Pith. sign in

REVIEW 5 major objections 6 minor 30 references

Learning-Based Hybrid Neural Receiver for 6G-V2X Communications

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A hybrid neural receiver that combines Transformer encoders with a Graph Neural Network decoder replaces the entire physical-layer receive chain in a 6G vehicle-to-network link, outperforming the previous best neural receiver by about 0.5…

desk verdict A plausible full-chain neural receiver, but the 0.5 dB claim lacks error bars and a clean GNN-versus-LDPC ablation. read the letter →

arxiv 2506.21983 v1 pith:PXXUDIJS submitted 2025-06-27 eess.SP

Osama Saleem , Mohammed Alfaqawi , Pierre Merdrignac , Abdelaziz Bensrhair , Soheyb Ribouh This is my paper
classification eess.SP
keywords hybridneuralreceiverTransformerencoderGraphNetwork6GV2XOFDMLDPCdecodingdomaingeneralization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Existing neural receivers replace only parts of the wireless receive chain — channel estimation, equalization, or decoding — while leaving the rest to classical signal processing. This paper claims to be the first to replace the entire physical-layer receiver with a single learned model: a hybrid neural receiver (H-NR) that takes the raw OFDM resource grid and outputs reconstructed information bits. The model combines five Transformer encoder blocks that produce soft bit estimates with a Graph Neural Network that performs belief-propagation-style decoding over the code's parity-check structure. The paper reports that this one model, trained only on a 28 GHz urban-macro (UMa) channel at speeds of 60–120 km/h, maintains its advantage over the previous best neural receiver (about 0.5 dB) when tested at 0–60 km/h, at 5.9 GHz, on a CDL channel, and across image, audio, GPS, LiDAR, and radar payloads. If true, it suggests a single data-driven receiver could serve multiple 6G V2X scenarios without redesign or retraining.

What carries the argument

The model is a hybrid neural receiver: a front-end of five Transformer encoder blocks with multi-head self-attention and 128-dimensional embeddings that maps the noisy OFDM resource grid (plus noise power) to log-likelihood ratios, followed by a bipartite Graph Neural Network that treats those LLRs as variable nodes and parity-check constraints as check nodes, passing messages along the edges of the LDPC parity-check matrix for 12 iterations. A three-stage training schedule is the mechanism: transformer trained with an LDPC teacher, GNN trained on frozen transformer LLRs, then joint fine-tuning with BCE loss.

What would settle it

Decode the transformer's soft outputs from the CDL and 5.9 GHz test sets with a standard LDPC decoder and compare BER against the GNN decoder; if the standard decoder matches or beats the GNN on identical soft inputs, the reported 0.5 dB gain comes from the front-end, not the learned decoder.

Watch

Extended reading notes

Core claim

The paper's central claim is that a full receiver substitution (FRS) is possible: a hybrid neural receiver, H-NR, can replace OFDM resource-grid demapping, channel estimation, signal equalization, demodulation, and channel decoding all at once, something the authors argue no previous neural receiver has done (Section I and Table I). H-NR is trained in three stages: first, the transformer front-end is trained to output LLRs on 12 million samples with a classical LDPC decoder as teacher; then the transformer is frozen and a bipartite GNN is trained on those LLRs using the parity-check matrix, with 12 message-passing iterations; finally, the whole receiver is fine-tuned jointly with a binary cross-entropy loss. Evaluated at 28 GHz on UMa, the model converges to minimal BER at 6.4 dB for speeds of 0–60 km/h and 6.5 dB for 60–120 km/h, improving on TransRx by about 0.4–0.5 dB and on DeepRx by 1.5–2 dB; at 5.9 GHz the gain over TransRx is 0.5 dB, and on a CDL channel the model converges at 5.60 dB versus 5.75 dB for TransRx. On multimodal data at 28 GHz, the model gives lower MSE and higher PSNR than TransRx and DeepRx at every tested SNR. The intended message is that an end-to-end learned receiver can match or beat classical baselines while staying robust to out-of-distribution channel conditions.

