acceptodds
Under review as a conference paper at ICLR 2027

G-Wave: Bit-to-Waveform Generation for Adaptive Wireless Communications

Abstract

Wireless waveforms are traditionally hand-designed for general propagation conditions, rather than adapted to each specific environment. Existing neural transceivers seek such adaptation, but often rely on relatively simple architectures with limited long-range modeling capability. More expressive generative models such as Transformers are difficult to apply directly because wireless communication demands extremely accurate bit recovery, often with BER below , while providing no ground-truth waveform supervision. To address these limitations, we introduce , a first generative transceiver that directly synthesizes environment-adaptive physical time-domain waveforms from bits. consists of a new Transformer-based transmitter and receiver models equipped with a multiscale architecture that combines long-range sequence modeling with fine-grained signal synthesis to meet the stringent precision required for communication. The two models are jointly optimized end-to-end using only bit-reconstruction error, allowing the waveform itself to be discovered without waveform labels. Across diverse 3GPP wireless environments, achieves lower BER than the SOTA neural transceiver (DeepOFDM) and a strong conventional baseline (OFDM) under estimated channel information, reducing mean BER by and , respectively.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Waveform design, AI for Wireless, Generative Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_d5StOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes G-Wave, an end-to-end neural transceiver that directly maps LDPC-coded bits to time-domain IQ waveforms and maps the received waveform back to coded-bit logits, without relying on a OFDM resource grid. G-Wave shows constantly better result compared to DeepOFDM, DeepOFW and all the other baselines.

Critical strengths

G-Wave consistently outperforms all baselines, even under perfect CSI. The paper provides good analyses of what the learned waveform actually does. Fig. 6 clearly shows it for several cases. The authors also honestly report cases when spectral adaptation is not correct in the appendix.

Critical weaknesses

The paper has several weaknesses. Fairness to the baselines, an overly simple setting, limited structural novelty in the model, and the use of BER as the only metric. I detail these below in the questions. My biggest question is this: if we only consider stationary users, why should we not just use OFDM? Where does all of G-Wave's gain actually come from?

Questions for authors
  1. Most critically, the "No conditioning" variant of G-Wave (9.22e-3) performs on par with DeepOFDM (9.52e-3). The proposed method should therefore be compared against OFDM with water-filling based on CSI, or at least on estimated CSI. In such a comparison, I suspect the benefit of the framework itself would be marginal. I am not claiming that there is no effect. However, if the improvement comes mainly from the conditioning itself, then most of the contribution cannot be attributed to the proposed framework.
  2. The experimental setting is overly simplistic. Single-user SISO, CDL-A~E channels, and a static channel at 0 m/s. The paper claims SOTA using only standardized LOS/NLOS profiles, without considering site-specific propagation (e.g., via ray tracing). This limits the practical applicability of the results. In particular, user mobility is crucial for this line of work, as emphasized in DeepOFDM, and the complete absence of mobility is a critical limitation.
  3. The delay spread used for training is 0.1–0.6 μs, which seems too long compared to typical indoor settings.
  4. Although a Transformer is used, isn't the model essentially trained in a supervised manner? It is neither a diffusion model nor even a GAN, so it is unclear in what sense the method is "generative", in what sense it learns a distribution. It is certainly a neural TX/RX model, but I do not think it qualifies as a generative model.
  5. DeepOFW was originally designed for PAPR reduction, so comparing against it solely in terms of BER seems unfair to DeepOFW.
  6. As I understand it, G-Wave's gain comes from channel-adaptive power allocation across frequency. However, this could also be achieved within DeepOFDM, and OFDM-based approaches that operate on a fine-grained per-subcarrier basis could likely achieve it even more explicitly. If gridity was the problem, ablation could be done with it. Yet no ablation addresses this. In other words, even if G-Wave is state of the art among the compared methods, it is not evident that the improvement stems from properties unique to G-Wave.
  7. The CP/pilot overhead is present only on the OFDM side, which makes the comparison unfair.
  8. Only the arithmetic mean of BER over the evaluated Eb/N0 points is reported. Neither BLER nor the dB gap at a target BER is provided.
  9. The architecture combines VAR-style multiscale modeling [A] with DiT's AdaLN-Zero conditioning [B], and training is an end-to-end AE with a BCE loss [C], which is not particularly novel. Compared with [A][B][C], there is nothing architecturally distinctive. This is a significant weakness for an ICLR submission. Had the same paper been submitted to a communications venue, this might not have been a major concern. References [A] K. Tian, Y. Jiang, Z. Yuan, B. Peng, and L. Wang. Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction. NeurIPS, 2024. [B] W. Peebles and S. Xie. Scalable Diffusion Models with Transformers. ICCV, 2023. [C] T. O'Shea and J. Hoydis. An Introduction to Deep Learning for the Physical Layer. IEEE Trans. Cognitive Communications and Networking, 3(4):563–575, 2017.
6 comments
AuthorsOpenReview ↗

Responses for Q1 and Q2

Q1: We thank the reviewer for raising this point and will clarify it in the revision. G-Wave consistently outperforms DeepOFDM under matched settings, both with and without CSI conditioning. With the same CSI input, G-Wave reduces mean BER by 48.5% and 42.4% over DeepOFDM under perfect and estimated CSI, respectively. More importantly, even without any CSI conditioning, G-Wave achieves a lower BER than CSI-conditioned DeepOFDM (9.22\times10^{-3}) vs. (9.52\times10^{-3})). Therefore, the gain cannot be attributed primarily to conditioning; the proposed framework itself provides a substantial advantage.

