RFDiT0.6330.9130.797
AverageSiT0.6400.9440.809
AverageDiT0.8550.9440.738

In summary, these additional experiments extend SensorGen beyond healthcare sensing, further demonstrating the generalizability of our benchmark and empirical findings.

We have expanded the benchmark with the vibration and RF sensing settings (Appendix D.12, line 1536; Table 36), noted this extension in the main text (Sec. 3.1, lines 156–157; Sec. 6, lines 503–504), and acknowledged in the new Limitations paragraph that broader sensor domains remain future work (Sec. 6, lines 490–495).

 

Q2: "Sensing applications contain …… less clear on the existing modeling techniques that explicitly consider such modality interaction or spatial interaction."

Thank you for the insightful feedback. We provide additional ablation results below.

Following the reviewer's suggestion, we further investigate explicit multisensor fusion by adapting GlobalFusion [3] with a flow-matching objective. We evaluate this approach on sleep-signal imputation under the same experimental protocol.

ModelMSE↓MAE↓SMSE↓
DiT0.2860.3890.079
SiT0.2700.3810.077
FractalGen0.2960.3850.092
TarFlow0.2900.4240.100
MAR0.2860.3980.091
GlobalFusion + Flow-matching0.2840.4020.084

As shown above, GlobalFusion with flow matching achieves top-performing results among the compared methods, providing additional empirical insights into the role of explicit multisensor fusion in sensor generation.

In summary, this additional comparison makes SensorGen a more comprehensive study by extending the evaluation to multisensor fusion. It complements our investigation of generative model families by examining interactions among sensor channels.

We have presented the explicit multisensor fusion analysis in a new subsection (Appendix D.9, line 1486), with results added to Table 12.

AuthorsOpenReview ↗

Response to Reviewer Fnjs (2/3)

Q3: "STFT Generation has been studied for sensing signals [4]. …… have you considered using STFT as another alternative qualitative analysis task?"

Thank you for the constructive feedback. Following [4], we visualize the STFT spectrograms of generated signals in sleep-signal imputation to examine their spectral characteristics. As shown in Fig. 17, the generated segments exhibit spectral patterns consistent with the paired ground-truth signals. This qualitative analysis complements our quantitative result that spectrogram conditioning improves sleep-signal imputation (Table 33).

We have added the STFT-based analysis to the revised paper (Appendix D.8, line 1479; Fig. 17, line 1856; Table 33) and discussed [4] in the related work (Sec. 2, lines 130–131).

 

Q4: "In sensing applications, …… The evaluation covers additive gaussian noises, but how might structured or non-stationary noise affect the performance?"

Thank you for the valuable suggestion. We further extend our ECG denoising evaluation to investigate model robustness under structured and non-stationary noise.

Specifically, beyond the Gaussian noise already evaluated in our paper, we consider baseline wander, time-varying variance, and motion artifacts. We evaluate DiT and SiT under the same experimental protocol and report the results below.

Noise TypeModelMSE↓MAE↓SMSE↓
Baseline wanderDiT0.03380.09990.0123
SiT0.00560.05130.0022
Time-varying varianceDiT0.03380.09940.0123
SiT0.00580.05040.0016
Motion artifactsDiT0.03430.10280.0124
SiT0.00690.05370.0022

As shown above, all models maintain consistently strong performance under non-stationary noise conditions, with flow matching retaining its leading position. Moreover, we have provided qualitative results in the revised manuscript, which further demonstrate the quality of denoised signals under such conditions.

By looking at disturbances beyond Gaussian noise, we believe that this analysis offers a more comprehensive view of generative models’ robustness to sensor noises and supports the stability under the tested conditions.

We have incorporated the ECG denoising analysis with these noise types (Appendix D.4, line 1376; Table 28) and provided qualitative examples of the denoised signals (Fig. 18, line 1881).