Load-bearing premise

The load-bearing premise is that the transformer's soft bit estimates, learned only on a 28 GHz urban channel at 60–120 km/h, remain statistically meaningful at lower speeds, at 5.9 GHz, and over the CDL channel, so the trained decoder still corrects them without retraining.

Editorial extensions

If this is right

  • If the central claim holds, a single learned receiver can replace the entire classical PHY receive chain, removing the need for separate demapping, estimation, equalization, and decoding modules in V2X base stations.
  • The reported out-of-distribution robustness (0–60 km/h, 5.9 GHz, CDL channel) suggests the model can be trained once in one environment and deployed across a range of speeds and frequencies without retraining.
  • The 0.4–0.5 dB improvement over TransRx on both in-distribution and OOD scenarios implies that jointly fine-tuning the GNN decoder with the transformer front-end yields a gain beyond what either module achieves alone.
  • On multimodal AV data, the receiver reconstructs images, audio, GPS, LiDAR, and radar from the same bit-level pipeline, so one receiver design could serve diverse sensor payloads in cooperative driving.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the GNN decoder's advantage depends on the LLR statistics produced by the UMa-trained transformer, then the same decoder may not transfer to a different front-end or a different code without retraining; a direct ablation against a classical LDPC decoder on identical LLRs would isolate where the 0.5 dB comes from.
  • The reported convergence SNRs are read off single BER curves; a distribution of runs or confidence intervals would be needed to know whether the 0.5 dB gap is significant as opposed to run-to-run variation.
  • The paper's first-FRS claim is relative to the cited baselines; since the GNN still uses the known parity-check matrix and the system assumes LDPC coding, the learned component augments rather than discovers the code structure.
  • A natural testable extension is to feed the same received resource grids to a receiver that keeps the classical LDPC decoder but uses the learned front-end, and vice versa, to separate the gains of the two halves.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes H-NR, a hybrid neural receiver that combines Transformer encoder blocks with a GNN-based channel decoder to replace the entire physical-layer receiver chain in a 6G V2X uplink. The model is trained on simulated UMa channels at 28 GHz and 60–120 km/h, then evaluated on out-of-distribution conditions (0–60 km/h, 5.9 GHz, CDL channel) and on multimodal data (image, audio, GPS, LiDAR, radar). The central empirical claim is that H-NR outperforms the previous best neural receiver TransRx by approximately 0.5 dB in BER, with consistent gains across the OOD scenarios.

Significance. If the results hold, the paper demonstrates a single jointly trained model that maps raw received resource grids directly to information bits, replacing demapping, channel estimation, equalization, demodulation, and channel decoding, with modest but systematic BER gains over a neural front-end plus LDPC decoding. The evaluation suite is broad, covering multiple OOD dimensions and several data modalities, and the use of the Sionna simulator provides reproducibility. The main significance is therefore in the engineering demonstration of a full learned receiver chain that generalizes without retraining, rather than in a fundamentally new architectural idea, since the Transformer front-end is the authors' own prior TransRx design.