To further isolate the benefit beyond conventional CSI-aware adaptation, we additionally compare against OFDM with CSI-based water-filling under both perfect and estimated CSI. G-Wave still consistently achieves lower BER, further confirming that its advantage extends beyond CSI conditioning or power allocation.

Table 1. Mean information-bit BER with CSI conditioning and water-filling.

MethodMean information-bit BER
G-Wave, perfect CSI4.90×10−34.90\times10^{-3}
DeepOFDM, perfect CSI9.52×10−39.52\times10^{-3}
OFDM + water-filling, perfect CSI1.07×10−21.07\times10^{-2}
G-Wave, no conditioning9.22×10−39.22\times10^{-3}
DeepOFDM, no conditioning13.3×10−313.3\times10^{-3}
G-Wave, estimated CSI5.74×10−35.74\times10^{-3}
OFDM + water-filling, estimated CSI2.44×10−22.44\times10^{-2}

G-Wave thus retains lower mean BER than CSI-based water-filling OFDM under both perfect and estimated CSI.

Q2: Prior neural communication and waveform-learning works, including DeepOFDM, have adopted similar single-user SISO settings with standardized channel models. We follow this established setup to enable a controlled and fair comparison with prior work. Moreover, 3GPP CDL-A--E models cover diverse LOS/NLOS conditions, multipath structures, and delay spreads [2], making them meaningful benchmarks for waveform design.

To further evaluate generalization beyond standardized channel models, we additionally test G-Wave, without retraining, on two unseen site-specific ray-tracing channel datasets: Sionna RT Munich [5] and DeepMIMO ASU campus [6]. G-Wave achieves the lowest mean information-bit BER on both, reducing BER by 46.9% and 49.0% relative to the strongest baselines, respectively.

Regarding mobility, although G-Wave is not explicitly trained for mobile channels, our additional evaluation shows that it maintains lower mean information-bit BER than DeepOFDM at all tested speeds up to 72 km/h under estimated CSI. At 72 km/h, G-Wave achieves a mean BER of 3.48×10−33.48\times10^{-3}, compared with 4.94×10−34.94\times10^{-3} for DeepOFDM. We will include these results in the revised manuscript to demonstrate that the framework is not inherently restricted to static channels.

Table 2. Mean BER on unseen ray-tracing channels.

Sionna RT Munich

MethodMean coded-bit BER ↓\downarrowMean information-bit BER ↓\downarrow
G-Wave2.19×10−3\boldsymbol{2.19\times10^{-3}}1.20×10−3\boldsymbol{1.20\times10^{-3}}
DeepOFW4.57×10−34.57\times10^{-3}3.30×10−33.30\times10^{-3}
DeepOFDM7.72×10−37.72\times10^{-3}2.26×10−32.26\times10^{-3}
OFDM2.45×10−22.45\times10^{-2}2.63×10−32.63\times10^{-3}

DeepMIMO ASU campus

MethodMean coded-bit BER ↓\downarrowMean information-bit BER ↓\downarrow
G-Wave4.94×10−3\boldsymbol{4.94\times10^{-3}}2.77×10−3\boldsymbol{2.77\times10^{-3}}
DeepOFW8.41×10−38.41\times10^{-3}5.43×10−35.43\times10^{-3}
DeepOFDM1.43×10−21.43\times10^{-2}5.93×10−35.93\times10^{-3}
OFDM3.14×10−23.14\times10^{-2}6.85×10−36.85\times10^{-3}

Table 3. Mean information-bit BER under mobility with estimated CSI (5, 10, and 15 dB).