AuthorsOpenReview ↗

Response to Reviewer Fnjs (3/3)

Q5: "Can authors provide some insights on subject-specific variability ….. which is critical for potential privacy leakage or personalized subject-related sensing tasks."

Thank you for raising this concern. We would like to provide detailed analysis experiments and discussion below.

First, following the reviewer’s suggestion, we quantify subject-specific variability in generated signals on two tasks: sleep signal imputation and invasive blood pressure translation. In both tasks, generated samples inherit the subject identity of their conditioning signals. Using MIRA-extracted features, we estimate within-subject and between-subject variability from pairwise feature distances for both real and generated signals. We compute within-subject and between-subject SDs separately for generated and real signals, and report their generated-to-real ratios below.

TaskMethodWithin-subject variability ratioBetween-subject variability ratio
Sleep signal imputationDiT0.940.85
SiT0.940.79
Blood pressure translationDiT1.000.97
SiT1.040.91

As shown above, within-subject SD ratios are close to 1, which indicates that generated signals preserve variability comparable to real recordings. In contrast, between-subject SD ratios are substantially below 1, suggesting that generative models capture only part of the subject-specific variability.

Second, we further assess demographic information retention by using k-NN regression (R-squared) to predict age and logistic regression (AUROC) to predict sex from real and generated ECGs.

SourceAge R2Sex AUROC
Real ECG+0.130.55
Generated (DiT)−0.090.46
Generated (SiT)+0.030.47

As confirmed above, demographic attributes are less predictable from generated ECGs than from real ECGs, indicating limited retention of the evaluated demographic information.

Taken together, generative models studied in SensorGen largely preserve signal-level variability while retaining less subject-specific information, including demographic attributes.

We have added both analyses in a new subsection (Appendix D.11, line 1503), with the variability ratios and demographic prediction results in Table 35.

 


References:

  1. FreqMAE: Frequency-Aware Masked Autoencoder for Multi-Modal IoT Sensing. WWW, 2024.
  2. Self-Contrastive Learning Based Semi-Supervised Radio Modulation Classification. MILCOM, 2021.
  3. GlobalFusion: A Global Attentional Deep Learning Framework for Multisensor Information Fusion. IMWUT, 2020.
  4. Fine-Grained Control of Generative Data Augmentation in IoT Sensing. NeurIPS, 2024.

 


We sincerely thank the reviewer once again for their time and constructive feedback. To support reproducibility and future research on sensor-signal generation, we are committed to releasing the following resources:

  • Benchmark. The SensorGen benchmark construction pipeline, including preprocessing, task configurations, and evaluation protocols, subject to the data use agreements of the source datasets.
  • Pretrained model. Checkpoints of the generative models trained for SensorGen’s task settings, where permitted by the applicable data use agreements, together with example code for inference.
  • Code. We will open-source the complete code to facilitate reuse by the broader research community.

We hope our detailed responses and the additional experiments have addressed all of your concerns, and we would be happy to answer any remaining questions during the discussion period. We would be very grateful if you might consider raising your score in light of these updates.

Reviewer_9KMDOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2(−1)
Read review
Paper summary

SensorGen studies real-world sensor time-series generation at 14 settings across 4 domains, 7 datasets, and 12 signal modalities to evaluate six models from five generative model families. Time-series generation tasks include semantic-to-signal generation, sensor interpolation and extrapolation, channel translation, and signal editing. The generative model families considered include diffusion models, flow-matching models, autoregressive models, normalizing-flow models, and hierarchical models. The study adapts these generative models, originally developed primarily for image generation, to 1D time-series settings. The study aims to provide a systematic understanding of when, how, and why generative models succeed or fail in real-world settings instead of designing application-specific models. One of the main findings of the paper is that flow-matching models consistently provide strong performance across most settings. The paper also points out the importance of preprocessing demographic covariates. Another finding is that target-signal context reduces degradation in long-sequence generation. The paper also suggests that high-frequency signal generation benefits from time-frequency modeling. Further analysis is carried out to investigate whether such generated signals are useful beyond visual realism. To this end, various experiments are conducted in which different amounts of synthetic ECG data are used in pretraining or supervised training of models, including ResNet classifiers on the PTB-XL dataset, to test zero-shot and fine-tuned performance. It is shown that adding synthetic data improves downstream performance on Myocardial Infarction (MI), ST/T Change (STTC), and Conduction Disturbance (CD) detection tasks in several settings, although some metrics degrade. It is also shown that sensor generative models benefit from increased training compute and model capacity in the text-to-ECG generation task.

