CARNet: Channel-Adaptive Receiver Network for Robust NextG Communications

Chao Jiang; Yongli Yan; Zhuo Xu

REVIEW 4 major objections 5 minor 13 references

A mixture-of-experts receiver trained on a handful of 5G channel models can decode unseen channel types by using a learned latent embedding of the coarse channel estimate to route each transmission to a specialist expert.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 12:56 UTC pith:G77FCQHU

load-bearing objection A plausible MoE-based neural receiver with a clever VAE router, but the zero-shot generalization claim is under-evidenced. the 4 major comments →

arxiv 2608.02172 v1 pith:G77FCQHU submitted 2026-08-03 cs.IT cs.LGmath.IT

CARNet: Channel-Adaptive Receiver Network for Robust NextG Communications

Chao Jiang , Zhuo Xu , Yongli Yan This is my paper

classification cs.IT cs.LGmath.IT

keywords neural receivermixture of expertschannel adaptationOFDM5G NRgeneralizationvariational autoencodersignal detection

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper sets out to show that a neural receiver does not have to be retrained for every propagation environment. The proposed architecture, CARNet, replaces the single static network with a library of specialist expert networks and a router that reads the coarse least-squares channel estimate, compresses it into a low-dimensional latent embedding, and uses that embedding to activate the best expert for the current channel. The authors train on several standardized 5G channel models and test, without any retraining, on others that were never seen during training; they report lower block error rates than two fixed neural baselines and near-perfect-CSI performance on trained channels. If the result holds, a single adaptive receiver could serve diverse 5G/6G deployments and new environments without per-scenario retraining.

Core claim

On its own terms, the paper's central finding is empirical: a mixture-of-experts receiver with channel-conditioned routing generalizes better than single-network neural receivers. The router takes the pilot-based least-squares channel estimate, passes it through a VAE-style encoder to obtain a compact latent representation, and produces routing weights; only the top expert is activated, and its bit-wise log-likelihood-ratio output is the receiver's soft decision. Two-stage training (pretraining the encoder, then jointly fine-tuning with the experts plus a load-balancing loss) prevents expert collapse. The plotted results show CARNet matching a perfect-CSI benchmark on trained CDL-C and CDL-D

What carries the argument

The load-bearing mechanism is the routing network's learned latent embedding of the coarse channel estimate. A lightweight VAE-style encoder projects the high-dimensional least-squares pilot estimate into a low-dimensional vector (a mean and a log-variance), trained with a reconstruction loss and a KL term; at inference the decoder is discarded and only the latent statistics feed an MLP that emits normalized routing weights. Because channels have strong time-frequency-space correlation, the authors argue this embedding is an efficient, task-oriented summary of the propagation condition. It selects the single most relevant expert from a library of three homogeneous ResNet-based receivers, so

Load-bearing premise

The generalization claim stands or falls on the assumption that the held-out channel models (CDL-A, CDL-E, UMi) are genuinely out-of-distribution for a network trained on CDL-B/C/D and UMa, and that these simulated channels represent real propagation.

What would settle it

Estimate the Wasserstein distance between the delay-Doppler profiles (or channel impulse responses) of the training channels and the held-out channels in the same simulator; if that distance is small, the zero-shot gap is not evidence of generalization. A stronger falsifier is an over-the-air campaign at 3.5 GHz under UMi-like conditions showing CARNet's BLER no longer staying below the static baselines.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

A single CARNet instance can serve multiple standardized 5G/NR channel profiles (CDL-B/C/D, UMa) with one set of weights, replacing per-scenario retraining.
On an unseen urban-microcell profile (UMi), CARNet still decodes with lower block error rate than DeepRx and CRNN-ResNet, especially at high signal-to-noise ratios.
Because only the top-1 expert is activated, inference FLOPs (~4.2 GFLOPs/slot) stay close to a single-network receiver despite a larger parameter count.
The learned latent embedding of the coarse channel estimate is what makes routing reliable; removing it (CARNet-Base) raises both BLER and complexity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The smooth, KL-regularized latent space created for routing is likely to interpolate between trained channel profiles, so the same receiver could generalize to mixed or transitional conditions (e.g., a UE moving from UMa to UMi) without special handling; this is not tested in the paper.
Because each expert is structurally identical and only one is activated, CARNet offers a natural upgrade path: add a new expert for a new channel model and let the latent router learn to send unfamiliar embeddings there, without disturbing the existing experts; the paper does not explore this continual-learning use.
A quantitative check the paper leaves implicit is the distribution gap between the 'unseen' test channels and the training set. If the simulator generates UMi samples that overlap heavily with training UMa samples, the zero-shot result would be less surprising; measuring that distance would make the generalization claim testable.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A plausible MoE-based neural receiver with a clever VAE router, but the zero-shot generalization claim is under-evidenced.