major comments (5)
  1. [Section VI-A, Figs 4-7] No error bars, confidence intervals, or multiple independent runs are reported for any BER curve. The claimed gains are small: Fig. 7 shows 5.60 dB versus 5.75 dB convergence SNR for H-NR versus TransRx on the CDL channel, a 0.15 dB margin that is within typical run-to-run variability for simulated BER at these SNRs. Without statistical uncertainty quantification, the headline 'approximately 0.5 dB' improvement is not established.
  2. [Section IV and Section VI-A] The H-NR versus TransRx comparison conflates two changes: replacing the LDPC decoder with the GNN and fine-tuning the Transformer front-end jointly with the GNN (Eq. 14). The paper never evaluates (i) TransRx's LLRs decoded by the GNN, (ii) H-NR's LLRs decoded by the standard LDPC decoder, or (iii) the frozen Transformer with the GNN but without joint fine-tuning. Without these ablations, the reported gains cannot be attributed to the learned GNN decoder, and the adaptability claim may be driven by front-end fine-tuning alone.
  3. [Section VI-A, Figs 4-7] The convergence SNR values are read from the figures without a stated criterion for 'minimal error rate' or a reported number of channel realizations or transmitted bits per point. This makes the 0.15–0.5 dB differences difficult to verify independently and prevents the reader from assessing whether the curves are in the noise floor region.
  4. [Section I and Table I] The claim of being the first 'full receiver substitution' rests on the authors' own classification in Table I, which marks all prior neural receivers as lacking channel decoding. In particular, Cammerer et al. [20] also employ a GNN in a neural receiver for 5G NR multi-user MIMO, and the text does not explain why that work does not count as a full receiver substitution. The novelty framing should be softened to an extension of the authors' own TransRx with a learned decoder and joint fine-tuning.
  5. [Section V] The training and evaluation protocol for the baseline receivers (DeepRx and TransRx) is not specified: it is not stated whether these models were retrained on the same Sionna data, with the same pilot pattern, subcarrier spacing, and code rate, or whether published weights were used. If the baselines were not trained on identical data, the comparison may be unfair, which is critical given the small margins claimed.
minor comments (6)
  1. [Section IV, Eq. (15)] The expression for the gradient of BER with respect to θ is a formal derivative of the error probability, but the actual optimization in Eq. (14) uses the BCE loss; the text does not explain the relationship between these two objectives or why the gradient expression is relevant to the training procedure.
  2. [Section VI-B] The multimodal evaluation transmits bits from each modality over the same simulated PHY and measures PSNR/MSE; it essentially replots the BER gains in different units. The authors should state explicitly that no modality-specific receiver processing is used, so that readers do not infer a claim about cross-modal generalization at the semantic level.
  3. [Section III and Table II] The baseline decoder in Eq. (8) uses belief propagation with 10 iterations, while the GNN decoder in Table II uses 12 message-passing iterations; the difference is not discussed and may affect the comparison's fairness.
  4. [Figures 4-7 captions] The captions are terse and do not indicate which conditions are out-of-distribution relative to training; a small summary table marking in-distribution vs. OOD (speed, carrier frequency, channel model) would improve clarity.
  5. [Section VI-B5] The heading 'Radar Data Trasmission' contains a typo; it should be 'Transmission'.
  6. [References] Reference [28] appears in the bibliography but is not cited in the body of the paper; either cite it or remove it.

Circularity Check

1 steps flagged · score 2.0 of 10

Self-referential novelty claim rests on the authors' own Table I classification, but the quantitative BER/PSNR/MSE evaluations are held-out and not circular.

  1. self citation load bearing [Section II (Table I) and Section IV]
    "While previous work has focused on neural receiver models that outperform conventional stochastic receiver architectures, their optimization is limited to multiple modules substitution (MMS) as shown in Table I. However, in this work, we propose a full receiver substitution (FRS) architecture, called H-NR... To the best of our knowledge, this is the first work that involves the proposition of a full receiver substitution."

    The central novelty premise, that no prior receiver substitutes the entire physical layer, is established by the authors' own Table I, which classifies their own prior works [21] (TransRx) and [22] as MMS (demapping and decoding marked ✗). Section IV then states 'The architecture of transformer encoder block is same as presented in [21]', so the 'FRS' contribution is partly defined by the authors' self-classification of their own prior architecture rather than by an externally established gap. This self-citation is load-bearing for the 'first work' claim, but it is minor because the quantitative comparisons are empirical and independent of this taxonomy.

