acceptodds
Under review as a conference paper at ICLR 2027

ALF: Spectrally Anchored Latent Flow Matching For Wireless Signal Generation

15 views

Abstract

Wireless time-series generation is useful for emerging applications that will drive the adoption and integration of machine learning tasks within next-generation networks, such as waveform classification in shared spectrum bands and interpretation of the physical world through integrated sensing and communications. Unlike a generic time series, wireless signals have time-frequency duality, where a frequency domain bandwidth constraint implicitly imposes latent structural constraints on the evolution of the time-domain sequence. In this paper, we develop Anchored Latent Flow (ALF), a latent space flow model for label-conditioned signal generation that exploits this duality. Our flow model ”reasons” through a spectrogram, which is a 2-D representation of the signal constructed by applying a windowed Fast Fourier Transform to successive, overlapping signal segments and mapping their frequency-domain power components across time. Specifically, we shape the latent space of the flow model through a training-time auxiliary objective that is supervised by the (latent) spectrogram (i.e. anchoring through the spectrogram). By guiding the flow field to reason through a spectrogram, the generated sequence maintains long-range coherence across time and greatly improves the signal quality at inference. Empirically, ALF exhibits state-of-the-art generation quality and reconstruction error across various signal lengths, waveform classes, signal-to-noise-ratio (SNR) and channel conditions, reaching a train on synthetic, test on real (TSTR) accuracy of 84.5%, a 1.9× average memory reduction, and 100× average compute speedup over existing models. Our design further enables downstream tasks including style transfer and data augmentation, and has broad utility for the wireless community.

Primary area:
generative models
Keywords:
flow matching, latent generative models, spectral anchoring, time-series generation, radio-frequency signals

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

This paper proposes Anchored Latent Flow (ALF), a conditional wireless signal generation framework that incorporates spectrogram-based auxiliary supervision into latent flow matching to improve generation quality and efficiency.

Critical strengths

The paper addresses an important problem: generating high-quality synthetic wireless signals that can potentially improve downstream model training, particularly in scenarios where collecting real-world data is costly or limited. This is a practically relevant research direction with potential value for wireless machine learning applications.

Critical weaknesses

1. Limited evaluation of generation quality and downstream utility.

The evaluation relies mostly on TSTR/TRTS experiments between ALF-generated samples and WaveScene, which is itself a synthetic dataset constructed by the authors. Since ALF is trained to reproduce the WaveScene distribution, both the generated and reference signals are tied to the same simulation framework. As a result, the reported classification performance may reflect simulator-specific patterns or shortcuts rather than physically meaningful signal properties that generalize to real-world wireless environments.

This limitation leaves two important claims insufficiently supported:

a) Conditional fidelity: The current TSTR/TRTS results do not convincingly establish whether ALF generates signals that faithfully satisfy the specified physical conditions generally. Evaluation using independent and reliable condition-specific measurements or pretrained evaluators (not on WaveScene, preferably real-word recordings) would provide stronger evidence that the generated signals capture the intended physical attributes rather than simulator-specific correlations.

b) Downstream utility: The authors highlight data augmentation as an important application and motivation for wireless signal generation in the abstract and introduction. However, the experiments do not establish whether ALF-generated data provides meaningful utility for real-world downstream tasks. The reported TSTR results only show that a classifier trained on ALF-generated signals can perform well on WaveScene samples, which are themselves simulator-generated.

More convincing evidence could be provided at two levels: (1) demonstrating that models trained solely on ALF-generated signals achieve competitive performance on real-world RF recordings, which would provide strong evidence of the synthetic data's practical utility; or (2) demonstrating that augmenting real-world RF training data with ALF-generated signals improves downstream performance compared to training on the same real-world data alone, particularly in limited-data settings. Without either evaluation, the practical motivation for generating these signals remains insufficiently substantiated.

2. Insufficient analysis of the spectral anchor.

Spectral anchoring is the central component of the proposed method, but the quality of the anchor representation itself is not sufficiently evaluated. The anchor is obtained from a separately trained spectrogram autoencoder, but the paper provides limited quantitative evidence of how faithfully its latent representations preserve the relevant spectral properties. While Figure 4 offers a qualitative visualization and Table 3 demonstrates improvements over unanchored flow matching, neither directly validates the quality of the learned anchors.

Furthermore, it remains unclear whether the improvements specifically arise from the proposed spectrogram latent anchoring or could be achieved with simpler spectral supervision, such as directly predicting spectrogram features or applying frequency-domain losses. Additional analysis and ablations would help substantiate the necessity of the proposed anchor design.