read the letter

The paper is worth engaging with, but the evidence as it stands does not fully support the headline claim. What is actually new: the use of a VAE trained on LS channel estimates to produce a latent embedding that drives top-k expert selection in a MoE-based neural receiver. That is a sensible idea — channel conditions are low-dimensional, LS estimates are noisy, and the probabilistic embedding is a natural way to stabilize routing. The architecture is coherent, the two-stage training with load-balancing is well-motivated, and the ablation against CARNet-Base gives some support for the VAE router specifically.

The paper does a few things well. The complexity comparison is informative: conditional computation keeps FLOPs close to the smaller baselines even though parameter count is higher. The simulation setup follows standard practice with Sionna, and the CDL-C/CDL-D results show CARNet beating DeepRx and CRNN-ResNet on in-distribution channels.

The soft spots are the ones the stress test flags. The zero-shot generalization claim rests entirely on Fig. 3c, a single UMi scenario. The promised CDL-A/E unseen results are never shown. There are no error bars, no seed variance, and no attempt to quantify how far UMi is from the training distribution. The reader's concern about statistical overlap with UMa is fair. Given that Table I fixes delay spread and UE speed ranges across training and test, UMi may not be meaningfully out-of-distribution. Also, the baselines are smaller networks, so the BLER gap at high Eb/N0 could be capacity rather than routing. None of this contradicts the architecture's usefulness, but the load-bearing empirical claim is under-specified.

The paper makes no mathematical derivation, so soundness rests entirely on simulation reproducibility. No code or data is released, which amplifies the lack of training hyperparameters and error bars. Still, the central idea is plausible and clearly communicated. It deserves a serious referee, not a desk reject, but it needs substantial revision: more unseen scenarios, uncertainty quantification, and either code release or full experimental details.

Who is this for? Researchers working on ML-based physical layer who are looking for ways to avoid retraining neural receivers across channel models. I would bring it to a reading group discussion on adaptive receivers, and I would cite it as related work if I were writing in that space.

Recommendation: send to peer review, with the expectation of major revisions focused on the generalization evidence.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes CARNet, a mixture-of-experts (MoE) neural receiver for 5G-NR/OFDM uplink reception. The architecture consists of K=3 homogeneous ResNet-based expert networks, a VAE-based representation learning module that maps LS channel estimates to a low-dimensional latent embedding, and a routing network that selects the top-1 expert. Training is done in two stages: VAE pretraining (Eq. 9) followed by joint detection and load-balancing optimization (Eqs. 10–12). Link-level simulations using Sionna compare BLER against LS-LMMSE, DeepRx, CRNN-ResNet, and perfect-CSI baselines on CDL-C, CDL-D, and zero-shot UMi. The paper claims superior performance and improved generalization across channel conditions, including an unseen UMi scenario in Fig. 3c.

Significance. If the zero-shot generalization results are robust, CARNet would be a practically useful contribution: a single neural receiver could serve diverse 5G/6G channel conditions without retraining. The paper's strengths are its use of a standard link-level simulator with realistic LDPC-coded OFDM, comparison against established neural receiver baselines, an ablation against CARNet-Base, and a complexity table. The architecture is coherent and the two-stage training with load balancing is a reasonable way to address expert collapse. However, the current empirical support is under-specified: only one unseen scenario is shown, no confidence intervals are provided, and the baselines are not capacity-matched. These issues must be addressed before the central generalization claim can be accepted.

major comments (4)