Critical strengths
  1. The paper conducts a large-scale study spanning 14 settings across 4 domains, 7 real-world datasets, 12 signal modalities, and 5 generative model families. In this work, these five generative model families which are primarily developed for image generation is adapted to sensor time-series generation.
  2. The paper reports failure cases, such as low recall in EEG generation and worsened downstream performance with high proportions of synthetic data.
Critical weaknesses
  1. The authors compare time-series-adapted generative models under various settings and datasets. However, there is a lack of strong baselines. Comparisons against specialist time-series generators are unavailable in the paper. The extrapolation task is a standard forecasting task, but the study lacks comparisons against such task specific models. The work offers, limited methodological novelty. The work adapts existing generative model families.
  2. The paper intends to provide a systematic understanding of "when, how, and why generative models succeed or fail in real-world settings" instead of designing application-specific models. However, it does not support this claim with enough experiments. Although the study has broad coverage, the range of datasets or tasks considered in several individual analyses is limited. Also, much of the study is designed to show which models perform well in particular settings rather than why they perform well.
Questions for authors
  1. How do these models perform against task specific time-series models for forecasting, imputation, etc., on the same tasks and datasets?
  2. The claim that flow matching performs consistently well across various tasks is not fully supported by the results (Tables 4 and 5). The paper compares only one implementation of flow matching with other generative models. How can the authors extrapolate this as a general conclusion?
  3. Do the differences in the reported results remain stable across different training runs? This is important because the results in many tables are close to each other.
  4. Are the models compared at fixed computational budget like parameters/inference time/ computation cost? Does this impact the ranking?
  5. Several takeaways and conclusions are based on a small subset of task or data. Can the analyses in Sections 4 and 5 of the main paper be extended to the wider range of settings?
  6. In Appendix B.2.1 and Table 15, semantic-to-signal performance is evaluated using distribution-level metrics. Would it be possible to investigate whether, in a task such as text-to-ECG generation, the generated waveform follows the clinical conditions described in the input text?
  7. In the comparison of min-max and z-score normalization, are the evaluation metrics computed on the same scale (either min-max or z-score normalized) for a fair comparison?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 9KMD (1/5)

Thank you for your detailed feedback and your recognition of the scale of our study. We respectfully believe this assessment overlooks several central contributions of the paper. Below, we clarify the scope, originality, and evidence supporting our findings to address each of your concerns and questions. We hope these clarifications better convey the paper’s contribution and merit a favorable reassessment.

 

W1 & Q1: "The authors compare time-series-adapted generative models under various settings and datasets. However, there is a lack of strong baselines. …… The work adapts existing generative model families"

Thank you for raising this concern. We respectfully disagree with the reviewer and would like to offer clarifications and results as follows.

First, instead of benchmarking task-specific models on a “standard forecasting task”, SensorGen aims to systematically investigate how representative generative modeling families perform across sensor-generation tasks. Our study covers 14 settings spanning semantic-to-signal generation, translation, interpolation/extrapolation, and editing. In particular, we have included tasks such as text-to-ECG generation that conventional forecasting models cannot directly address. This study design aligns with established generation benchmarks of general time-series, including TSGBench [1] and ConTSG-Bench [2], which focus on broad generative capabilities of model families rather than forecasting specialists.