Speed (km/h)G-WaveDeepOFDM
02.72×10−3\mathbf{2.72\times10^{-3}}4.72×10−34.72\times10^{-3}
92.85×10−3\mathbf{2.85\times10^{-3}}4.97×10−34.97\times10^{-3}
182.67×10−3\mathbf{2.67\times10^{-3}}4.80×10−34.80\times10^{-3}
272.65×10−3\mathbf{2.65\times10^{-3}}4.39×10−34.39\times10^{-3}
362.84×10−3\mathbf{2.84\times10^{-3}}5.01×10−35.01\times10^{-3}
542.95×10−3\mathbf{2.95\times10^{-3}}4.58×10−34.58\times10^{-3}
723.48×10−3\mathbf{3.48\times10^{-3}}4.94×10−34.94\times10^{-3}
1084.06×10−24.06\times10^{-2}1.78×10−2\mathbf{1.78\times10^{-2}}
AuthorsOpenReview ↗

Responses for Q3 to Q7

Q3: We use the (0.10.1)–(0.60.6,μs\mu\text{s}) delay-spread range during training to expose G-Wave to more challenging, highly frequency-selective channels and better explore its adaptation capability. For indoor evaluation, we follow 3GPP TR 38.901 [2] and use the more typical (2020)–(6060) ns range. Despite being outside the training distribution, G-Wave still achieves the lowest mean BER under both perfect and estimated CSI, demonstrating good generalization to typical indoor channels.

Q4: G-Wave connects the bit space and waveform space through a learned latent representation. The tokenizer first encodes the payload bits into latent features. The channel condition then guides the transformation of these features, and the synthesis head converts them into a physical waveform. The bits specify the information to be transmitted, while the condition guides how that information is represented in the waveform.

For a fixed channel condition, different bit sequences produce a collection of waveforms whose distribution is induced by the bit distribution. Changing the condition changes the latent transformation and consequently the waveform distribution. Through joint training with the receiver, G-Wave learns these condition-dependent representations for reliable bit recovery, without predefined waveform labels. This learned transformation from a bit distribution, through conditioned latent representations, to physical waveforms is the mechanism underlying our use of “generative waveform”. We will make this relationship explicit in the revised manuscript.

Q5: For a fair comparison using the same training objective, we specifically trained DeepOFW without a PAPR penalty. We now also report PAPR for this baseline and G-Wave. With 4×4\times oversampling, G-Wave has a mean PAPR of 7.327.32 dB, compared with 8.738.73 dB for BCE-only DeepOFW; at a CCDF level of 10−310^{-3}, the corresponding values are 10.7610.76 and 11.5211.52 dB. Thus, the additional comparison reports peak-power behavior as well as error rates, with the baseline training objective explicitly specified.

Table 4. PAPR with 4×4\times oversampling (dB).

MethodMeanCCDF 10−210^{-2}CCDF 10−310^{-3}
G-Wave7.3210.1610.76
DeepOFW, BCE only8.7310.6411.52
DeepOFDM, with PAPR penalty10.3411.5911.82
DeepOFDM, without PAPR penalty10.5812.5613.34
OFDM (QPSK)8.5110.3310.99

Q6: We clarify that G-Wave’s gain does not come solely from channel-adaptive power allocation. Rather, the framework jointly learns several aspects of waveform design: (1) a flexible bit-to-waveform representation, where information is distributed across multiple waveform samples; (2) channel-adaptive power allocation, which reshapes transmitted energy according to the channel condition; and (3) channel-adaptive overhead allocation, which avoids the fixed pilot/CP overhead of conventional OFDM and can instead adjust the amount of redundancy according to the channel condition.

These effects are learned jointly through end-to-end optimization and cannot be reduced to water-filling alone. To isolate the contribution beyond channel-adaptive power allocation, we additionally evaluate OFDM with CSI-based water-filling. It still exhibits substantially higher mean BER than G-Wave under both perfect and estimated CSI ((1.07×10−21.07\times10^{-2}) vs. (4.90×10−34.90\times10^{-3}), and (2.44×10−22.44\times10^{-2}) vs. (5.74×10−35.74\times10^{-3}), respectively).

Regarding resource-grid constraints, the requested ablation is already included in Table 2. The “Frequency-domain I/O” and “OFDM resource-grid I/O” variants retain the same neural backbone and conditioning while progressively constraining waveform generation, yielding BERs of (5.04×10−35.04\times10^{-3}) and (9.69×10−39.69\times10^{-3}), respectively, compared with (4.57×10−34.57\times10^{-3}) for full G-Wave. These results provide further evidence that G-Wave’s gain comes from the joint optimization of waveform representation, channel-adaptive power allocation, and adaptive resource overhead, rather than from power allocation alone. We will clarify this distinction more explicitly in the revised manuscript.

Q7: We clarify that the comparison uses the same number of bits, the same transmission duration, bandwidth, and the same frame-energy budget for all methods. Conventional OFDM allocates part of this fixed time-frequency resource to CP and pilots because they are required for channel estimation and equalization. G-Wave does not explicitly reserve such structures, but must instead learn the required robustness and channel adaptation end-to-end within the same transmission budget. Therefore, G-Wave is not given additional bits, time, or energy; the difference lies only in how each method utilizes the same available transmission resources.

