- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
Helm-Diffusion is a conditional diffusion framework for wireless signal generation. It constructs static cross-frequency cues using leave-one-frequency-out Loewner prediction and dynamic cues using normalized local temporal spectra, motivated by a Helmholtz-based propagation model and local multipath approximations. Frequency-domain Gaussian smoothing and Gaussian noise define the forward process; a denoising network and learned physics-guided refinement predict the clean response during reverse sampling. Training combines reconstruction with matching static and dynamic consistency scores to measured responses. The paper reports evaluations on seven datasets spanning wireless sensing and uplink-conditioned downlink channel generation.
Critical strengths
-
A coherent physics-informed generative design for an important shared problem. Wireless sensing and communications both need synthetic signals that reproduce time-frequency statistics while retaining useful propagation structure. The paper addresses this recurring problem without requiring explicit scene geometry. Its complex-valued Transformer backbone, task conditioning, and attention-based refinement are well matched to phase-sensitive, structured RF responses. The contribution is the combination of observation-derived static/dynamic representations with a modified diffusion process (Section 3, pp. 4–6). Appendices A.3 and A.5 account for the structured Gaussian covariance introduced by frequency mixing when defining initialization and reverse sampling. This is a substantive domain-adapted learning design of interest to ICLR readers working on structured generation and scientific machine learning.
-
Broad within-domain evidence with consistent improvements and several complementary ablations. For separately trained dataset-specific models, Table 1 (p. 8) reports lower WD and MMD on all five sensing datasets and higher SNR on both FDD datasets than the listed baselines. Relative to RF-Diffusion, the reductions in the five-dataset mean WD and MMD are approximately 34.6% and 28.0%; the SNR gains are 1.84 dB on Argos and 1.73 dB on COST2100. Appendix A.6 documents five independent training and sampling runs, with means and standard deviations in Tables 4–6 (pp. 27–28). Figure 5 compares degradation processes and backbones, and Figure 7 tests the static and dynamic branches separately. Together, these results support the usefulness of the complete method across multiple wireless tasks and give the contribution more weight than a single-dataset demonstration.
Critical weaknesses
-
The physical interpretation is less firmly established than the empirical improvement. The architecture and component ablations support an effective combination of propagation-inspired features and generative modeling. They leave some uncertainty about how much improvement specifically comes from the proposed physical information. Figure 5(a) holds the architecture fixed, but No-Physics removes the entire learned refinement module (Section 4.3, p. 9), and Figure 7 does not fully specify the corresponding changes to trainable components and consistency losses. Table 2 (p. 17) demonstrates useful held-out-frequency prediction, while Equations (143)–(145) (p. 27) match observation-level consistency scores. These findings support structured inductive biases, but do not fully distinguish their propagation-specific value from generic frequency/temporal modeling or additional refinement capacity. This limits the strength of the mechanism attribution. A matched generic refinement control would strengthen that attribution; the existing results already provide evidence for the usefulness of the combined method.
-
Dataset-specific task definitions need clarification to assess the scope of the evaluation. Appendix A.6 (p. 22, Equation (112)) defines the localization condition as the mean receiver position. However, the cited espargos-0007 dataset has stationary transmitters and receivers and a moving passive target. The same appendix describes the HAR condition as a six-dimensional activity-related category, whereas the cited GaitSense study concerns gait-based identity recognition. The paper does not make clear how the position and category labels are mapped for these datasets. These could be terminology issues or valid subset/task choices, rather than implementation errors, but the distinction matters for interpreting the breadth and conditional validity of the results. The localization “Accuracy” endpoint in Figure 4(e) (p. 8) also needs an operational definition. The coherent design, consistent reported gains, and existing component studies support my current weak acceptance; clarification of these task definitions is important, and a substantive mismatch with the implemented experiments could change that recommendation.
Questions for authors
-
Dataset-to-task mappings. For ESPARGOS, does the spatial condition represent the passive target position or the receiver position? Please identify the raw position field used and explain the receiver-position description accompanying Equation (112). For GaitSense, which labels and encoding define the six-dimensional categorical condition, and which participant or activity subset is used? Please provide a concise per-dataset label mapping and define the localization “Accuracy” endpoint in Figure 4(e). These answers would distinguish reporting issues from task-definition differences that could affect the recommendation.
-
Attribution of the existing ablation results. For No-Physics and the static/dynamic branch ablations, which trainable components and loss terms are removed, and how are retraining and computational budgets matched? If a capacity-matched generic refinement control is available, how does it compare? These answers could strengthen the physical interpretation or indicate that the contribution should be framed more narrowly as an effective structured generative design.
2 comments
Thank you for the detailed response. The clarifications are helpful, but they do not substantially change my overall assessment. I will therefore maintain my original rating.
Weaknesses & Questions
W1 & Q2. Attribution of Physical Guidance and Ablation Studies
We thank the reviewer for the positive assessment of our overall design and experimental results, as well as for pointing out the need to further isolate the contribution of physical information. We agree that removing the Physics-Guided Module alone is insufficient to fully exclude the influence of additional neural refinement capacity. Therefore, we have added a parameter- and computational-complexity-matched generic refinement control in Appendix A.8, and clarified the component and loss modifications in each ablation.
Specifically:
All variants use identical dataset splits, training iterations, optimizers, random seeds, and diffusion steps, with matched trainable parameter counts. The results are presented in Table 1 (Sec. 4.4; Appendix A.8).
Table 1. Ablation results of the proposed physics-consistency designs and their generic alternatives.
The results show that Generic Refinement, despite having matched parameter count and computational complexity, performs substantially worse than the complete model (WD: 0.338 vs. 0.240; MMD: 0.158 vs. 0.055). Generic wireless feature replacements also fail to match the performance of the complete model. These findings indicate that the improvements are not merely attributable to additional neural refinement capacity, but are closely associated with the proposed static and dynamic propagation features and their consistency constraints.
W2 & Q1. Dataset-to-Task Mapping and Conditional Labels
We thank the reviewer for identifying the inaccuracies and omissions in our original descriptions of dataset conditions and task definitions. We have re-examined the datasets and experimental settings and corrected the corresponding descriptions in Appendix A.7.
ESPARGOS-0007: The original description incorrectly referred to the receiver position. We actually use the original
Posfield associated with each CSI sample: c(i)=Pos(i)=(xi,yi,zi). This field represents the passive target position, rather than the transmitter or receiver position. Both transmitters and receivers remain stationary in the experimental setup.GaitSense: We use the original gait recordings from all participants, without selecting a participant subset. The six-dimensional condition consists of the original acquisition metadata: User ID, Room ID, Position ID, Orientation ID, Instance ID, and Receiver ID. These fields describe the acquisition conditions rather than six activity classes.
Other datasets: WiDAR and FallDar use their original activity and event labels, respectively; DICHASUS uses its original spatial position annotations; and Argos and COST2100 use uplink CSI as the condition and downlink CSI as the generation target.
We have also clarified the Accuracy metrics in Figure 4(e): for FallDar, Accuracy represents fall-detection classification accuracy; for ESPARGOS, it represents the localization success rate, defined as the proportion of test samples whose Euclidean distance between predicted and ground-truth positions is no greater than 1.0 m.
We sincerely thank the reviewer for these detailed suggestions, which have helped improve the accuracy and reproducibility of our dataset descriptions and experimental evaluations.