Questions for authors

The presentation of Section 6 (Results) could be substantially improved. Currently, experiments addressing different objectives are presented together without a clear structure, making it difficult to follow the evaluation logic and understand what each experiment is intended to demonstrate. For example, Table 2 is introduced before the corresponding evaluation protocol, baselines, and metrics are explained. I would suggest organizing the results into separate subsections according to their evaluation objectives (e.g., generation quality, ablation studies, computational efficiency, and downstream applications). Each subsection should first clearly state the research question or claim being evaluated, followed by the experimental setup, evaluation metrics, and results. A dedicated experimental setup subsection would also help readers understand the overall evaluation protocol before interpreting the results.

5 comments
AuthorsOpenReview ↗

Response to reviewer Ndrf - Part 1/3

Dear Reviewer Ndrf,

We thank the reviewer for the constructive feedback, particularly regarding the necessity of spectral anchoring and the real-world utility of ALF. We have conducted additional experiments to address these concerns and clarify the design choices underlying our framework.

(Q1) Limited evaluation of real-world generation quality and downstream utility. Since existing TSTR/TRTS experiments use WaveScene as both the source and target distribution, how can the authors establish that ALF-generated signals provide meaningful benefits on real RF recordings rather than merely reproducing simulator-specific characteristics?

(A1) Real-world validation and utility.

We strongly agree with the reviewer’s concern, and have added extensive real-world validation in Appendix H. We provide an overview of the topics covered in Appendix H below:

To evaluate whether ALF-generated signals provide practical benefits beyond the WaveScene simulation environment, we conducted Over-The-Air (OTA) experiments in two distinct settings: a controlled indoor laboratory environment and the outdoor NSF PAWR POWDER testbed.

Results from laboratory measurements (Indoor Hotspot/InH).

We collected 13,776 OTA recordings across 13 capture configurations in an indoor-hotspot setting, including LOS, NLOS, reflector/blockage, and moving conditions, with TX–RX separations of 2.16, 3.90, and 9.14 m. ALF-generated augmentation improves classification accuracy on held-out InH recordings by 8.66 percentage points over limited-real-only training, achieving performance comparable to training on the complete real dataset.

Results from outdoor urban measurements (NSF PAWR POWDER platform).

We further evaluated ALF using 23,621 OTA recordings classes across 3 directed links on the NSF PAWR POWDER testbed. Unlike the laboratory measurements, outdoor transmitter and receiver positions remained fixed while transmit gain was varied to capture different received-signal conditions. This set of data has a longer transmission distance (~200m) and exhibits lower measured SNR, presenting a significant distribution shift compared to InH data. ALF augmentation still improves classification accuracy by 8.57 percentage points over limited-real-only training, approaching the performance obtained using the complete real training set.

Details of the two augmentation experiments can be found in Appendix H.3 (Tables 20–21), where we further show comparisons of ALF and other traditional data augmentation methods. We also included label-only decoding as a baseline to isolate the learned flow's contribution. Directly decoding metadata-derived latents from the metadata encoder through the frozen Stage-1 decoder, without flow matching, achieves much worse results on both experiments. These results demonstrate that the learned latent flow contributes substantially beyond the pretrained decoder and label conditioning alone.

Synthetic-to-real encoder transfer.

Appendix H.4 (Tables 22–25) separately evaluates the frozen WaveScene-trained encoders using both synthetic-trained and OTA-trained classification probes. ALF achieves the strongest accuracy among the evaluated encoders in both settings. Together with the embedding visualizations in Figures 20-21, these results demonstrate that ALF's learned representations retain meaningful waveform information after transfer to real OTA measurements.

AuthorsOpenReview ↗

Response to reviewer Ndrf - Part 2/3

(Q2 (a)) Insufficient evaluation of spectral anchoring. The learned spectrogram anchor is a central contribution, but its quality and benefits are not sufficiently established.

(A2 (a)) Gains from spectral anchoring.

Spectral anchoring helps preserve fine-grained fidelity in the spectral domain, particularly in bandwidth accuracy and the preservation of structured time–frequency features. With newly expanded Appendix E.11, we now demonstrate these benefits using 2 additional experiments and highlight theoretical benefits that were previously not sufficiently emphasized in Appendix A. Both bandwidth accuracy and preservation of structured time–frequency features are important for functional wireless signal generation. For example, frequency-modulated chirps, a fundamental tool in radar technology, enable longer-duration, higher-energy transmissions while maintaining range resolution through pulse compression. Their practical utility depends critically on preserving their characteristic frequency evolution.