[Sec. IV-A, IV-B; Fig. 3] The central zero-shot claim is not established. Sec. IV-A states that unseen scenarios include CDL-A/E and UMi, but only UMi results are reported; the promised CDL-A/E curves are absent. Moreover, Table I fixes identical RMS delay spread (10–300 ns) and UE speed (0–35 m/s) ranges for training and test, and no shift metric is provided to quantify how far UMi is from the training set (CDL-B/C/D + UMa). Without such a diagnostic, Fig. 3c may be an in-distribution comparison rather than a generalization test. Please add the missing CDL-A/E results, error bars, and a quantitative distribution-shift analysis (e.g., over LS estimates or latent embeddings).
[Fig. 3; Table II] No confidence intervals, seed variance, or number of Monte Carlo blocks are reported. All BLER curves appear to be single runs. The performance gap at high Eb/N0 in Fig. 3c could be within seed variance, especially where BLER is low and few error events occur. Report results over multiple seeds (or at least error bars) and state the number of simulated slots/blocks per BLER point.
[Sec. IV-B; Table II] The comparison is not capacity-matched. DeepRx and CRNN-ResNet have 1.196M and 1.273M parameters, while CARNet has 3.810M. The observed gains may reflect model capacity rather than channel-adaptive routing. The CARNet-Base ablation is also an MoE variant, so it does not isolate routing benefit against a static single-network receiver of comparable capacity. Include a matched-capacity static ResNet baseline (or a single-expert CARNet variant) and report expert activation distributions to demonstrate that routing, not just extra parameters, drives the improvement.
[Sec. III-D; Sec. IV-A] Key training hyperparameters are omitted: β in Eq. (9), α in Eq. (12), learning rates for both training stages, optimizer, batch size, number of epochs, and channel-sampling details (e.g., how delay spread/Doppler are drawn for each CDL model). Without these, the experiments are not reproducible, and the sensitivity of Eq. (12)'s load-balancing term to expert collapse cannot be assessed. Please provide full hyperparameters and a brief sensitivity analysis.

minor comments (5)

[Sec. IV-A] Typo: 'perfroms' should be 'performs'.
[Fig. 3] Axis label 'Eb=N0' should be 'Eb/N0'.
[Eq. (9)] The KL divergence term is written ambiguously; the 1/2 factor should be explicit, and the notation should clarify that σ² is the variance (not the standard deviation) throughout.
[Eq. (4); Fig. 2] Define X_G and X_E precisely: it is unclear whether the routing input is only pilot-region LS estimates or the interpolated full channel estimate.
[Sec. IV-B] The CARNet-Base ablation description is vague; specify the input and architecture of the DNN used in place of the representation learning module.

Circularity Check

0 steps flagged

No significant circularity; the empirical comparison is self-contained.

full rationale

The paper's central claim is empirical: CARNet (Eq. 4 with VAE-based routing Eq. 8 and ResNet experts) is evaluated on 5G NR link-level simulations against external baselines (DeepRx, CRNN-ResNet, LS-LMMSE, perfect CSI). Training uses CDL-B/C/D and UMa; the zero-shot UMi test is held out and is not used to fit any parameter or equation. There is no derivation chain in which an output is equivalent to an input by construction, no fitted parameter is renamed a prediction, and no uniqueness theorem or ansatz is imported from self-citations. The only self-citation, [9], is related prior work on energy-efficient receivers and is not load-bearing for CARNet's architecture or results. The CARNet-Base ablation compares two variants of the same system, but it is not circular because it is anchored by external baselines and does not define the main result in terms of itself. Concerns about whether UMi is truly out-of-distribution, missing CDL-A/E curves, and absence of error bars are validity or reporting issues, not circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 1 invented entities

The central claim is empirical, so the ledger is dominated by simulation and experimental-design assumptions rather than mathematical axioms. The only mathematical ingredients are standard neural-network operations and the VAE ELBO; no physical law is asserted beyond the standard 5G NR signal model in Eq. (1). The unstated hyperparameters beta, alpha, and D_z are free parameters that are not reported.

free parameters (6)