AuthorsOpenReview ↗

Response to Q8

Q8: We thank the reviewer for this valuable suggestion. We have conducted additional BLER evaluations and provide a more detailed analysis of the SNR gains at target error rates. At an information-bit BER of 10−310^{-3}, G-Wave achieves gains of approximately 2.142.14 dB and 1.501.50 dB over DeepOFDM under perfect and estimated CSI, respectively. Furthermore, under perfect CSI, G-Wave reaches 1%1\% BLER at 7.837.83 dB, compared with 10.4210.42 dB for DeepOFDM, yielding a 2.592.59 dB advantage. These additional results further demonstrate G-Wave's reliability improvements beyond arithmetic-mean BER. We will incorporate these analyses into the revised manuscript.

Table 5. Mean BLER and SNR at 1%1\% BLER (Exp. 1).

Perfect CSI

MethodMean BLEREb/N0E_b/N_0 at 1%1\% BLER (dB)Gap to G-Wave (dB)
G-Wave1.90×10−21.90\times10^{-2}7.830.00
DeepOFW, BCE only2.24×10−22.24\times10^{-2}9.031.20
DeepOFDM3.73×10−23.73\times10^{-2}10.422.59
DeepOFDM, no conditioning5.25×10−25.25\times10^{-2}12.244.41
DeepOFDM, BCE only3.12×10−23.12\times10^{-2}10.062.23
OFDM-LMMSE4.37×10−24.37\times10^{-2}12.314.48

Estimated CSI

MethodMean BLEREb/N0E_b/N_0 at 1%1\% BLER (dB)Gap to G-Wave (dB)
G-Wave2.10×10−22.10\times10^{-2}8.250.00
DeepOFW4.41×10−14.41\times10^{-1}>15>15—
DeepOFDM3.89×10−23.89\times10^{-2}11.152.90
OFDM-LS9.53×10−29.53\times10^{-2}>15>15—

Table 6. SNR and gaps at information-bit BER 10−310^{-3}.

MethodPerfect CSI: Eb/N0E_b/N_0 at BER 10−310^{-3} (dB)Perfect CSI: Gap to G-Wave (dB)Estimated CSI: Eb/N0E_b/N_0 at BER 10−310^{-3} (dB)Estimated CSI: Gap to G-Wave (dB)
G-Wave12.260.0012.440.00
DeepOFDM14.402.1413.931.50
DeepOFW12.850.59>15>15—
OFDM14.952.68>15>15—
Reviewer_tsyEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the G-Wave architecture to address the challenge of environmental adaptability in physical-layer waveform design for wireless communications. The architecture models waveform generation as a condition-guided sequence transformation task, employing a multi-scale residual Transformer with an AdaLN-Zero conditional injection mechanism. Conditioned on Channel State Information (CSI) and noise variance, it directly synthesizes time-domain complex baseband waveforms from encoded bits and recovers the bits at the receiver. Experiments are conducted under 3GPP CDL-A to E channel models, evaluating the Bit Error Rate (BER) performance under both perfect CSI and estimated CSI conditions. The results demonstrate that, on a mixed channel test distribution, G-Wave reduces the average BER by 42.4% and 72.6% compared to DeepOFDM and conventional OFDM, respectively. The paper also provides spectrum adaptation analysis, multi-scale branch contribution analysis, and influence/perturbation analysis to enhance the interpretability of the model.

Critical strengths
  1. The multi-scale residual decomposition mechanism draws inspiration from the coarse-to-fine structure of multi-carrier systems in classical communications and combines it with the long-range modeling capability of Transformers, making it a rational architectural choice.
  2. The experimental results are outstanding; G-Wave achieves significant BER performance improvements compared to DeepOFDM and conventional OFDM.
  3. Other experimental results and analyses are relatively comprehensive, and the relevant ablation studies validate the soundness of the model architecture design.
Critical weaknesses
  1. The authors repeatedly cite generative models as the theoretical motivation in the abstract and introduction. However, generative models inherently focus on modeling data distributions, whereas the essence of G-Wave does not involve data distribution modeling or sampling. It is essentially a conditional autoencoder rather than a generative model in the strict sense. More importantly, if the authors approach this from the perspective of generative modeling, they should at least introduce generative models such as GANs or Diffusion models at the experimental level. Otherwise, it cannot be proven whether current generative models already possess the capability to accomplish efficient communications.
  2. The authors ambitiously claim in the introduction that the trained model can "automatically discover the optimal waveform for a given environment," yet the paper lacks theoretical analysis supporting this claim. First, the authors have not proven that the G-Wave architecture is capable of discovering the optimal waveform; merely providing experimental results cannot substantiate the claim of "optimality." Second, if the paper claims to achieve "optimal" results, it should provide the theoretical upper bound of this "optimality" and analyze whether the model has reached such an upper bound.
  3. The authors claim that "generalization ability" is the core contribution of G-Wave; however, the current theoretical analysis, experimental validation, and scenario settings are far from adequate. First, the authors have not theoretically proven why the model possesses generalization capabilities. Second, although the authors' experiments are validated under three scenarios—indoor, rural, and urban—the corresponding CDL models (A-E) and delay spread ranges (0.1-0.6 µs) for these scenarios are entirely contained within the training distribution. This does not constitute true generalization capability in the machine learning sense. Furthermore, out-of-distribution (OOD) scenarios, such as extreme delay spreads exceeding the training range or channel models unseen during training, have not been evaluated.
  4. The paper lacks an explanation of overheads such as computational complexity. As a signal model, practicality is a crucial aspect; if the model overhead is too high, it will directly compromise its practicality. It is recommended to supplement the paper with quantitative analyses of the inference latency, memory footprint, etc., for both G-Wave and the baselines.
Questions for authors
  1. Please clarify the "generative" positioning of G-Wave. Given that the training objective is solely the BCE Loss and does not involve distribution modeling or sampling, could the authors justify from the perspective of machine learning theory why G-Wave should be categorized as a generative model rather than a conditional autoencoder?
  2. Please provide a performance comparison between at least one SOTA generative model and G-Wave under the same task and experimental conditions.
  3. Referring to Weakness 3, please supplement the paper with experiments in realistic out-of-distribution (OOD) scenarios.
  4. Referring to Weakness 4, please supplement the paper with quantitative analysis results regarding the inference latency, memory footprint, etc., for both G-Wave and the baselines.
5 comments
AuthorsOpenReview ↗