Bandwidth fidelity. We demonstrate how spectral anchoring improves adherence to requested occupied bandwidth. Narrowband signals are particularly sensitive to relative bandwidth errors: energy leakage outside the desired band can interfere with neighboring channels. To quantify bandwidth fidelity, we consider identical subsets of the held-out 100k test sets for ALF and LF-M (without spectral anchoring), with occupancy labels between 0 and 0.2 and SNR above 20 dB. These subsets contain 24,272 records. We measure bandwidth noncompliance by the absolute difference between requested occupancy and measured 99%-energy occupied bandwidth, normalized by sampling rate, where \mathrm{OccBW}_{99\\%}=\frac{\mathrm{BW}_{99\\%}}{f_s}. As shown in Table 14, under identical conditions, ALF reduces the mean occupied-bandwidth error by 14.23% relative to LF-M, demonstrating improved adherence to spectral occupancy conditions.

Spectrogram fidelity. Spectral anchoring also reduces spectrogram errors for nonstationary waveforms such as radar chirps and hopping sequences. Following the evaluation setting of Table 3, we compare ALF and LF-M on the same held-out 100k test set using identical conditioning labels. We restrict the analysis to SNR ≥ 0 dB to include low-SNR regimes while excluding signals below the noise floor. We measure spectrogram MSE (Spec-MSE) between each generated recording and its reference waveform, then analyze 10 nonstationary and 16 stationary waveform classes separately, averaging within each class before averaging across classes. Shown in Table 15, ALF reduces Spec-MSE by 7.8 dB² on nonstationary waveforms, compared with 1.2 dB² on stationary waveforms, representing a 5.4× larger relative improvement. This demonstrates that spectral anchoring is particularly beneficial for signals with evolving spectral structure.

Theoretical benefits. In the revised paper, we have emphasized the theoretical analysis included in Appendix A. From a theoretical standpoint, spectral anchoring eases the generative modeling task by encouraging structural constraints that are challenging to learn purely in the time domain. For a sequence containing n discrete points spanning d levels, an unconstrained joint distribution demands dn−1d^n-1 free parameters. RF waveforms are inherently band-limited: despite containing n temporal samples, their spectral representations have restricted support. Through the inverse Fourier transform, the retained spectral coefficients determine the sequence, illustrating that time-domain samples are tightly coupled rather than independent parameters. Incorporating a spectral anchor s leverages this structure through p(x,s)=p(s)p(x∣s)p(x,s)=p(s)p(x\mid s). As discussed in Appendix A, a support-conditioned formulation reduces the required parameter space to at most ∣K∣dk−1|\mathcal{K}|d^k-1, where K\mathcal{K} denotes valid frequency support, compared with dn−1d^n-1 in an unconstrained formulation. This motivates spectral anchoring as a physically informed inductive bias for flow learning.

AuthorsOpenReview ↗

Response to reviewer Ndrf - Part 3/3

(Q2 (b)) Could simpler spectral supervision, such as direct spectrogram features or frequency-domain losses, provide similar improvements?

(A2 (b)) Why a separate latent spectral encoder. Regarding the choice of spectral anchor, Stage 2 learns to transport representations toward the reconstruction-oriented IQ latent space established in Stage 1. The Stage 1 spectral branch is fused with raw-IQ features and jointly optimized for reconstruction and label alignment; its information is therefore already incorporated into the final latent representation that the flow aims to learn. Reusing these features as intermediate supervision would provide a target closely related to the flow's endpoint, rather than an independent spectral constraint. Instead, we seek an intermediate target that retains important time-frequency structure while abstracting away finer IQ details. The separately trained spectrogram autoencoder provides such a compact, spectrally focused representation, allowing the flow to learn global spectral structure before completing the transport toward the full IQ latent. As for direct STFT log-power supervision, it would require an additional mapping to produce compact, context-aware spectral embeddings compatible with the flow's intermediate representations. This is precisely the role of our separately trained spectrogram autoencoder, which learns such embeddings through a dedicated spectral reconstruction objective rather than relying on an arbitrary projection.

(Q3) Organization of Section 6. The reviewer suggests introducing the evaluation setup before the results and reorganizing the experiments into clearly separated objectives, such as generation quality, ablations, computational efficiency, and downstream applications.

(A3) Organization of Section 6.

We thank the reviewer for this helpful suggestion. We have restructured Section 6 of the revised manuscript accordingly, introducing the experimental setup and evaluation protocol before presenting the results, and organizing the experiments into clearer subsections based on their objectives: experiment setup, generation quality, ablation studies, computational efficiency, and downstream applications. This organization is intended to make the evaluation logic and the purpose of each experiment easier to follow.

