Pith. sign in

REVIEW 3 major objections 4 minor 28 references

DL-Based Beam Management for mmWave Vehicular Networks Exploring Temporal Correlation

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read An LSTM that reads a four-step history of beam powers can keep a millimeter-wave vehicle link aligned while replacing up to three-quarters of physical beam measurements with its own predictions.

desk verdict Useful empirical study of RNN-based beam tracking on NLOS-rich simulated datasets, but the advertised position-aware pre-selection is absent from the experiments and the perfect-knowledge input assumption limits the strength of the accuracy claims. read the letter →

arxiv 2511.02260 v2 pith:PIBDMQXB submitted 2025-11-04 eess.SP

classification eess.SP
keywords beamtrackingmmWavedeeplearningLSTMvehicle-to-infrastructurenon-line-of-sightautoregressiveinferencemeasurementoverheadreduction
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

This paper establishes that a recurrent neural network, fed only a short sliding window of per-beam power values and beam indices, can keep a millimeter-wave vehicle-to-infrastructure link aligned with high top-K accuracy even when half the scenes are non-line-of-sight. The authors propose two model variants — DeepBT-C, which classifies the next optimal beam, and DeepBT-R, which regresses the received power of every beam — and show that the regression variant can substitute its own predictions for real measurements in autoregressive fashion. On ray-traced urban datasets, this substitution cuts measurement overhead by 50% to 75% with only a small accuracy penalty, and the classification variant reaches near-99% top-10 accuracy in the most obstructed scenario. The practical stake is that 6G beam management could run with far fewer physical measurements, using only data already available in standard architectures.

What carries the argument

The key mechanism is an LSTM-based recurrent network that maps a compact input vector (about 1.25 KB) of recent RSRP values per beam to either a beam-index distribution (DeepBT-C) or per-beam RSRP estimates (DeepBT-R). A sliding window of four time steps suffices. The authors also introduce Mean Absolute First Difference (MAFD), a circular-distance metric that quantifies how abruptly the optimal beam index changes, and use it to characterize datasets as smooth (M10%) or erratic (R10%, R50%).

What would settle it

Run the same DeepBT models in closed loop: at each step, feed back the predicted beam (or the beam actually used) instead of the oracle-best beam, and measure top-1/top-5 accuracy and throughput ratio. If the accuracy drops sharply relative to the oracle-fed evaluation — especially under 50% NLOS and 3:1 substitution — the central claim of viable low-overhead tracking would fail.

Watch

Extended reading notes

Core claim

The paper's central claim is that temporal correlation in beam-power sequences is sufficient for robust beam tracking: an LSTM that sees the last few RSRP values per beam can predict the optimal beam index with over 90% top-5 accuracy even under 50% non-line-of-sight conditions, beating heavier LIDAR-based and GRU baselines. A second finding is that the regression variant's output can be fed back as input — 'autoregressive inference' — so that the system only measures every second, third, or fourth time step. This preserves most of the accuracy while cutting measurement overhead by approximately 50%, 66.7%, and 75% respectively. The classification variant is more accurate, but only the regre

Load-bearing premise

The results assume that the model is always told the true optimal beam from the previous time step; in a real system the current beam may be suboptimal, so the reported accuracy and autoregressive stability are likely optimistic.

Editorial extensions

If this is right

  • Autoregressive inference at a 3:1 prediction-to-measurement ratio reduces sensing overhead by roughly 75% while keeping top-K accuracy acceptable, especially in line-of-sight-dominant scenarios.
  • The classification variant sustains over 95% top-5 accuracy in a dataset with 50% non-line-of-sight samples, suggesting that sensor-free tracking is viable in obstructed urban settings.
  • Substituting predictions for measurements shifts the accuracy/overhead trade-off: the regression model loses only a few points of top-K accuracy but cuts measurements by half to three-quarters.
  • Because the input is two orders of magnitude smaller than vision-based alternatives, the approach is better suited to bandwidth-constrained uplink reporting and real-time edge deployment.