full rationale

The derivation chain for the quantitative claims is self-contained. H-NR is trained on Sionna-generated UMa 28 GHz data at 60-120 km/h and then evaluated on held-out conditions (0-60 km/h, 5.9 GHz, CDL) without retraining; these are genuine predictions benchmarked against DeepRx, the baseline receiver, and the authors' prior TransRx. The GNN decoder is a structured message-passing network whose graph is defined by the LDPC parity-check matrix, and the final joint loss (Eq. 14) is BCE against transmitted bits, so the outputs are not fitted to the test conditions. The only self-referential element is the novelty claim: Section II/Table I classifies the authors' own [21] and [22] as MMS, and Section IV reuses the [21] transformer block unchanged, so 'first full receiver substitution' rests in part on the authors' own taxonomy. This does not affect the empirical BER/PSNR/MSE comparisons, which remain independent evidence. The absence of an ablation separating the GNN decoder's contribution from joint fine-tuning is a correctness or attribution concern, not a circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on empirically trained weights plus four hand-chosen hyperparameter groups and four domain assumptions. The heaviest debt is the assumption that Sionna-simulated OOD conditions and the frozen-LLR training pipeline transfer to real V2X channels, which the paper admittedly does not test. No new physical entities are introduced.

free parameters (4)
  • GNN staged training schedule = [4M, 40M, 40M] samples; LRs [5e-4, 1e-4, 1e-5]
    Hand-chosen schedule in Table II with no sensitivity analysis; the 40M-sample stages dominate training and are tuned without stated criteria.
  • GNN message-passing iterations = 12
    MP iteration count in Table II controls decoder performance and complexity; no ablation justifies it.
  • Joint fine-tune budget = LR 1e-5, 1M samples
    The final joint fine-tuning (Section IV, Eq. 14) uses a hand-picked learning rate and data budget; results may depend on this choice.
  • Transformer depth and width = 5 blocks, 4 heads, 128 embedding
    Inherited unchanged from the authors' prior TransRx [21]; the paper states the encoder is 'same as presented in [21]', so these hyperparameters are not independently justified here.
assumptions (4)
  • domain assumption 3GPP UMa and CDL channel models in Sionna faithfully represent real V2X propagation.
    All evaluation runs inside the Sionna simulator (Section V); the conclusion defers real-world validation to a future USRP/OAI testbed, so simulation fidelity is load-bearing.
  • domain assumption The frozen transformer's LLR outputs carry enough information to train an effective GNN decoder.
    Section IV trains the GNN on LLRs from the frozen transformer, whose own training used a standard LDPC decoder as teacher; poorly calibrated LLRs would break the decoder stage.
  • ad hoc to paper The jointly fine-tuned model generalizes to OOD conditions without retraining.
    Section VI asserts consistent performance at 0-60 km/h, 5.9 GHz, and CDL channels using single BER curves with no error bars; this is the paper's key adaptability claim.
  • standard math Standard machine-learning building blocks (attention, message passing, BCE training) behave as assumed.
    Equations (9)-(14) rely on standard transformer and MPNN definitions from [21] and [25]; no convergence or calibration properties are proved.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning-Based Hybrid Neural Receiver for 6G-V2X Communications." pith.science (2026). https://pith.science/paper/PXXUDIJS

@misc{pith2026250621983,
  author       = {Pith},
  title        = {Pith review of: Learning-Based Hybrid Neural Receiver for 6G-V2X Communications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PXXUDIJS}},
  note         = {Machine review of arXiv:2506.21983}
}
read the original abstract