Concluding Remark

We hope the clarifications and additional results above address your concerns. If you are satisfied with our response, we kindly ask you to consider revisiting your score. We welcome any further questions and are committed to addressing additional concerns during the discussion period.

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

First, I would like to disclose that I am not an expert of wireless signal generation. This issue has been discussed with AC. Due to the lack of reviewers in this field, I was encouraged to review this manuscript, so I will leave my comments with low confidence, and will read other reviews to learn more context during the rebuttal period.

Next, I would like to summarize my thoughts on the paper.

  • I like the outline of this paper:
    • it starts from a signal domain (wireless signal) that is less discussed in the generative model community.
    • It then built a latent flow-matching model inside the latent space created by a powerful 768-dim auto-encoder.
    • In the evaluation phase, the authors covered the data augmentation application under downstream classification tasks (table 2), then go to more signal-level metrics (table 3), and resource consumption (table 4)
    • In the last part, the author mentioned zero-shot style-transfer (e.g., denoise) capability.
    • The mentioning of the block-AR long-form generation further makes the paper more solid.
  • However, the evaluation protocal is confusing to me
    • In Appendix C.2 says "A deterministic canonical bit stream is generated once, and the modulation and protocol waveforms draw length-appropriate prefixes from this stream before class specific framing, coding, scrambling, and modulation."
      • Does this mean, the bit stream (i.e., the content) of this huge dataset is always the same? If so, I think the diversity of this dataset has a huge issue which may make the whole evaluation unreliable.
    • I don't see some important metrics that can evaluate the generation quality (e.g., Frechet Distance) or the content preservation (decoded bit error rate), which is also confusing: if the bit error rate is high, I imagine the method can only be used as a data augmentation tool for classification tasks, as the content accuracy cannot be guaranteed.

Other strengths and weaknesses will be mentioned in their own sessions.

Given the strengths (a very complete tech report regarding a less discussed signal domain) and weaknesses (questions regarding the dataset and evaluation metrics), I think the paper is at the boundary. I am tentatively giving a "weak reject", and is willing to see if the above concerns can be explained or handled.

Critical strengths

Details can be found in the Summary.

  • A very comprehensive tech report to explain how to train a generative model for a minor signal domain
Critical weaknesses

Details can be found in the summary

  • Skeptical bit stream (transmission content)

  • Evaluation metrics do not include bit error rate, so that we don't know if the content is well preserved.

Questions for authors

There is no reproducibility statement. Is there an open-source plan regarding your training data, evaluation protocal, and trainng/inference code?

Without open-source, I cannot imagine how to reproduce this work and benefit the community with the insights from this work.

3 comments
AuthorsOpenReview ↗

Response to reviewer NVaN - Part 1/2

Dear Reviewer NVaN,

We thank the reviewer for the careful reading and constructive feedback. We address the concerns regarding dataset diversity, evaluation metrics, and reproducibility below.

(Q1) Deterministic payload and dataset diversity. Appendix C.2 states that modulation and protocol waveforms draw prefixes from a canonical bitstream. Does this mean the dataset repeatedly uses the same content, potentially limiting diversity and making the evaluations unreliable?

(A1) Deterministic payload and dataset diversity.

WaveScene uses a canonical fixed bitstream for communication waveforms and truncates it to each time-series length. This choice follows an established practice in physical-layer learning tasks that focus on propagation channels and hardware characteristics. For example, Arvinte and Tamir (2023) use a fixed QPSK pilot matrix across test samples for score-based generative MIMO channel estimation, while He et al. (2025) similarly employ known pilot signals in diffusion-based channel estimation. In RF fingerprinting, Morin et al. (2019) use identical modulated bit sequences across transmitters, while Deng et al. (2017) investigate hardware-dependent characteristics using fixed protocol payloads.

Although these tasks differ from general-purpose waveform generation, they illustrate a common principle: controlling transmitted content allows physical-layer variability to be studied independently of message-content variation. We adopt this principle in WaveScene to focus on learning waveform characteristics and propagation conditions while avoiding the additional complexity of arbitrary payload generation.

To put this additional complexity into perspective, a single WiFi recording in WaveScene can contain up to 154,400 encoded payload bits, representing an enormous space of possible messages. Randomizing these payloads across recordings would introduce another high-dimensional source of variation, making it more difficult for the model to isolate and learn waveform configurations and propagation channels.

Nevertheless, WaveScene recordings exhibit substantial feature diversity despite the fixed source bitstream. Radar waveforms do not encode communication payloads and are therefore unaffected by this choice. Standard modulation and protocol waveforms vary in symbol rates, pulse shaping, framing, coding, and other transmission parameters, while different recording lengths produce different payload prefixes.