Reading between the lines

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

  • The 3:1 ratio is a tested configuration, not a ceiling; in slowly varying LOS environments even longer prediction runs may hold accuracy, which this paper does not explore.
  • Since only beam indices and RSRP are used, the core model should transfer across deployments if retrained per site; cross-site generalization is not tested and could be a worthwhile follow-up.
  • A natural test is closed-loop evaluation where the model's own previous output, not the oracle-best beam, feeds back; the paper's limitation note implies this as the decisive next experiment.
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

3 major / 4 minor

Summary. The paper proposes two LSTM-based beam tracking models for mmWave V2I systems: DeepBT-C (beam-index classification) and DeepBT-R (RSRP regression). The models consume short time-series windows of beam measurements and are evaluated on three public ray-tracing datasets with varying LOS/NLOS proportions. The authors also propose an autoregressive measurement-replacement strategy that reduces the number of real beam measurements, reporting 50%, 66.7%, and 75% overhead reductions under 1:1, 2:1, and 3:1 prediction-to-measurement ratios. A new dataset-dynamics metric, MAFD, is introduced. The paper claims in the abstract that the framework combines a position-aware beam pre-selection strategy with sequential prediction, but no experiment in Section V implements that pre-selection; Section VI states that it is preliminary and left for future work.

Significance. If the reported results hold, the paper is a useful empirical contribution: it uses publicly available datasets, provides a clear comparison against three baselines, introduces a meaningful overhead-reduction mechanism, and explicitly tests NLOS-heavy scenarios, which are often missing in beam-tracking evaluations. The datasets and MAFD characterization are reusable assets. However, the advertised position-aware pre-selection is not evaluated, and the perfect-knowledge assumption in the baseline comparison needs clarification before the central claims can be fully accepted. With revision, the sequential-prediction and measurement-replacement findings could support a solid journal paper.

major comments (3)
  1. [Abstract, Section I, Section VI] The abstract and introduction present 'a position-aware beam pre-selection strategy' as part of the proposed framework, but Section V contains no experiment involving such a stage. Section VI explicitly says the filtering strategy is preliminary, was 'omitted for brevity', and remains future work. The reported Top-K accuracy and measurement-overhead reductions therefore evaluate only the sequential prediction component, not the complete advertised framework. This is a load-bearing mismatch: readers will interpret the accuracy/overhead results as validating the full system. Please either include the pre-selection stage in the experiments or revise the abstract and contributions to state that the pre-selection is not part of the current evaluation.
  2. [Section II-C, Section IV-A, Section V-D] The paper acknowledges reproducing the Jiang et al. baseline 'including the one of perfect knowledge of previously selected optimal beams.' It is not stated whether the same assumption applies to DeepBT-C and DeepBT-R, particularly in the autoregressive inference experiments of Section V-D. If the historical beam indices or 'previously selected optimal beams' used as model inputs are ground-truth labels rather than the model's own previous predictions or the actually deployed beam, the reported accuracy and the 50-75% overhead reduction are optimistic upper bounds. Please specify exactly what feedback signal is used during autoregressive testing, and, if ground-truth optimal beams are used, quantify the sensitivity to incorrect or suboptimal beam inputs.
  3. [Section IV-C, Section V] The baseline models are reimplemented/adapted to the authors' datasets, but the manuscript does not report training hyperparameters, number of random seeds, or confidence intervals. All figures show point estimates only. Since the paper's claims are comparative ('outperforms all baselines'), the absence of variance information and any reported baseline tuning makes it difficult to judge whether the observed margins are robust or stem from specific implementation choices. Please add multi-seed error bars or otherwise justify that all models were trained under comparable, converged settings.
minor comments (4)
  1. [Section III-B, Eq. (5)] In the Throughput Ratio definition, the denominator should be indexed by the test example, i.e., y_{\hat{i}(i)} rather than y_{\hat{i}}. As written, the sum over N of a single global best-beam value is not the intended throughput ratio.
  2. [Section IV-B, Ref. [26]] The 'autoregressive inference' reference [26] is an econometrics paper on structural VAR identification; a time-series forecasting or signal-processing reference would be more appropriate for this use of autoregressive prediction.
  3. [Figure 11] The caption says 'First row: Best Index Classification; Second row: RSRP Regression. Top to bottom: (a, b) R50%, (c, d) M10%, (e, f) R10%.' The row/column organization is confusing. Label each subplot with its dataset and model type directly in the figure.
  4. [Various] Minor typos: 'around to 66%' in the introduction should be 'around 66%'; 'Full Measur.' in Figure 11 should be 'Full Measurement'; 'miss-election' in Section V-E should be 'mis-selection'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the self-admitted omission of the position-aware pre-selection step is a scope limitation, not a circularity.