Responses for Q1–Q3

Q1: We thank the reviewer for highlighting this distinction. G-Wave connects the bit space and waveform space through a learned latent representation. Specifically, coded bits are projected into latent features, which are transformed by a channel-conditioned Transformer and synthesized into physical waveforms. Mathematically, given coded bits C∼PCC\sim P_C and a channel condition d\mathbf{d}, the transmitter generates waveforms as X=fθ(C,d)X=f_\theta(C,\mathbf{d}), where fθf_\theta is the learned transmitter mapping. Since the input bits are randomly sampled, this mapping induces a conditional waveform distribution Pθ(⋅∣d)P_\theta(\cdot\mid\mathbf{d}). Changing d\mathbf{d} modulates the latent representations and consequently changes the induced waveform distribution. Thus, the bits determine the information content, while the channel condition guides its waveform representation. This transformation is learned through end-to-end bit reconstruction without waveform-level supervision. We use “generative” to describe this condition-guided waveform synthesis, rather than explicit target-distribution modeling, and will clarify this distinction in the revised manuscript.

Q2: We thank the reviewer for this suggestion. As discussed in Related Work (Sec. 2), existing generative approaches in wireless communications primarily focus on individual tasks, such as RF signal generation and channel estimation, rather than end-to-end bit-to-waveform synthesis and recovery. To provide a meaningful comparison, we integrate the diffusion-based channel estimator of Zhou et al. [1] into a conventional OFDM transceiver under the same settings. Although diffusion improves channel estimation NMSE by 5.3 dB over LS estimation, G-Wave still achieves substantially lower information-bit BER (5.46×10−35.46\times10^{-3} vs. 1.44×10−21.44\times10^{-2}). This demonstrates the advantage of our end-to-end learned waveform synthesis framework over the evaluated diffusion-assisted OFDM pipeline.

Table 1. Comparison with diffusion-based channel estimation.

MethodNMSE ↓\downarrow (dB)Coded BERInfo BERBLEREb/N0E_b/N_0 at 1%1\% BLER (dB)
G-Wave−7.6-7.68.05×10−3\mathbf{8.05\times10^{-3}}5.46×10−3\mathbf{5.46\times10^{-3}}2.10×10−2\mathbf{2.10\times10^{-2}}8.25
OFDM + diffusion estimation−12.9-12.94.01×10−24.01\times10^{-2}1.44×10−21.44\times10^{-2}7.33×10−27.33\times10^{-2}>15>15

Q3: We thank the reviewer for this suggestion. We clarify that our original evaluation already includes delay-spread OOD scenarios: the training range is 0.1–0.6 μ\mus, whereas the indoor-office range (20–60 ns) lies entirely outside it (Sec. 5.2). To further evaluate generalization, we additionally test G-Wave and all baselines under perfect CSI on two unseen site-specific ray-tracing channel sources, Sionna RT Munich [3] and DeepMIMO ASU campus [4], without retraining or fine-tuning. G-Wave still achieves the lowest mean information-bit BER on both.

Table 2. Mean BER on unseen ray-tracing channels (perfect CSI).