Altogether, WaveScene includes 1,066 parameterized waveform variants across 32 classes; these are further diversified by sampling rates, transmission parameters, channels, and noise conditions. Thus, sharing a canonical source bitstream does not produce identical IQ recordings or eliminate waveform diversity. We further validate the utility of ALF beyond simulated channel conditions through independent indoor and outdoor Over-The-Air (OTA) experiments in the new Appendix H, demonstrating that its learned waveform representations remain useful under real-world propagation conditions.

We integrate the controlled-payload discussion above into a dedicated subsection, Appendix C.3, to help clarify similar concerns for future readers.

(Q2) Missing distributional evaluation metrics. Why does the paper not report Fréchet Distance or comparable distributional metrics for generated signals?

(A2) FID evaluation.

FID measures distributional similarity in a chosen feature space, but does not necessarily reflect the functional correctness or downstream utility of generated wireless signals. These properties are especially important in communications, as discussed by Baur et al. (2025) and highlighted in Section 6 of our paper. This is why we prioritize downstream evaluations.

Nevertheless, to provide a more complete comparison, we added Appendix E.13 (Table 17), and evaluated FID using three feature extractors: Inception-v3 applied to spectrogram images, the spectrogram autoencoder used to provide spectral anchoring targets in ALF training, and our Stage-1 IQ encoder. Notably, TimeWeaver achieves better FID with Inception-v3 and the IQ encoder despite performing substantially worse than ALF on downstream waveform classification (61% versus 84% TSTR, Table 2).

Overall, FID rankings do not align with downstream utility rankings, further illustrating why distributional metrics alone are insufficient to characterize RF generation quality.

AuthorsOpenReview ↗

Response to reviewer NVaN - Part 2/2

(Q3) Missing bit-error-rate evaluation. If generated waveforms cannot preserve or recover communication payloads, does that limit their usefulness beyond classification-based augmentation?

(A3) BER evaluation.

Following the reviewer's suggestion, we evaluated payload decodability using blind, standard-specific receivers across 16 digital waveform classes, and results are documented in the new Appendix E.12 (Table 16). Each receiver knows the relevant PHY configuration, but has no access to the transmitted payload. The true payload is used only to score BER after receiver acceptance has been determined.

At SNR ≥ −20 dB, 53% of ALF recordings are decodable, compared with 66% of WaveScene reference recordings, 51% of LF-M, 6% of TimeWeaver, and 0% of WaveStitch. Among decodable ALF recordings, we achieve a BER of 1.3%. These results demonstrate that ALF-generated signals can preserve recoverable communication content.

(Q4) Reproducibility and open-source availability. Is there a plan to release the training data, evaluation protocols, and training/inference code so the community can reproduce this work?

(A4) Reproducibility and code availability.

The value of open source has been a guiding principle throughout this project. Appendix C.7 details the reproducibility measures used in WaveScene construction and our commitment to releasing the dataset-generation pipeline. We have also added a Reproducibility Statement to the revised manuscript on page 11, outlining our intention to publicly release ALF's implementation, training and inference pipelines, model weights, evaluation scripts, and exact experiment configurations upon acceptance. We intend to maintain these resources publicly to facilitate further research. If additional implementation details would be helpful during the review process, we are also willing to explore sharing relevant portions of the code in an anonymized form, consistent with the conference's double-blind review requirements.

References for (A1)

  1. M. Arvinte and J. I. Tamir, “MIMO Channel Estimation Using Score-Based Generative Models,” IEEE Transactions on Wireless Communications, 22(6), 3698–3713, 2023.
  2. Z. He, F. Héliot, and Y. Ma, “Deterministic Score-based Diffusion Model for Channel Estimation in RIS-assisted MIMO Systems,” IEEE SPAWC, 2025.
  3. C. Morin, L. Cardoso, J. Hoydis, J.-M. Gorce, and T. Vial, “Transmitter Classification with Supervised Deep Learning,” CrownCom, 2019.
  4. S. Deng, Z. Huang, X. Wang, and G. Huang, “Radio Frequency Fingerprint Extraction Based on Multidimension Permutation Entropy,” International Journal of Antennas and Propagation, 2017, Article 1538728.

Concluding Remark

We hope the clarifications and additional results above address your concerns. We also invite the reader to explore the newly added Appendix H for real world Over-The-Air (OTA) evaluations of ALF’s performances. If you are satisfied with our response, we kindly ask you to consider revisiting your score. We welcome any further questions and are committed to addressing additional concerns during the discussion period.