Second, we respectfully disagree that adapting existing generative models implies limited originality. SensorGen establishes new empirical insights into how generative model families behave across sensor-generation tasks, including the effects of signal properties, conditioning strategies, and modeling choices.

This contribution aligns with the ICLR 2027 Reviewer Guidelines, which explicitly recognizes new empirical knowledge and practical insights as substantive scientific contributions, rather than requiring a novel model architecture.

”Submissions bring value to the ICLR community when they convincingly demonstrate new, relevant, impactful knowledge (incl., empirical, theoretical, for practitioners, etc)."

Finally, to further substantiate our findings, following the reviewer’s request, we evaluate Moirai [3] and TS-ICL [4], two state-of-the-art time-series forecasting models, on our three forecasting and imputation settings. As shown below, our leading generative models consistently outperform both forecasting foundation models. These additional comparisons broaden the evaluation and complement our central investigation of how different generative model families perform across sensor-generation tasks.

MethodSleep imputationMedication-aware forecastingCardiac forecasting
MSE ↓MAE ↓SMSE ↓MSE ↓MAE ↓SMSE ↓MSE ↓MAE ↓SMSE ↓
DiT0.2860.3890.0790.4940.5820.1630.2240.3720.046
SiT0.2700.3810.0770.2820.4260.0760.2030.3450.039
Moirai0.3250.4420.1220.5200.5810.0950.2050.2920.045
TS-ICL0.3820.5020.1140.4330.5030.0800.1570.2610.044

In summary, we believe the clarifications and additional results above address the reviewer’s concerns regarding baseline strength and originality. SensorGen’s contribution lies in the new empirical knowledge established through its systematic study, rather than solely through task-specific forecasting comparisons or introducing new model architectures. We have added the comparison with Moirai and TS-ICL to the revised paper (Appendix C.2, lines 1129–1133 and lines 1156–1157, Table 17).

AuthorsOpenReview ↗

Response to Reviewer 9KMD (2/5)

W2 & Q5: "The paper intends to provide a systematic understanding …… the range of datasets or tasks considered in several individual analyses is limited. Also, much of the study is designed to show which models perform well in particular settings rather than why they perform well."

Thank you for the feedback. We provide further discussion and additional results below.

First, SensorGen systematically investigates how signal properties and modeling choices affect generation quality through controlled analyses across models and task settings. These analyses reveal the factors driving model performance, beyond simply comparing model rankings.

Moreover, these analyses extend beyond individual models and settings. In Sec. 4.2–4.4 and Appendix C, we systematically investigate demographic conditioning, temporal context, frequency structure, and failure modes. Our analyses span multiple models and task settings. For example, we have already included two model families across three generation tasks in the long-sequence study (Fig.5).

Following the reviewer’s suggestion, we have further expanded task coverage of the two following analyses to further strengthen the generality of our findings: (1) high-frequency modeling, from emotion-to-EEG generation to sleep-signal imputation, and (2) normalization strategies, from blood pressure reconstruction to CGM translation.

As shown below, both findings consistently hold across the expanded tasks and datasets: fixed-range normalization improves sensor generation, while spectrogram conditioning enhances high-frequency signal modeling.

ModelConditioningMSE ↓MAE ↓SMSE ↓
DiTRaw waveform0.2860.3890.079
DiT+ STFT0.2790.3860.079
SiTRaw waveform0.2700.3810.077
SiT+ STFT0.2570.3800.078

Cgm translation

ModelNormalizationMSE ↓MAE ↓SMSE ↓
DiTMin-max0.09740.22300.0320
DiTZ-score0.1110.2430.042
SiTMin-max0.09100.21560.0302
SiTZ-score0.1180.2460.039

Second, explicit “why” analyses are already provided through controlled ablation studies in Appendix D.4–D.5 and detailed failure-case analyses in Appendix C.3. Specifically, we demonstrate how noise severity affects model robustness and how cross-channel dependencies improve signal translation. Our failure-case analysis further reveals why models struggle with high-frequency signals.