full rationale

The paper's contribution is empirical: DeepBT-C and DeepBT-R are trained and evaluated on three public ray-tracing datasets, with baselines that include a same-group CNN (Suzuki et al.) and external LSTM/GRU methods. No reported result is a fitted parameter renamed as a prediction; Top-K accuracy, throughput ratio, and measurement-overhead reductions are measured outcomes. The autoregressive replacement strategy is evaluated by feeding back model predictions, not deduced from the training objective. The overhead-reduction percentages follow arithmetically from the defined 1:1, 2:1, and 3:1 replacement ratios via Eq. (8), which is a metric definition rather than a derived claim. Self-citations appear for the datasets and the Suzuki et al. baseline, but those resources are public and externally available; they are not used to justify the central result. The only notable gap is a scope mismatch: the abstract advertises a position-aware beam pre-selection strategy, while Section VI explicitly states that the pre-selection experiments were omitted and remain future work ('Preliminary experiments (now omitted for brevity) explored a filtering strategy ... A more systematic exploration ... remains a valuable avenue for future research'). This is a limitation in supporting the abstract's full claim, not a circular step. The evaluation assumption of 'perfect knowledge of previously selected optimal beams' (Section II-C) is a stated input assumption common to the reproduced baseline, and it may inflate accuracy, but it does not make the prediction equivalent to its input by construction. Therefore, no circularity is exhibited.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper's empirical claims rest on the realism of the simulated datasets and the assumption that past optimal beams are known. No new physical entities are introduced.

assumptions (3)
  • standard math Narrowband geometric channel model with L multipath components (Eq. 1)
    Used in Section III-A to define the beam selection problem; standard in mmWave literature.
  • domain assumption Perfect knowledge of previously selected optimal beams
    Stated in Section II-C; used to construct input sequences and evaluate autoregressive inference. In practice the optimal beam is not perfectly known.
  • domain assumption Ray-tracing simulations (Wireless Insite) are representative of real V2I mmWave channels
    The three datasets are simulated; the claimed robustness to NLOS is demonstrated only in simulation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DL-Based Beam Management for mmWave Vehicular Networks Exploring Temporal Correlation." pith.science (2026). https://pith.science/paper/PIBDMQXB

@misc{pith2026251102260,
  author       = {Pith},
  title        = {Pith review of: DL-Based Beam Management for mmWave Vehicular Networks Exploring Temporal Correlation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PIBDMQXB}},
  note         = {Machine review of arXiv:2511.02260}
}
read the original abstract

Millimeter wave communications are essential for modern wireless networks. It supports high data rates but suffers from severe path loss, which requires precise beam alignment to maintain reliable links. This beam management is particularly challenging in highly dynamic scenarios such as vehicle-to-infrastructure, and several methods have been presented. In this work, we propose a deep learning-based beam tracking framework that combines a position-aware beam pre-selection strategy with sequential prediction using recurrent neural networks. The proposed architecture can support deep learning models trained for both classification and regression. In contrast to many existing studies that evaluate beam tracking under predominantly line-of-sight (LOS) conditions, our work explicitly includes highly challenging non-LOS scenarios - with up to 50% non-LOS incidence in certain datasets - to rigorously assess model robustness. Experimental results demonstrate that our approach maintains high top-K accuracy, even under adverse conditions, while reducing the beam measurement overhead by up to 50%.