Dear Reviewer,

I hope you got the chance to take a look at our response. If you have any additional questions, we would be happy to address them. Thank you for your time and useful feedback.

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

This paper proposes Anchored Latent Flow (ALF), a two-stage latent flow-matching framework for conditional generation of multi-antenna (SIMO) wireless RF IQ waveforms. Stage 1 trains a patch-based IQ autoencoder (1024 samples per patch) aligned with 6 RF metadata labels via a SoftCLIP and structural objective. Stage 2 trains a rectified flow model initialized from a label-conditioned Gaussian prior, where an intermediate Transformer layer is supervised to predict latents from a separately trained spectrogram autoencoder ("spectral anchoring"). The authors also introduce WaveScene, a 7.82M synthetic SIMO IQ dataset, and show strong gains in TSTR/TRTS accuracy and inference speed over baselines like Time Weaver.

Critical strengths
  1. Well-engineered pipeline with strong empirical speed and quality gains: Moving long-sequence RF IQ generation from raw sample space into a patch-compressed, label-aligned latent space makes generating 32k-length multi-antenna waveforms practical (100x faster than Time Weaver in Table 4, with TSTR jumping from 0.61 to 0.84 in Table 2). Adding a training-time spectral anchor (similar in spirit to REPA) is also a clean way to inject time-frequency inductive bias without adding inference overhead.
  2. Large-scale benchmark contribution: The proposed WaveScene dataset (7.82M recordings across 32 waveform classes, 4 antennas, and multiple lengths/3GPP channel models) is comprehensive and will be a useful resource for the wireless ML community.
Critical weaknesses
  1. The actual gain from "Spectral Anchoring" is quite modest compared to the Stage-1 latent/prior design (Table 3a & Table 12): While spectral anchoring is framed as the core contribution, Table 3(a) shows that unanchored latent flow with the metadata prior (LF-M) already reaches 83.49% TSTR (and LF-N reaches 81.93%), while adding spectral anchoring (ALF) only brings it to 84.49% (+1.0% absolute). Similarly, in Table 12, the spectrogram MSE improvement over Baseline w/ Prior is only 1.7% (119.08 vs. 121.09). Most of the heavy lifting over external baselines seems to come from the Stage-1 patch autoencoder and label-prior initialization rather than spectral anchoring itself.
  2. Weak performance on propagation/channel tasks and reliance on a post-hoc seam refiner (Table 15 & Appendix E.6): In Table 15, TSTR/TRTS on channel-related labels (Delay Spread, Urban/Rural) is barely above random chance (+0.06 to +0.11). Also, the per-patch autoencoder still leaves seam artifacts at patch boundaries, requiring a separate post-hoc 1D CNN refiner for SNR >= 30 dB (Appendix E.6). Finally, WaveScene is purely simulated with static channels and a fixed deterministic payload bitstream (Appendix C.2), without real over-the-air (OTA) validation.
Questions for authors
  1. Attribution of gains (Stage 1 + Prior vs. Spectral Anchoring): Given that LF-M (without spectral anchoring) already achieves 83.49% TSTR vs. 84.49% for ALF in Table 3(a), and training the spectral anchor requires pretraining a separate 38.6M spectrogram autoencoder (Appendix E.3), what happens if you simply compute the STFT log-power spectrogram directly on the target (or use the Stage-1 Butterfly spectral branch features) as the anchor target instead of training a separate 38.6M spectral AE?
  2. Encoder probing fairness (Table 7 & Table 9): In Table 9, the 60.3M ALF encoder is explicitly trained with supervised label-alignment (L_CLIP and L_struct with 6 factor probes on the exact target labels) and sees the full recording length, whereas external baselines (WavesFM, IQFM, 0.34M–6.4M params) are self-supervised and restricted to 256-sample views. Isn't this probing comparison inherently tilted in ALF's favor?
  3. Near-chance channel classification (Table 15): Why does the model struggle so much on Delay Spread and Urban/Rural classification (Table 15) even though RMS delay spread and scene type are explicitly supervised in Stage 1 and injected into the Stage-2 prior?
5 comments
AuthorsOpenReview ↗

Response to reviewer sPqo - Part 1/3

Dear Reviewer sPqo,

We thank the reviewer for the careful reading and constructive suggestions. We address the concerns below.

(Q1 (a)) Attribution of gains from spectral anchoring. Given the modest improvement over LF-M in aggregate TSTR and spectrogram MSE, what additional benefits does spectral anchoring provide?

(A1 (a)) Gains from spectral anchoring.