Neural receiver models are proposed to jointly optimize multiple functionalities of wireless receivers; however, a comprehensive receiver model that replaces the entire physical layer blocks has not yet been presented in the literature. In this work, we introduce a novel hybrid neural receiver (H-NR) built on Transformer encoder blocks and Graph Neural Network (GNN), as part of an end-to-end wireless communication framework. In our communication framework, we assume vehicle to network (V2N) uplink scenario where information is transmitted by vehicle and received at the base station (BS). Our proposed H-NR model replace OFDM resource grid demapping, channel estimation, signal equalization, demodulation, and channel decoding. To test the adaptability of our proposed model on unseen conditions, we evaluate its performance for various scenarios, including a vehicle speed of range [0-60] km/h, a carrier frequency of 5.9GHz, and a cluster delay line (CDL) channel model. Furthermore, we assess the performance of our proposed H-NR on multimodal data, such as images, audio, GPS, radar, and LiDAR, to examine its adaptability in real-world use cases. The simulation results clearly demonstrate that our proposed model outperforms the state-of-the-art neural receiver by approximately 0.5 dB in terms of reconstruction and error correction.

Figures

Figures reproduced from arXiv: 2506.21983 by the authors.

Figure 1
Figure 1. Vehicle-to-Everything (V2X) Communication in Urban Environment [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. PHY of Proposed H-NR model and state of the art methods based end to end wireless communication framework. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. H-NR Architecture built around the multi-head self-attention mechanism, which enables the model to capture long-range dependencies and complex signal patterns within the input data. The attention mechanism is mathematically defined as: Attention(Q, K, V ) = softmax  QK⊤ √ vk  V (9) where: • Q, K, and V represent the query, key, and value matrices derived from the input, • vk is the dimension of the key vectors use… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: BER w.r.t SNR for UMa Channel Model with vehicle speed range of [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: BER w.r.t SNR for UMa Channel Model with vehicle speed range of [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: BER w.r.t SNR for UMa Channel Model with vehicle speed range of [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: Visual comparison of image reconstruction and corresponding L1 heat map analysis across different receiver architectures. The scenario is conducted [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: PSNR w.r.t SNR for image transmission over UMa Channel Model [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: MSE w.r.t SNR for audio transmission over UMa Channel Model [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 12
Figure 12. Figure 12: MSE w.r.t SNR for radar transmission over UMa Channel Model [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 24 canonical work pages

  1. [20]

    Cammerer et al., ”A Neural Receiver for 5G NR Multi-User MIMO,” 2023 IEEE Globecom Workshops (GC Wkshps), Kuala Lumpur, Malaysia, 2023, pp

    S. Cammerer et al., ”A Neural Receiver for 5G NR Multi-User MIMO,” 2023 IEEE Globecom Workshops (GC Wkshps), Kuala Lumpur, Malaysia, 2023, pp. 329-334, doi: 10.1109/GCWkshps58843.2023.10464486

    arXiv 2023

  2. [1]

    ”Key focus areas and enabling technologies for 6G.” IEEE Communications Magazine 63.3 (2025): 84-91

    Brinton, Christopher G., et al. ”Key focus areas and enabling technologies for 6G.” IEEE Communications Magazine 63.3 (2025): 84-91

    work page 2025

  3. [2]

    Study on Artificial Intelligence (AI)/Machine Learn- ing (ML) for NR Air Interface RAN,

    3GPP Release 18, “Study on Artificial Intelligence (AI)/Machine Learn- ing (ML) for NR Air Interface RAN,” Meeting #112 , Athens, Greece, Tech. Rep., 27th February – 3rd March 2023

    work page 2023

  4. [3]

    The Case for Cellular V2X for Safety and Cooperative Driving,

    “The Case for Cellular V2X for Safety and Cooperative Driving,” 5G Automotive Association, Tech. Rep. 23- Nov-2016, 2016

    work page 2016

  5. [4]

    ”6G for vehicle-to-everything (V2X) commu- nications: Enabling technologies, challenges, and opportunities.” Proceed- ings of the IEEE 110.6 (2022): 712-734

    Noor-A-Rahim, Md, et al. ”6G for vehicle-to-everything (V2X) commu- nications: Enabling technologies, challenges, and opportunities.” Proceed- ings of the IEEE 110.6 (2022): 712-734

    work page 2022

  6. [5]