K (number of experts) = 3
Set to 3 without sensitivity analysis; expert count directly determines capacity and is central to MoE performance.
top-k (activated experts) = 1
Only one expert is activated; no comparison to k>1 or soft mixing. This choice heavily affects the routing claim.
D_z (latent embedding dimension) = not reported
Dimensionality of the VAE routing representation; critical to routing accuracy but not given.
beta (VAE regularization weight, Eq. 9) = not reported
Controls the KL term in the ELBO; unspecified in experiments.
alpha (load-balancing weight, Eq. 12) = not reported
Controls expert-collapse mitigation; unspecified in experiments.
Training schedule hyperparameters = not reported
Learning rates (encoder fine-tuning vs. experts), batch size, epochs, channel sampling distribution not stated; required to reproduce the curves.

axioms (4)

domain assumption The 3GPP TR 38.901 channel models and Sionna link-level simulator are faithful proxies for real NextG propagation.
The only evidence is simulation; no over-the-air validation is offered (Sections II, IV-A).
domain assumption The 'unseen' test channels (CDL-A/E, UMi) are sufficiently distinct from training channels (CDL-B/C/D, UMa) for the zero-shot evaluation to be meaningful.
The generalization claim depends on this split; the paper provides no analysis of channel-statistic overlap (Section IV-A).
domain assumption The LS channel estimate at DMRS positions contains enough task-relevant information for the VAE router to select a good expert.
Routing uses only this coarse estimate; if the estimate is poor, routing decisions would be arbitrary and the method would fail (Section III-B, Eq. 2).
ad hoc to paper The two-stage training objective (Eq. 12) with load balancing avoids expert collapse in the reported runs.
The paper asserts two-stage training is needed but gives no diagnostic evidence of expert utilization or collapse (Section III-D).

invented entities (1)

Task-oriented latent channel embedding z (or mu/log-sigma^2) no independent evidence
purpose: Low-dimensional representation of channel conditions used to route received blocks to experts.
It is an internal learned variable with no measurement outside the paper; the ablation CARNet vs CARNet-Base is an internal comparison, not independent evidence.

pith-pipeline@v1.3.0-daily-deepseek · 7251 in / 15605 out tokens · 125752 ms · 2026-08-04T12:56:04.758897+00:00 · methodology

0 comments

read the original abstract

Neural receivers have been recognized as a promising paradigm for the next-generation (NextG) communications. However, due to the reliance on a static network optimized for specific channel conditions, their generalization capability across diverse scenarios remains a significant challenge. To address this issue, this paper proposes a novel channel-adaptive neural receiver network (CARNet) based on the mixture-of-experts (MoE) framework. The proposed architecture employs multiple expert networks together with an efficient routing mechanism to enable signal detection in various scenarios. The experts are constructed via stacked ResNet blocks and specialize in robust signal detection within specific channel conditions, while the routing mechanism incorporates a lightweight representation learning module, which projects the coarse channel estimate into a low-dimensional latent embedding. The learned embedding characterizes task-relevant channel conditions and provides efficient guidance for accurate expert selection. Link-level simulation experiments demonstrate that the proposed CARNet achieves superior performance across diverse channel conditions.

Figures

Figures reproduced from arXiv: 2608.02172 by Chao Jiang, Yongli Yan, Zhuo Xu.

**Figure 2.** Figure 2: Overall architecture and detailed sub-network structures of the proposed CARNet. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: BLER performance versus Eb/N0 under diverse channel conditions: (a) NLoS CDL-C, (b) LoS CDL-D ,and zero-shot generalization to (c) UMi. B. Simulation Results As illustrated in Fig. 3a and Fig. 3b, we first evaluate the block error rate (BLER) performance under CDL-C and CDLD, which are included in the training dataset. The former represents a NLoS scenario with rich multipath propagation, while the latter… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 1 linked inside Pith

[1]

Toward a 6G AI-native air interface,

J. Hoydis, F. A. Aoudia, A. Valcarce, and H. Viswanathan, “Toward a 6G AI-native air interface,”IEEE Commun. Mag., vol. 59, no. 5, pp. 76–81, May 2021

2021
[2]

A tale of two mobile generations: 5G-Advanced and 6G in 3GPP Release 20,

X. Lin, “A tale of two mobile generations: 5G-Advanced and 6G in 3GPP Release 20,”IEEE Commun. Stand. Mag., pp. 1–9, Oct. 2025