Sionna RT Munich

MethodMean coded-bit BER ↓\downarrowMean information-bit BER ↓\downarrow
G-Wave2.19×10−3\boldsymbol{2.19\times10^{-3}}1.20×10−3\boldsymbol{1.20\times10^{-3}}
DeepOFW4.57×10−34.57\times10^{-3}3.30×10−33.30\times10^{-3}
DeepOFDM7.72×10−37.72\times10^{-3}2.26×10−32.26\times10^{-3}
OFDM2.45×10−22.45\times10^{-2}2.63×10−32.63\times10^{-3}

DeepMIMO ASU campus

MethodMean coded-bit BER ↓\downarrowMean information-bit BER ↓\downarrow
G-Wave4.94×10−3\boldsymbol{4.94\times10^{-3}}2.77×10−3\boldsymbol{2.77\times10^{-3}}
DeepOFW8.41×10−38.41\times10^{-3}5.43×10−35.43\times10^{-3}
DeepOFDM1.43×10−21.43\times10^{-2}5.93×10−35.93\times10^{-3}
OFDM3.14×10−23.14\times10^{-2}6.85×10−36.85\times10^{-3}
AuthorsOpenReview ↗

Response for Q4 and reference

Q4: We thank the reviewer for suggesting a quantitative efficiency analysis. We additionally profile all methods on an NVIDIA A100 with batch size one under identical frame settings. While G-Wave has higher inference latency than conventional OFDM and the lightweight DeepOFW, it requires 34.0% fewer parameters and only 46.0% of DeepOFDM's FLOPs. With XLA compilation, G-Wave achieves a latency of 3.68 ms, close to DeepOFDM's 3.23 ms. These results demonstrate that G-Wave achieves improved communication reliability with lower computational complexity and comparable optimized inference latency relative to DeepOFDM, offering a favorable performance–complexity trade-off. Notably, XLA provides general-purpose compiler optimization without G-Wave-specific kernel tuning, leaving further opportunities to reduce inference latency through customized CUDA kernels, operator fusion, and optimization for newer GPU architectures.

Table 3. Model complexity and inference latency (A100, batch size 1).

MethodParameters (millions)GFLOP per frameLatency (ms)Relative to G-Wave
G-Wave17.5911.03.68 (XLA)1.00
DeepOFW17.680.072.00 (GRU, non-XLA)0.54
DeepOFDM26.6723.93.23 (XLA)0.88
OFDM-LMMSE0≈0\approx 00.69 (XLA)0.19

Table 4 reports FP32 weight storage and additional inference GPU memory measured without XLA. G-Wave requires 67 MB for model weights, compared with 102 MB for DeepOFDM. At batch size one, its additional GPU memory is 9 MB, versus 10 MB for DeepOFDM and 47 MB for DeepOFW. At batch size 64, the corresponding values increase to 594, 242, and 122 MB, respectively. Thus, G-Wave has a smaller weight footprint than DeepOFDM and modest single-frame inference memory requirements, while its batched inference uses more memory.

Table 4. FP32 weight storage and inference GPU memory (non-XLA).

MethodFP32 weights (MB)Extra GPU memory batch size 1 (MB)Extra GPU memory batch size 64 (MB)
G-Wave679594
DeepOFW6847122
DeepOFDM10210242
OFDM-LMMSE002

References

[1] Xingyu Zhou, Le Liang, Jing Zhang, Peiwen Jiang, Yong Li, and Shi Jin. Generative diffusion models for high dimensional channel estimation. IEEE Transactions on Wireless Communications, 24(7):5840–5854, July 2025. doi: 10.1109/TWC.2025.3549592. URL https://ieeexplore.ieee.org/document/10930691.

[2] S. Ashwin Hebbar, Sravan Kumar Ankireddy, Harshithanjani Athi, Brandon Nguyen, Pramod Viswanath, and Hyeji Kim. Deep-OFDM: Neural modulation for high mobility. arXiv:2506.17530, 2025. URL https://arxiv.org/abs/2506.17530.

[3] NVIDIA. Introduction to Sionna RT. Sionna documentation. URL https://nvlabs.github.io/sionna/rt/tutorials/Introduction.html.

[4] DeepMIMO. Quickstart. DeepMIMO documentation. URL https://www.deepmimo.net/docs/quickstart.html.

[5] Ran Greidi and Kobi Cohen. DeepOFW: Deep learning-driven OFDM-flexible waveform modulation for peak-to-average power ratio reduction. arXiv:2603.23544, 2026. URL https://arxiv.org/abs/2603.23544.

The author address some of my concerns. However, my main concern still exists. The authors should explain and demonstrate why G-Wave achieves the best results. Please refer to Weakness 2 for further details.

Reviewer_K1FaOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces G-Wave, a wireless-environment-conditioned end-to-end label-free bit-to-waveform generative transceiver. G-Wave adopts a new residual multiscale architecture that captures coarse-to-fine temporal information modulation structures, facilitating precise waveform synthesis and reliable bit recovery. G-Wave achieves lower BER than Deep-OFDM, DeepOFW, and conventional OFDM.