Figures

Figures reproduced from arXiv: 2511.02260 by the authors.

Figure 1
Figure 1. Representation of beam tracking in a vehicular network [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Aerial visualization of the propagation environments [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Structure of historical beam data for time-series models. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 6
Figure 6. Figure 6: Proposed beam tracking architecture based on LSTM [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Top-K accuracy comparison within proposed models vs. baselines in datasets M10% and R10% (10% NLOS) multi-path propagation. These transitions contrast sharply with the more stable index behavior observed during LOS, where beam direction remains largely consistent [PIT…
Figure 8
Figure 8. Figure 8: Temporal evolution of the optimal beam index across [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 10
Figure 10. Figure 10: Top-K accuracy comparison between RSRP Regres￾sion and Best Index Classification in R10% vs. R50%. Specifically, we analyze three replacement configurations: the 1:1 prediction strategy, where each model prediction is followed by one real measurement; the 2:1 predicti…
Figure 11
Figure 11. Figure 11: Top-K accuracy under different measurement replacement strategies. First row: Best Index Classification; Second row: RSRP Regression. Top to bottom: (a, b) R50%, (c, d) M10%, (e, f) R10% [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]
Figure 12
Figure 12. Figure 12: Top-1 Throughput Ratio for RSRP Regression and [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 2 linked inside Pith

  1. [1]

    Millimeter-wave beamforming as an enabling technology for 5G cellular communications: theoretical feasibility and prototype results,

    W. Roh, J. Seol, J. Park, B. Lee, J. Lee, Y . Kim, J. Cho, K. Cheun, and F. Aryanfar, “Millimeter-wave beamforming as an enabling technology for 5G cellular communications: theoretical feasibility and prototype results,”IEEE Communications Magazine, vol. 52, pp. 106–113, 2014

  2. [2]

    Multi-Frequency Multi-Scenario Millimeter Wave MIMO Channel Measurements and Modeling for B5G Wireless Communication Systems,

    J. Huang, C.-X. Wang, H. Chang, J. Sun, and X. Gao, “Multi-Frequency Multi-Scenario Millimeter Wave MIMO Channel Measurements and Modeling for B5G Wireless Communication Systems,”IEEE Journal on Selected Areas in Communications, vol. 38, pp. 2010–2025, 2020

  3. [3]

    Massive MIMO in Sub-6 GHz and mmWave: Physical, Practical, and Use-Case Differences,

    E. Bjornson, L. Van der Perre, S. Buzzi, and E. G. Larsson, “Massive MIMO in Sub-6 GHz and mmWave: Physical, Practical, and Use-Case Differences,”IEEE Wireless Commun., vol. 26, no. 2, pp. 100–108, 2019

  4. [4]

    Beam training and tracking in mmwave communication: A survey,

    W. Yi, W. Zhiqing, and F. Zhiyong, “Beam training and tracking in mmwave communication: A survey,”China Communications, 2024

  5. [5]

    Feature lead summary #3 evaluation of AI/ML for beam management,

    3GPP, “Feature lead summary #3 evaluation of AI/ML for beam management,” Meeting Document R1-2306199, 3GPP TSG RAN WG1 Meeting #112b-e, e-Meeting, Apr. 2023

  6. [6]

    A survey of beam management for mmWave and THz communications towards 6G,

    Q. Xue, C. Ji, S. Ma, J. Guo, Y . Xu, Q. Chen, and W. Zhang, “A survey of beam management for mmWave and THz communications towards 6G,”IEEE Communications Surveys & Tutorials, 2024