2025
[3]

DeepRx: Fully convolu- tional deep learning receiver,

M. Honkala, D. Korpi, and J. M. Huttunen, “DeepRx: Fully convolu- tional deep learning receiver,”IEEE Trans. Wireless Commun., vol. 20, no. 6, pp. 3925–3940, Jun. 2021

2021
[4]

DARNet: Deep attention receiver network for 5G,

Z. Wu et al, “DARNet: Deep attention receiver network for 5G,”IEEE Commun. Lett., vol. 30, pp. 572–576, 2026

2026
[5]

CRNN-ResNet: Combined CRNN and ResNet networks for OFDM receivers,

R. Mei, Z. Wang, and X. Chen, “CRNN-ResNet: Combined CRNN and ResNet networks for OFDM receivers,”IEEE Trans. Cogn. Commun. Netw., vol. 10, no. 4, pp. 1294–1306, Aug. 2024

2024
[6]

DRNet: Deep learning-based dual- branch receiver for OFDM with nonlinear distortions,

Q. Jin, X. Li, R. Guo, and F. Liu, “DRNet: Deep learning-based dual- branch receiver for OFDM with nonlinear distortions,”IEEE Commun. Lett., vol. 30, pp. 357–361, 2026

2026
[7]

Adaptive end-to-end transceiver design for NextG pilot-free and CP-free wireless systems,

J. Cheng, W. Chen, and B. Ai, “Adaptive end-to-end transceiver design for NextG pilot-free and CP-free wireless systems,”IEEE J. Sel. Areas Commun., vol. 44, pp. 3055–3069, Feb. 2026

2026
[8]

Learning-aided iterative receiver for superimposed pilots: Design and experimental evaluation,

X. Li, X. Zhou, Y . Cao, J. Zhang, C.-K. Wen, X. Li, and S. Jin, “Learning-aided iterative receiver for superimposed pilots: Design and experimental evaluation,”IEEE Trans. Wireless Commun., vol. 25, pp. 13 864–13 880, Mar. 2026

2026
[9]

Towards green 6G networks: An energy- efficient neural receiver for MIMO systems,

C. Jiang, Z. Xu, and Y . Yan, “Towards green 6G networks: An energy- efficient neural receiver for MIMO systems,”IEEE Trans. Veh. Technol., early access, Jul. 2026, doi: 10.1109/TVT.2026.3717959

arXiv 2026
[10]

GNN-assisted BiG-AMP: Joint channel estimation and data detection for massive MIMO receiver,

Z. Liu, N. Wu, D. He, W. Yuan, Y . Li, and T. Q. S. Quek, “GNN-assisted BiG-AMP: Joint channel estimation and data detection for massive MIMO receiver,”IEEE Trans. Wireless Commun., vol. 24, no. 6, pp. 4631–4646, Jun. 2025

2025
[11]

A hybrid inference architecture incorporating neural network with belief propagation for AI receivers,

Y . Yang, Z. Zhang, Z. Chen, Z. Yang, L. Liu, C. Huang, and M. Debbah, “A hybrid inference architecture incorporating neural network with belief propagation for AI receivers,”IEEE Trans. Wireless Commun., vol. 24, no. 8, pp. 6365–6381, Aug. 2025

2025
[12]

Artificial intelligence for wireless physical-layer technolo- gies (AI4PHY): A comprehensive survey,

N. Ye et al., “Artificial intelligence for wireless physical-layer technolo- gies (AI4PHY): A comprehensive survey,”IEEE Trans. Cogn. Commun. Netw., vol. 10, no. 3, pp. 729–755, Jun. 2024

2024
[13]

Sionna: An open-source library for next-generation physical layer research,

J. Hoydis et al., “Sionna: An open-source library for next-generation physical layer research,”arXiv:2203.11854, Mar. 2023

Pith/arXiv arXiv 2023

[1] [1]

Toward a 6G AI-native air interface,

J. Hoydis, F. A. Aoudia, A. Valcarce, and H. Viswanathan, “Toward a 6G AI-native air interface,”IEEE Commun. Mag., vol. 59, no. 5, pp. 76–81, May 2021

2021

[2] [2]