Critical strengths
  • The framework directly optimizes transmitted waveforms without imposing a predefined modulation format.
  • Explicit CSI and noise conditioning enables coordinated adaptation of the transmitter and receiver.
Critical weaknesses
  • The motivation for the residual multiscale architecture is insufficiently developed. The stated inspiration does not clearly explain why coarse-to-fine processing is well suited to waveform generation and bit recovery.
  • The limited performance degradation of the single-scale variant raises questions about the benefit of the multiscale architecture.
  • The evaluation reports only post-LDPC information-bit BER, making it difficult to separate the transceiver’s detection performance from the effects of LDPC decoding.
  • It is unclear whether the baselines were retrained under the same channel conditions as the proposed method. If not, the reported gains may partly reflect a mismatch between training and evaluation conditions rather than architectural advantages.
  • A comparison of inference latency or computational cost against the baselines would help assess the significance of the reported performance gains.
  • The number of training steps and channel realizations should be provided to clarify the overall training budget.
Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Responses to Q1– Q4

Q1: We thank the reviewer for this valuable comment. The motivation for our residual multiscale architecture comes from both the time-frequency characteristics of OFDM signals and the challenges we encountered when applying conventional Transformer architectures to direct bit-to-waveform generation and bit recovery.

From a signal perspective, OFDM waveforms are formed by superimposing multiple orthogonal subcarriers, resulting in time-domain signals that contain variations across different temporal scales. This characteristic inspired us to explore multiscale representations for waveform modeling. From an architectural perspective, during our early-stage investigations, we found that conventional Transformer architectures with fixed patch/token representations overly constrained how bit information could be represented and distributed across the waveform, limiting their flexibility in modeling waveform characteristics at different temporal resolutions.

These considerations motivated our residual multiscale architecture, which enables the Transformer to progressively learn waveform representations at multiple temporal resolutions while preserving information through residual connections. We will further clarify this motivation in Sec. 4.2 to better explain the rationale behind our architectural design.

Q2: We agree that the absolute BER difference appears modest. However, OFDM is already a highly optimized communication framework, making further reliability gains difficult to achieve yet still highly valuable. G-Wave reduces mean BER by 72.6% over conventional OFDM under estimated CSI, while the multiscale design provides a further (∼10\sim10%) BER reduction over the single-scale variant (Table 2). At wireless network scale, even such incremental gains can be meaningful: for example, if they translate into only a 1% reduction in radio-resource usage, a 3-kW base station could save roughly 22 kWh per month. This illustrates the practical value of the multiscale design.

Q3: We thank the reviewer for this suggestion. As described in Sec. 4.2, LDPC decoding is not involved in training, and all methods use the same LDPC code and decoder for fair evaluation. To address this concern, we additionally report pre-LDPC coded-bit BER for all baselines, eliminating the influence of LDPC error correction. The results show that G-Wave consistently achieves the lowest BER under both perfect and estimated CSI, confirming that its performance advantage persists independently of LDPC decoding.

Table 1. Mean coded-bit and information-bit BER on mixed CDL-A–E channels.

MethodPerfect CSI: Coded BER (without LDPC)Perfect CSI: post LDPC BEREstimated CSI: Coded BER (without LDPC)Estimated CSI: post LDPC BER
G-Wave7.34×10−3\mathbf{7.34\times10^{-3}}4.83×10−3\mathbf{4.83\times10^{-3}}8.34×10−3\mathbf{8.34\times10^{-3}}5.74×10−3\mathbf{5.74\times10^{-3}}
DeepOFW9.08×10−39.08\times10^{-3}6.03×10−36.03\times10^{-3}1.30×10−11.30\times10^{-1}1.41×10−11.41\times10^{-1}
DeepOFDM1.64×10−21.64\times10^{-2}9.39×10−39.39\times10^{-3}1.62×10−21.62\times10^{-2}9.97×10−39.97\times10^{-3}
OFDM-LMMSE3.37×10−23.37\times10^{-2}1.03×10−21.03\times10^{-2}6.01×10−26.01\times10^{-2}2.11×10−22.11\times10^{-2}

Q4: All neural baselines were retrained under the same channel conditions as G-Wave and evaluated using the identical test protocol, ensuring a fair comparison. Under these matched training and evaluation conditions, G-Wave consistently achieves lower BER than the baselines across diverse channel scenarios under both perfect and estimated CSI, demonstrating that its performance advantages are not due to mismatched channel conditions. We will explicitly clarify these shared training settings in the revised manuscript to avoid ambiguity.

AuthorsOpenReview ↗

Responses to Q5 and Q6

Q5: We thank the reviewer for suggesting a computational efficiency comparison. We additionally profile all methods on an NVIDIA A100 with batch size one under identical frame settings. While G-Wave incurs higher inference latency than conventional OFDM and the lightweight DeepOFW, it requires 34.0% fewer parameters and only 46.0% of DeepOFDM's FLOPs. With XLA compilation, G-Wave's latency is 3.68 ms, close to DeepOFDM's 3.23 ms. These results demonstrate that G-Wave achieves substantial BER improvements with lower computational complexity and comparable optimized inference latency relative to DeepOFDM, offering a favorable performance–complexity trade-off. Moreover, our implementation has only undergone basic compiler-level optimization, leaving further opportunities for latency reduction through hardware-specific optimizations, such as custom CUDA kernels. Table 2. Model complexity and inference latency (A100, batch size 1).