Spectral anchoring helps preserve fine-grained fidelity in the spectral domain, particularly in bandwidth accuracy and the preservation of structured time–frequency features. With newly expanded Appendix E.11, we now demonstrate these benefits using two additional experiments and highlight theoretical benefits that were previously not sufficiently emphasized in Appendix A. Both bandwidth accuracy and preservation of structured time–frequency features are important for functional wireless signal generation. For example, frequency-modulated chirps, a fundamental tool in radar technology, enable longer-duration, higher-energy transmissions while maintaining range resolution through pulse compression. Their practical utility depends critically on preserving their characteristic frequency evolution.

Bandwidth fidelity. We demonstrate how spectral anchoring improves adherence to requested occupied bandwidth in Table 14 of Appendix E.11. Narrowband signals are particularly sensitive to relative bandwidth errors: energy leakage outside the desired band can interfere with neighboring channels. To quantify bandwidth fidelity, we consider identical subsets of the held-out 100k test sets for ALF and LF-M (without spectral anchoring), with occupancy labels between 0 and 0.2 and SNR above 20 dB. These subsets contain 24,272 records. We measure bandwidth noncompliance by the absolute difference between requested occupancy and measured 99%-energy occupied bandwidth, normalized by sampling rate, where \mathrm{OccBW}_{99\\%}=\frac{\mathrm{BW}_{99\\%}}{f_s}. As shown in Table 14, under identical conditions, ALF reduces the mean occupied-bandwidth error by 14.23% relative to LF-M, demonstrating improved adherence to spectral occupancy conditions.

Spectrogram fidelity. Spectral anchoring also reduces spectrogram errors for nonstationary waveforms such as radar chirps and hopping sequences, as shown in Table 15. Following the evaluation setting of Table 3, we compare ALF and LF-M on the same held-out 100k test set using identical conditioning labels. We restrict the analysis to SNR ≥ 0 dB to include low-SNR regimes while excluding signals below the noise floor. We measure spectrogram MSE (Spec-MSE) between each generated recording and its reference waveform, then analyze 10 nonstationary and 16 stationary waveform classes separately, averaging within each class before averaging across classes. ALF reduces Spec-MSE by 7.8 dB² on nonstationary waveforms, compared with 1.2 dB² on stationary waveforms, representing a 5.4× larger relative improvement. This demonstrates that spectral anchoring is particularly beneficial for signals with evolving spectral structure.

Theoretical benefits. In the revised paper, we have emphasized the theoretical analysis included in Appendix A. From a theoretical standpoint, spectral anchoring eases the generative modeling task by encouraging structural constraints that are challenging to learn purely in the time domain. For a sequence containing n discrete points spanning d levels, an unconstrained joint distribution demands dn−1d^n-1 free parameters. RF waveforms are inherently band-limited: despite containing n temporal samples, their spectral representations have restricted support. Through the inverse Fourier transform, the retained spectral coefficients determine the sequence, illustrating that time-domain samples are tightly coupled rather than independent parameters. Incorporating a spectral anchor s leverages this structure through p(x,s)=p(s)p(x∣s)p(x,s)=p(s)p(x\mid s). As discussed in Appendix A, a support-conditioned formulation reduces the required parameter space to at most ∣K∣dk−1|\mathcal{K}|d^k-1, where K\mathcal{K} denotes valid frequency support, compared with dn−1d^n-1 in an unconstrained formulation. This motivates spectral anchoring as a physically informed inductive bias for flow learning.

AuthorsOpenReview ↗

Response to reviewer sPqo - Part 2/3

(Q1 (b)) Could direct STFT log-power spectrograms or Stage-1 spectral features replace the separately trained spectrogram autoencoder?

(A1 (b)) Why a separate latent spectral encoder. Regarding the choice of spectral anchor, Stage 2 learns to transport representations toward the reconstruction-oriented IQ latent space established in Stage 1. The Stage 1 spectral branch is fused with raw-IQ features and jointly optimized for reconstruction and label alignment; its information is therefore already incorporated into the final latent representation that the flow aims to learn. Reusing these features as intermediate supervision would provide a target closely related to the flow's endpoint, rather than an independent spectral constraint. Instead, we seek an intermediate target that retains important time-frequency structure while abstracting away finer IQ details. The separately trained spectrogram autoencoder provides such a compact, spectrally focused representation, allowing the flow to learn global spectral structure before completing the transport toward the full IQ latent. As for direct STFT log-power supervision, it would require an additional mapping to produce compact, context-aware spectral embeddings compatible with the flow's intermediate representations. This is precisely the role of our separately trained spectrogram autoencoder, which learns such embeddings through a dedicated spectral reconstruction objective rather than relying on an arbitrary projection.