  7. [7]

    Study on New Radio Access Technology: Physical Layer Aspects,

    “Study on New Radio Access Technology: Physical Layer Aspects,” 3rd Generation Partnership Project (3GPP), Tech. Rep. TR 38.802, 2017, version 14.2.0

  8. [8]

    NR; Physical layer procedures for control,

    “NR; Physical layer procedures for control,” 3rd Generation Partnership Project (3GPP), Tech. Rep. TS 38.213, 2024, version 18.3.0

Show all 28 references
  1. [9]

    Extended Kalman filter beam tracking for millimeter wave vehicular communications,

    S. Shaham, M. Kokshoorn, M. Ding, Z. Lin, and M. Shirvanimoghad- dam, “Extended Kalman filter beam tracking for millimeter wave vehicular communications,” in2020 IEEE International Conference on Communications Workshops (ICC Workshops). IEEE, 2020, pp. 1–6

  2. [10]

    Beam tracking channel for millimeter- wave communication system using least mean square algorithm,

    B. A. Asi and F. E. Mohmood, “Beam tracking channel for millimeter- wave communication system using least mean square algorithm,”Al- Rafidain Engineering Journal (AREJ), vol. 26, no. 2, pp. 118–123, 2021

  3. [11]

    LSTM- Based Predictive mmWave Beam Tracking via Sub-6 GHz Channels for V2I Communications,

    Y . Zhao, X. Zhang, X. Gao, K. Yang, Z. Xiong, and Z. Han, “LSTM- Based Predictive mmWave Beam Tracking via Sub-6 GHz Channels for V2I Communications,”IEEE Transactions on Communications, 2024

  4. [12]

    Deep learning-based beam tracking for millimeter-wave communications under mobility,

    S. H. Lim, S. Kim, B. Shim, and J. W. Choi, “Deep learning-based beam tracking for millimeter-wave communications under mobility,” IEEE Trans. Commun., vol. 69, no. 11, pp. 7458–7469, 2021

  5. [13]

    6g virtualized beamforming: a novel framework for optimizing massive mimo in 6g networks,

    A. M. Alwakeel, “6g virtualized beamforming: a novel framework for optimizing massive mimo in 6g networks,”EURASIP Journal on Wireless Communications and Networking, vol. 2025, no. 1, p. 23, 2025

  6. [14]

    Image-Based Beam Tracking With Deep Learning for mmWave V2I Communication Systems,

    W. Zhong, L. Zhang, H. Jin, X. Liu, Q. Zhu, Y . He, F. Ali, Z. Lin, K. Mao, and T. S. Durrani, “Image-Based Beam Tracking With Deep Learning for mmWave V2I Communication Systems,”IEEE Transac- tions on Intelligent Transportation Systems, 2024

  7. [15]

    Ray-Tracing MIMO Channel Dataset for Machine Learning Applied to V2V Communication,

    D. Suzuki, A. Oliveira, L. Gonc ¸alves, I. Correa, A. Klautau, S. Lins, and P. Batista, “Ray-Tracing MIMO Channel Dataset for Machine Learning Applied to V2V Communication,” in2022 IEEE Latin-American Con- ference on Communications (LATINCOM). IEEE, 2022, pp. 1–6

  8. [16]

    Ma- chine learning-based mmwave mimo beam tracking in V2I scenarios: Algorithms and datasets,

    A. Oliveira, D. Suzuki, S. Bastos, I. Correa, and A. Klautau, “Ma- chine learning-based mmwave mimo beam tracking in V2I scenarios: Algorithms and datasets,” in2024 IEEE Latin-American Conference on Communications (LATINCOM), 2024, pp. 1–5

  9. [17]

    Lidar aided future beam prediction in real-world millimeter wave V2I communications,