A tale of two mobile generations: 5G-Advanced and 6G in 3GPP Release 20,

X. Lin, “A tale of two mobile generations: 5G-Advanced and 6G in 3GPP Release 20,”IEEE Commun. Stand. Mag., pp. 1–9, Oct. 2025

2025

[3] [3]

DeepRx: Fully convolu- tional deep learning receiver,

M. Honkala, D. Korpi, and J. M. Huttunen, “DeepRx: Fully convolu- tional deep learning receiver,”IEEE Trans. Wireless Commun., vol. 20, no. 6, pp. 3925–3940, Jun. 2021

2021

[4] [4]

DARNet: Deep attention receiver network for 5G,

Z. Wu et al, “DARNet: Deep attention receiver network for 5G,”IEEE Commun. Lett., vol. 30, pp. 572–576, 2026

2026

[5] [5]

CRNN-ResNet: Combined CRNN and ResNet networks for OFDM receivers,

R. Mei, Z. Wang, and X. Chen, “CRNN-ResNet: Combined CRNN and ResNet networks for OFDM receivers,”IEEE Trans. Cogn. Commun. Netw., vol. 10, no. 4, pp. 1294–1306, Aug. 2024

2024

[6] [6]

DRNet: Deep learning-based dual- branch receiver for OFDM with nonlinear distortions,

Q. Jin, X. Li, R. Guo, and F. Liu, “DRNet: Deep learning-based dual- branch receiver for OFDM with nonlinear distortions,”IEEE Commun. Lett., vol. 30, pp. 357–361, 2026

2026

[7] [7]

Adaptive end-to-end transceiver design for NextG pilot-free and CP-free wireless systems,

J. Cheng, W. Chen, and B. Ai, “Adaptive end-to-end transceiver design for NextG pilot-free and CP-free wireless systems,”IEEE J. Sel. Areas Commun., vol. 44, pp. 3055–3069, Feb. 2026

2026

[8] [8]

Learning-aided iterative receiver for superimposed pilots: Design and experimental evaluation,

X. Li, X. Zhou, Y . Cao, J. Zhang, C.-K. Wen, X. Li, and S. Jin, “Learning-aided iterative receiver for superimposed pilots: Design and experimental evaluation,”IEEE Trans. Wireless Commun., vol. 25, pp. 13 864–13 880, Mar. 2026

2026

[9] [9]

Towards green 6G networks: An energy- efficient neural receiver for MIMO systems,

C. Jiang, Z. Xu, and Y . Yan, “Towards green 6G networks: An energy- efficient neural receiver for MIMO systems,”IEEE Trans. Veh. Technol., early access, Jul. 2026, doi: 10.1109/TVT.2026.3717959

arXiv 2026

[10] [10]

GNN-assisted BiG-AMP: Joint channel estimation and data detection for massive MIMO receiver,

Z. Liu, N. Wu, D. He, W. Yuan, Y . Li, and T. Q. S. Quek, “GNN-assisted BiG-AMP: Joint channel estimation and data detection for massive MIMO receiver,”IEEE Trans. Wireless Commun., vol. 24, no. 6, pp. 4631–4646, Jun. 2025

2025

[11] [11]

A hybrid inference architecture incorporating neural network with belief propagation for AI receivers,

Y . Yang, Z. Zhang, Z. Chen, Z. Yang, L. Liu, C. Huang, and M. Debbah, “A hybrid inference architecture incorporating neural network with belief propagation for AI receivers,”IEEE Trans. Wireless Commun., vol. 24, no. 8, pp. 6365–6381, Aug. 2025

2025

[12] [12]

Artificial intelligence for wireless physical-layer technolo- gies (AI4PHY): A comprehensive survey,

N. Ye et al., “Artificial intelligence for wireless physical-layer technolo- gies (AI4PHY): A comprehensive survey,”IEEE Trans. Cogn. Commun. Netw., vol. 10, no. 3, pp. 729–755, Jun. 2024

2024

[13] [13]

Sionna: An open-source library for next-generation physical layer research,

J. Hoydis et al., “Sionna: An open-source library for next-generation physical layer research,”arXiv:2203.11854, Mar. 2023

Pith/arXiv arXiv 2023