    O’Shea & J

    T. O’Shea & J. Hoydis, ”An Introduction to Deep Learning for the Physical Layer,” in IEEE Transactions on Cognitive Communica- tions and Networking , vol. 3, no. 4, pp. 563-575, Dec. 2017, doi: 10.1109/TCCN.2017.2758370

    arXiv 2017

  7. [6]

    Learning the MMSE channel estimator,

    D. Neumann, T. Wiese, & W. Utschick, “Learning the MMSE channel estimator,” IEEE Transactions on Signal Processing , vol. 66, no. 11, pp. 2905–2917, June 2018

    work page 2018

  8. [7]

    Deep learning-based channel estima- tion for beamspace mmWave massive MIMO systems,

    H. He, C. Wen, S. Jin, & G. Li, “Deep learning-based channel estima- tion for beamspace mmWave massive MIMO systems,” IEEE Wireless Communications Letters, vol. 7, no. 5, 2018

    work page 2018

Show all 30 references
  1. [8]

    Complex CNN-based equalization for communication signal,

    Z. Chang, Y . Wang, H. Li, & Z. Wang, “Complex CNN-based equalization for communication signal,” in 2019 IEEE 4th International Conference on Signal and Image Processing (ICSIP) , July 2019, pp. 513–517

    2019

  2. [9]

    Shental & J

    O. Shental & J. Hoydis, ””Machine LLRning”: Learning to Softly Demodulate,” 2019 IEEE Globecom Workshops (GC Wkshps) , Waikoloa, HI, USA, 2019, pp. 1-7, doi: 10.1109/GCWkshps45667.2019.9024433

    2019

  3. [10]

    Michon, F

    A. Michon, F. Ait Aoudia, & K.P. Srinath (2022, May). Convolutional Self-Attention-Based Multi-User MIMO Demapper. In ICC 2022-IEEE International Conference on Communications (pp. 2621-2626). IEEE

    2022

  4. [11]

    Cammerer, J

    S. Cammerer, J. Hoydis, F. A ¨ıt Aoudia, & A. Keller (2022, December). Graph neural networks for channel decoding. In 2022 IEEE Globecom Workshops (GC Wkshps) (pp. 486-491). IEEE

    2022

  5. [12]

    ComNet: Combination of deep learning and expert knowledge in OFDM receivers,

    X. Gao, S. Jin, C. Wen, & G. Y . Li, “ComNet: Combination of deep learning and expert knowledge in OFDM receivers,” IEEE Communica- tions Letters, vol. 22, no. 12, pp. 2627–2630, Dec 2018

    2018

  6. [13]

    Model-driven deep learning for physical layer communications,

    H. He, S. Jin, C. Wen, F. Gao, G. Y . Li, & Z. Xu, “Model-driven deep learning for physical layer communications,” IEEE Wireless Communi- cations, vol. 26, no. 5, pp. 77–83, 2019

    2019

  7. [14]

    Learning to detect,

    N. Samuel, T. Diskin, & A. Wiesel, “Learning to detect,” IEEE Trans- actions on Signal Processing, vol. 67, no. 10, pp. 2554–2564, May 2019

    2019

  8. [15]

    Power of deep learning for channel es- timation and signal detection in OFDM systems,

    H. Ye, G. Y . Li, & B.-H. Juang, “Power of deep learning for channel es- timation and signal detection in OFDM systems,” IEEE Communications Letters, vol. 7, no. 1, pp. 114–117, Feb. 2018

    2018

  9. [16]

    Z. Zhao, M. C. Vuran, F. Guo & S. D. Scott, ”Deep-Waveform: A Learned OFDM Receiver Based on Deep Complex-Valued Convolutional Networks,” in IEEE Journal on Selected Areas in Communications , vol. 39, no. 8, pp. 2407-2420, Aug. 2021, doi: 10.1109/JSAC.2021.3087241

    2021

  10. [17]