MethodParameters (millions)GFLOP per frameLatency (ms)Relative to G-Wave
G-Wave17.5911.03.68 (XLA)1.00
DeepOFW17.680.072.00 (GRU, non-XLA)0.54
DeepOFDM26.6723.93.23 (XLA)0.88
OFDM-LMMSE0≈0\approx 00.69 (XLA)0.19

Q6: We train G-Wave for 120,000120{,}000 optimizer steps with a batch size of 1616, corresponding to 1.92×1061.92\times10^{6} training-frame exposures. Channel realizations are sampled during training according to the distribution described in Sec. 5.1. We will explicitly state the training budget and channel-sampling procedure in the revised manuscript.

I appreciate the authors' efforts to address my concerns. However, my main concern regarding the motivation and necessity of the multiscale architecture remains unresolved.

The authors motivate the multiscale architecture based on the time-frequency characteristics of OFDM signals. However, G-Wave does not explicitly follow the OFDM subcarrier structure, and the proposed decomposition operates on learned latent representations, not on physical waveforms. The time-frequency characteristics of OFDM alone do not provide a convincing justification for the proposed coarse-to-fine architecture. The authors have not adequately established how this design benefits waveform generation and bit recovery beyond the observed BER improvement.

AuthorsOpenReview ↗

Motivation and coarse-to-fine architecture analysis

Thank reviewer for this valuable comment. We would like to clarify that the primary contribution of G-Wave is the introduction of a Transformer-based generative transceiver for direct bit-to-waveform generation, rather than the multiscale architecture alone. The multiscale representation is introduced as an architectural enhancement to facilitate more effective waveform generation and bit recovery within this framework.

We agree that our multiscale decomposition operates in the learned latent space and does not explicitly correspond to OFDM subcarriers or physical frequency components. The time-frequency characteristics of OFDM serve as a source of inspiration rather than a direct architectural correspondence. Specifically, they motivate us to explore coarse-to-fine representations that allow the Transformer to jointly model long-range temporal dependencies and fine-grained signal structures, both of which are important for accurate waveform generation and bit recovery. Therefore, the performance improvement of G-Wave should be attributed to the overall Transformer-based generative framework, with multiscale modeling serving as a complementary enhancement rather than the sole source of improvement.

Importantly, our analysis of the multiscale architecture goes beyond the overall BER comparison by further investigating how coarse-to-fine latent representations influence the structure of the generated physical waveform and contribute to reliable bit recovery. In Appendix E (Figure 8), we have already investigated the contributions of individual scales to waveform generation and bit recovery by examining the waveform-energy contributions of transmitter branches, the bit-logit contributions of receiver branches, and the effects of removing individual scales on BER. The results demonstrate that different scales play meaningful roles in the learned transceiver. For example, although the coarser transmitter branches contribute relatively little waveform energy when activated individually, removing the two coarsest branches increases coded-bit BER by factors of 3.8 and 49, respectively, at 15 dB, highlighting their importance for reliable bit recovery.

To further clarify how the multiscale latent representations affect the generated physical waveform, we conducted an additional scale-wise gradient sensitivity analysis to quantify the influence of individual latent scales on waveform generation.

Table 1. Scale-wise gradient sensitivity of the generated waveform (Eb/N0=10E_b/N_0=10 dB).

TX scaleTokensShare of waveform-gradient energyPer-token share
T0124.4%89.1%
T1717.3%9.0%
T24917.2%1.3%
T324541.1%0.6%

Here, the gradient share is the normalized E∥∂x/∂tk∥22\mathbb E\lVert\partial\mathbf x/\partial\mathbf t_k\rVert_2^2 across scales, estimated from 512 frames. The per-token share is obtained by dividing each scale's gradient share by its token count and renormalizing.

Table 1 shows that all four latent scales exhibit measurable waveform-gradient sensitivity. Despite containing substantially fewer tokens, the three coarser scales (T0–T2) collectively account for 58.9% of the total waveform-gradient energy. In particular, T0 contains only one token yet accounts for 24.4% of the gradient energy, indicating a substantial influence on the generated waveform. These results show that coarse-scale representations play a meaningful role alongside fine-scale features in waveform synthesis.

Taken together, the above scale-wise gradient analysis and the branch-removal experiments already presented in Appendix E provide complementary evidence for understanding the contributions of different scales to waveform generation and bit recovery within G-Wave. We will incorporate the additional scale-wise sensitivity analysis and further clarify the architectural motivation in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.