We further analyze model behavior on a few representative tasks to provide mechanism-level insights into why different generative models achieve their observed performance. Following prior work [5–7], we quantify conditioning utilization using precision at rank R (R-precision) [5] and boundary continuity using horizon-wise MSE [6] across representative generation tasks.

MethodTranslationImputationForecasting
MSE ↓R-precision↑MSE ↓Horizon MSE ↓MSE ↓Horizon MSE ↓
SiT0.0910.890.2700.750.2820.84
DiT0.0970.860.2860.830.4940.76
FractalGen0.1490.850.2960.840.5321.03
TarFlow0.3070.110.2900.990.5471.02
MAR0.2980.100.2861.050.5701.01

The table reveals that better conditioning utilization and stronger context continuity are associated with superior generation performance. We will add the additional analysis into our final manuscripts.

Taken together, these clarifications and additional results address the reviewer’s concerns regarding the breadth and explanatory depth of our analyses.

We have extended the high-frequency modeling analysis to sleep-signal imputation (Sec. 4.4, lines 399–400; Appendix D.8, line 1479; Table 33; Fig. 17) and the normalization analysis to glucose signal translation (Sec. 5, lines 467–472; Appendix D.2, line 1324; Table 24), and added the analysis of conditioning utilization and boundary continuity (Appendix D.10, line 1494; Table 34).

AuthorsOpenReview ↗

Response to Reviewer 9KMD (3/5)

Q2: "The claim that flow matching performs consistently well …… compares only one implementation of flow matching with other generative models. How can the authors extrapolate this as a general conclusion?"

Thank you for the feedback. We would like to clarify our key contribution and provide additional results.

First, SensorGen does not aim to develop a best-performing model that dominates every generation task. Rather, our goal is to systematically study representative generative model families across heterogeneous sensor-generation settings and establish generalizable empirical insights. Our implementations therefore preserve each model’s defining generative objective and generation procedure while only making necessary adaptation to the sensor domain, and evaluate the resulting models using shared task definitions, preprocessing, data splits, and evaluation procedures.

Second, we would like to point out that our paper never claims that flow matching always achieves the best performance. As explicitly stated in Takeaway 1, our conclusion is that “flow-matching models provide a strong baseline for sensor time series generation.” This is distinct from claiming universal superiority, and our paper explicitly discusses settings where other generative models perform better in Sec. 4.1. Moreover, our empirical results directly support this conclusion. Across 14 generation settings, SiT achieves the best in 8/14 settings and ranks among the top two in 10/14, while no other model leads in more than two settings. SiT also achieves the strongest aggregate performance in Translation and Interpolation & Extrapolation (Tables 2–3), and leads in two of three Semantic-to-Signal metrics (Table 4).

Finally, SiT represents an established transformer-based implementation of flow matching. To further examine whether our finding extends beyond SiT, we additionally evaluate FM [8] across semantic-to-signal generation and Interpolation & Extrapolation.

ModelActivity-to-IMUEmotion-to-EEG
FID ↓Precision ↑Recall ↑FID ↓Precision ↑Recall ↑
DiT0.320.8900.6184.770.7880.098
SiT0.250.8930.6834.410.8610.232
FractalGen1.100.8840.1518.830.0490.008
TarFlow0.790.8570.2257.340.0240.000
MAR0.470.9030.4638.330.6370.011
Imagen0.540.8300.5505.930.3950.010
FM[8]0.450.8640.4742.590.7700.447
ModelCGM super-resolution
MSE ↓MAE ↓SMSE ↓
DiT0.0590.1510.032
SiT0.1340.2600.088
FractalGen0.1270.2560.042
TarFlow0.0340.1440.008
MAR0.2660.3930.121
Imagen0.0080.0640.002
FM[8]0.0140.0880.003
ModelText-to-ECG
FID ↓Precision ↑Recall ↑
DiT1.070.9820.373
SiT2.030.9920.325
FractalGen11.890.9630.059
TarFlow4.660.9710.107
MAR14.480.7020.001
Imagen1.070.9760.478
FM[8]1.920.9930.157

As confirmed, flow matching families still lead the performance across the compared tasks. Taken together, we have clarified that SensorGen provides a systematic study of generative model behavior across diverse sensor-generation tasks, with the additional comparisons strengthening the empirical support for our findings.