    S. Jiang, G. Charan, and A. Alkhateeb, “Lidar aided future beam prediction in real-world millimeter wave V2I communications,”IEEE Wireless Communications Letters, vol. 12, no. 2, pp. 212–216, 2022

  10. [18]

    An Overview of Signal Processing Techniques for Millimeter Wave MIMO Systems,

    R. W. Heath, N. Gonz ´alez-Prelcic, S. Rangan, W. Roh, and A. M. Sayeed, “An Overview of Signal Processing Techniques for Millimeter Wave MIMO Systems,”IEEE Journal of Selected Topics in Signal Processing, vol. 10, no. 3, pp. 436–453, 2016

  11. [19]

    A tutorial on beam management for 3GPP NR at mmWave frequencies,

    M. Giordani, M. Polese, A. Roy, D. Castor, and M. Zorzi, “A tutorial on beam management for 3GPP NR at mmWave frequencies,”IEEE Communications Surveys & Tutorials, vol. 21, no. 1, pp. 173–196, 2018

  12. [20]

    Codebook- based hybrid precoding for millimeter wave multiuser systems,

    S. He, J. Wang, Y . Huang, B. Ottersten, and W. Hong, “Codebook- based hybrid precoding for millimeter wave multiuser systems,”IEEE Transactions on Signal Processing, vol. 65, no. 20, pp. 5289–5304, 2017

  13. [21]

    FL summary #5 for AI/ML in beam management,

    3GPP, “FL summary #5 for AI/ML in beam management,” Meeting Document R1-2407554, 3GPP TSG RAN WG1 Meeting #118, Maas- tricht, NL, Apr. 2023

  14. [22]

    Bengio, I

    Y . Bengio, I. Goodfellow, and A. Courville,Deep Learning. MIT Press, 2015

  15. [23]

    Position and LIDAR-aided mmWave beam selection using deep learning,

    M. Dias, A. Klautau, N. Gonz ´alez-Prelcic, and R. W. Heath, “Position and LIDAR-aided mmWave beam selection using deep learning,” in 2019 IEEE 20th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC). IEEE, 2019, pp. 1–5

  16. [24]

    Resnet in resnet: Generalizing residual architectures,

    S. Targ, D. Almeida, and K. Lyman, “Resnet in resnet: Generalizing residual architectures,”arXiv preprint arXiv:1603.08029, 2016

  17. [25]

    Deep learning on multimodal sensor data at the wireless edge for vehicular network,

    B. Salehi, G. Reus-Muns, D. Roy, Z. Wang, T. Jian, J. Dy, S. Ioannidis, and K. Chowdhury, “Deep learning on multimodal sensor data at the wireless edge for vehicular network,”IEEE Transactions on Vehicular Technology, vol. 71, no. 7, pp. 7639–7655, 2022

  18. [26]

    Structural vector autoregressions: Theory of identification and algorithms for inference,

    J. F. Rubio-Ramirez, D. F. Waggoner, and T. Zha, “Structural vector autoregressions: Theory of identification and algorithms for inference,” The Review of Economic Studies, vol. 77, no. 2, pp. 665–696, 2010

  19. [27]

    Study on artificial intelligence (AI)/machine learning (ML) for NR air interface,

    “Study on artificial intelligence (AI)/machine learning (ML) for NR air interface,” 3rd Generation Partnership Project (3GPP), Technical Report TR 38.843, December 2023, version 18.0.0. [Online]. Available: https: //www.3gpp.org/ftp/Specs/archive/38 series/38.843/38843-1800.zip

  20. [28]

    5G-Advanced AI/ML beam management: Performance evaluation with integrated ML models,

    N. Jayaweera, A. Bonfante, M. Schamberger, A. M. A. Tehrani, T. San- guanpuak, P. Tilak, K. Jayasinghe, F. W. V ook, and N. Rajatheva, “5G-Advanced AI/ML beam management: Performance evaluation with integrated ML models,”arXiv preprint arXiv:2404.15326, 2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.