(Q2) Encoder probing fairness. The ALF encoder uses supervised label alignment, has greater capacity, and processes longer recordings than the external pretrained baselines. Does this make the comparison inherently favorable to ALF?

(A2) Encoder comparison and probing fairness.

Wireless recordings often contain structures spanning thousands of samples, including packet framing, symbol sequences, and temporally varying waveform patterns. For example, a 40-MHz IEEE 802.11ac WiFi packet includes synchronization preambles, channel-estimation training fields, and PHY signaling headers spanning over 1,600 IQ samples before the payload begins. Our goal is to compare the practical capabilities of RF encoders on realistic recordings, including the longer protocols. Models restricted to short input windows may miss information needed to represent extended structures. The ability to process and represent longer recordings is therefore an important capability in its own right, rather than merely an experimental advantage to be controlled away.

In Table 9, we use the same downstream training dataset, classification tasks, and probing protocol for all encoders, with best-effort adaptations preserving the baselines' original architectures and input requirements. ALF achieves substantially stronger performance, demonstrating the practical value of its long-context representation. We recognize that the models differ in supervision, capacity, and input context, so this comparison evaluates practical downstream capabilities rather than claiming that ALF's architecture or training objective is intrinsically superior under matched resources.

We now clarify the scope of this comparison in Appendix D.2, where we detail the baselines' native training and input procedures and explicitly distinguish practical representation performance from architectural superiority under matched resources. We also additionally report synthetic-to-real representation evaluations of the encoders in Appendix H.4.

(Q3) Near-chance channel classification. Why does ALF struggle with delay-spread and scene-related classification despite conditioning on these physical attributes?

(A3) Channel-related classification.

Some channel-conditioned generation attributes depend strongly on received signal power, while others require recovery of finer multipath structure (see Section 6: Limitations and Appendix G.2). Our results in Table 19 (previous Table 15) demonstrate the former: after restoring received-power information using WaveScene oracle rescaling, ALF's LOS/NLOS TSTR increases from 0.57 to 0.85, and indoor/outdoor TSTR increases from 0.56 to 0.75.

Delay spread presents a more challenging objective because it requires fine-grained classification and is calculated from the power-weighted distribution of multipath delays. Different propagation environments can also exhibit overlapping delay-spread distributions, as illustrated in Figure 7. Our primary objective is to learn wireless-signal distributions under diverse physical conditions, and we acknowledge that fine-grained channel-conditioning fidelity remains an area for improvement.

AuthorsOpenReview ↗

Response to reviewer sPqo - Part 3/3

(Q4) Reliance on a high-SNR seam refiner. The patch-based autoencoder leaves boundary artifacts, requiring a separate post-processing network. Are ALF's improvements dependent on this refiner?

(A4) Seam-refiner ablation.

The high-SNR seam refiner is an optional post-processing component that reduces residual discontinuities at patch boundaries. To isolate its contribution, we evaluated ALF and LF-M under matched settings without applying the refiner. For waveform classification, ALF achieves 83.91% TSTR and 80.32% TRTS, compared with 83.01% TSTR and 77.32% TRTS for LF-M, averaged over five seeds. These results confirm that ALF's improvements over LF-M persist independently of the seam refiner and are not attributable to post-processing.

(Q5) Lack of real over-the-air validation. WaveScene is simulated; does ALF provide meaningful utility on real RF recordings?

(A5) Real-world validation.

We have added Appendix H to evaluate ALF on real over-the-air measurements collected in a controlled indoor environment and on the outdoor NSF PAWR POWDER testbed. Appendix H.1–H.2 document the collections and preprocessing procedures. In Appendix H.3, Tables 20–21 show that ALF-generated augmentation substantially improves classification accuracy when real training records are limited, approaching the performance obtained with full-real training. We further evaluate synthetic-to-real representation transfer in Appendix H.4 (Tables 22–25). The frozen ALF encoder outperforms the evaluated RF encoder baselines both when synthetic-trained classification probes are applied directly to Over-The-Air (OTA) data and when new lightweight probes are trained on OTA recordings. Figures 20–21 additionally illustrate that the learned waveform embedding structure remains meaningfully organized after transfer to real measurements. These results provide complementary evidence of practical generation utility and transferable representations beyond WaveScene.

Concluding Remark

We hope the clarifications and additional results above address your concerns. If you are satisfied with our response, we kindly ask you to consider revisiting your score. We welcome any further questions and are committed to addressing additional concerns during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.