We have added FM to the comparisons on semantic-to-signal generation and glucose signal super-resolution (Tables 4, 5, 13 and 15), described it in Sec. 3.3 (lines 253–255) and Appendix G.2 (lines 2057–2059), and summarized its results in Sec. 4.1 (lines 308–309) and Appendix C.1 (lines 1040–1042). We have also stated the setting-level ranking of SiT in Sec. 4.1 (lines 302–304).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all the reviewers for acknowledging the contributions of our work and providing insightful comments and suggestions! We are glad to see that the reviewers found that:

  1. The problem we study is important and well-motivated (Reviewer Fnjs).
  2. SensorGen is a large-scale (Reviewer 9KMD), systematic benchmark (Reviewer Fnjs), and is a valuable contribution to the community (Reviewer Fnjs, HWGB).
  3. The evaluation is comprehensive (Reviewer Fnjs) and sound (Reviewer 9KMD, Fnjs, HWGB).
  4. The analyses are thorough (Reviewer 9KMD), and the findings are insightful and practical (Reviewer Fnjs, HWGB).
  5. The manuscript is clearly written and easy to follow (Reviewer 9KMD, Fnjs).

 

To address concerns raised by the reviewers, we have made the following changes and updated the manuscript accordingly (major changes are highlighted in red in the updated paper):

  1. We have added more models and baselines, including task-specific time-series models (Appendix C.2, Table 17), an additional flow-matching model (Tables 4, 5, 13 and 15), and an explicit multisensor fusion variant (Appendix D.9, Table 12), as suggested by Reviewer 9KMD, Fnjs.
  2. We have added results across training seeds and bootstrap confidence intervals (Appendix D.7, Tables 31–32), and pointed to model capacity and training throughput (Section 4, Table 37), to support the reliability of our comparisons, as suggested by Reviewer 9KMD.
  3. We have extended our analysis studies to additional tasks (Sections 4.4 and 5, Appendices D.2 and D.8, Tables 24 and 33), and added new analyses of conditioning utilization, boundary continuity, and clinical-condition adherence (Appendices C.5 and D.10, Tables 20 and 34) to better understand model behavior, as suggested by Reviewer 9KMD.
  4. We have extended SensorGen beyond healthcare to vibration and RF sensing in Appendix D.12 and Table 36, as suggested by Reviewer Fnjs.
  5. We have added further evaluations of generated signals, including robustness to structured and non-stationary noise (Appendix D.4, Table 28, Fig. 18), STFT spectrogram visualizations (Appendix D.8, Fig. 17), and subject-specific variability (Appendix D.11, Table 35), as suggested by Reviewer Fnjs.
  6. We have clarified the aggregation protocol (Section 4.1), the evaluation scale of the normalization analysis (Section 5, Appendix F.2, Table 24), and the continuous-token formulation of MAR (Appendix G.3), as suggested by Reviewer 9KMD, HWGB.
  7. Beyond the points above, we have added a Limitations paragraph discussing clinical use, dataset coverage, evaluation challenges, and future directions (Sec. 6, lines 488–496).

 

We hope our responses have addressed all concerns from the reviewers adequately. We thank all reviewers again for their time and feedback, and please feel free to let us know if there are other clarifications or experiments we can offer. We would really appreciate it if the reviewers could consider raising their scores after evaluating our updates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.