    Honkala, D

    M. Honkala, D. Korpi & J. M. J. Huttunen, ”DeepRx: Fully Con- volutional Deep Learning Receiver,” in IEEE Transactions on Wire- less Communications , vol. 20, no. 6, pp. 3925-3940, June 2021, doi: 10.1109/TWC.2021.3054520

    2021

  11. [18]

    J. Pihlajasalo et al., ”HybridDeepRx: Deep Learning Receiver for High- EVM Signals,” 2021 IEEE 32nd Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC) , Helsinki, Finland, 2021, pp. 622-627, doi: 10.1109/PIMRC50174.2021.9569393

    2021

  12. [19]

    Akrout, A

    M. Akrout, A. Feriani, F. Bellili, A. Mezghani & E. Hossain, ”Domain Generalization in Machine Learning Models for Wireless Communica- tions: Concepts, State of-the-Art, and Open Issues,” in IEEE Communica- tions Surveys and Tutorials , vol. 25, no. 4, pp. 3014-3037, Fourthqua...

    2023

  13. [21]

    Saleem, S

    O. Saleem, S. Ribouh, M. Alfaqawi, A. Bensrhair & P. Mer- drignac, ”TransRx-6G-V2X : Transformer Encoder-Based Deep Neu- ral Receiver For Next Generation of Cellular Vehicular Communica- tions,” 2024 IEEE 100th Vehicular Technology Conference (VTC2024- Fall), Washington, DC, U...

    2024

  14. [22]

    Saleem, M

    O. Saleem, M. Alfaqawi, P. Merdrignac, A. Bensrhair & S. Ribouh (2025). Deep Multi-modal Neural Receiver for 6G Vehicular Communi- cation. arXiv preprint arXiv:2501.13464

    2025 arXiv

  15. [23]

    C ¸ evik, E

    H. C ¸ evik, E. Karakoca, I. H ¨okelek, & G ¨orc ¸in, A. (2025). Design and Evaluation of Neural Network-Based Receiver Architectures for Reliable Communication. arXiv preprint arXiv:2503.20500

    2025 arXiv

  16. [24]

    https://www.atis.org/wp-content/uploads/3gpp- documents/Rel15/ATIS.3GPP.37.885.V1530.pdf

  17. [25]

    Neural enhanced belief propagation on factor graphs,

    V . G. Satorras & M. Welling, “Neural enhanced belief propagation on factor graphs,” in International Conference on Artificial Intelligence and Statistics, pp. 685-693. PMLR, 2021

    2021

  18. [26]

    Sionna: An Open-Source Library for Next-Generation Physical Layer Research,

    J. Hoydis, S. Cammerer, F. Ait Aoudia, A. Vem, N. Binder, G. Marcus, and A. Keller, “Sionna: An Open-Source Library for Next-Generation Physical Layer Research,” arxiv preprint arXiv:2203.11854, 2022

    2022 arXiv

  19. [27]

    Alkhateeb, G

    A. Alkhateeb, G. Charan, T. Osman, A. Hredzak, J. Morais, U. Demirhan and N. Srinivas (2023). DeepSense 6G: A large-scale real-world multi- modal sensing and communication dataset. IEEE Communications Mag- azine, 61(9), 122-128

    2023

  20. [28]

    Y . Lu, Y . Hu, Y . Zhong, D. Wang, Y . Wang, & S. Chen (2024). An extensible framework for open heterogeneous collaborative perception. arXiv preprint arXiv:2401.13964

    2024 arXiv

  21. [29]

    ”nuscenes: A multimodal dataset for autonomous driving.” Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Holger, Caesar, et al. ”nuscenes: A multimodal dataset for autonomous driving.” Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020

    2020

  22. [30]

    Technical Specification Group Services and System Aspects; Release description; Release 15,

    3GPP, “Technical Specification Group Services and System Aspects; Release description; Release 15,” 3GPP TR 21.915 V15.0.0, Sep. 2019. [Online]. Available: https://www.3gpp.org/DynaReport/21915.htm

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.