pith. sign in

arxiv: 2605.16735 · v1 · pith:WGFKAZR3new · submitted 2026-05-16 · 💻 cs.NI · cs.LG

Transformer-Based MCS Prediction for 5G Multicast-Broadcast Services (MBS)

Kasidis Arunruangsirilert , Jiro Katto This is my paper

Pith reviewed 2026-05-19 20:00 UTC · model grok-4.3

classification 💻 cs.NI cs.LG
keywords 5GMBSMCS predictionTransformerlink adaptationmulticastbroadcast serviceschannel estimation
0
0 comments X

The pith

Transformer model forecasts safe MCS for 5G video multicast

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

This paper seeks to demonstrate that a transformer model can effectively predict the probability of success for various modulation and coding schemes in 5G multicast-broadcast services over a future time horizon. Such predictions matter because these services operate without retransmission feedback, so any mismatch between chosen parameters and actual channel conditions causes irreversible packet loss and degraded video quality. Conventional methods focus on maximizing data rate and often lead to failures in this setting. By incorporating a loss function designed to penalize overconfident channel estimates, the model learns to favor stable choices that reduce the risk of stalls.

Core claim

The paper establishes that training a lightweight transformer on high-granularity network data with an asymmetric safety loss enables it to output success probabilities for all 28 MCS indices in a way that maintains a conservative bias appropriate for risk-intolerant broadcast transmissions.

What carries the argument

A lightweight Transformer architecture trained using a custom Asymmetric Safety Loss that penalizes overestimation of channel quality.

Load-bearing premise

The collected commercial dataset at 0.5 ms granularity reflects typical conditions in MBS deployments and the safety loss function ensures predictions generalize without overfitting.

What would settle it

A deployment test showing that selected MCS values based on the model predictions frequently result in packet losses exceeding acceptable thresholds for video playback would falsify the reliability of the approach.

Figures

Figures reproduced from arXiv: 2605.16735 by Jiro Katto, Kasidis Arunruangsirilert.

Figure 1
Figure 1. Figure 1: System Architecture Overview (Multicast Part) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: AIS 4T4R Passive Antenna at Dusit Central Park [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: NSG confirming the 4T4R passive panel antenna [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Example temporal configuration of the proposed framework based on evaluated parameters. The model receives a [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: MCS Index vs SS-SINR Heatmap the preliminary experiments, we used the Network Signal Guru (NSG) to execute a high-throughput HTTP GET stress test against the AIS Ookla SpeedTest server, ensuring the downlink channel remained saturated to capture the maximum MCS under varying channel conditions. B. Data Pre-Processing and Feature Engineering The raw logs captured by NSG were converted into Qual￾comm’s Diagn… view at source ↗
Figure 6
Figure 6. Figure 6: Proposed Model Architecture indices (k < m) would also have succeeded. Consequently, we increment both the trial and success counts for m and all k < m. Conversely, if MCS m fails, we do not assume that lower indices would have passed, as the gNodeB may have significantly overestimated the channel quality. In this case, we increment only the trial count for the specific MCS m, without inferring success for… view at source ↗
read the original abstract

The deployment of 5G Multicast-Broadcast Services (MBS) is emerging as a critical technology for spectral-efficient UHD content delivery and serving as a promising solution to modernize CATV deployment. However, unlike unicast networks that rely on RLC-AM with HARQ retransmissions, MBS broadcast operates in RLC Unacknowledged Mode (RLC-UM), where the absence of a feedback loop means packet loss is permanent and immediately impacts user QoE. Conventional link adaptation algorithms, designed for unicast, typically aggressively maximize throughput and fail in this risk-intolerant environment, resulting in severe video stalls and rebuffering. To address this, we propose a lightweight Transformer-based framework that predicts the success probability of all 28 MCS indices over an upcoming video segment horizon. Utilizing a unique commercial network dataset with 0.5 ms slot-level granularity, we train our model using a custom Asymmetric Safety Loss function that penalizes channel overestimation to prioritize link stability. Experimental results show that our approach achieves a reliability score of 86.89%, significantly outperforming standard AI baselines optimized for raw throughput (31.65%) while maintaining a safe conservative bias. Furthermore, the model is optimized for real-time applications, demonstrating an inference time of less than 0.07 ms on COTS 5G-era smartphones.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes a lightweight Transformer-based framework to predict success probabilities for all 28 MCS indices over a video segment horizon in 5G Multicast-Broadcast Services (MBS). Operating in RLC Unacknowledged Mode without feedback, the model is trained on a commercial 0.5 ms slot-level dataset using a custom Asymmetric Safety Loss that penalizes overestimation to enforce conservative predictions. It reports a reliability score of 86.89% versus 31.65% for throughput-optimized AI baselines, with inference latency below 0.07 ms on COTS smartphones.

Significance. If the performance claims are substantiated, the work addresses a practical gap in risk-intolerant MBS deployments by shifting link adaptation from aggressive throughput maximization to reliability-focused prediction. The combination of Transformer architecture with an asymmetric loss and demonstrated real-time inference on mobile hardware could support more stable UHD delivery in broadcast scenarios where packet loss is permanent.

major comments (2)
  1. [Experimental Results] Experimental Results section: the manuscript reports a reliability score of 86.89% but provides no information on dataset size, train-test split ratios, temporal or spatial hold-out strategy, or cross-validation procedure. Without these details the central performance claim cannot be evaluated for robustness or potential overfitting to the specific commercial traces.
  2. [Methodology] Methodology section: no ablation study or quantitative analysis isolates the contribution of the Asymmetric Safety Loss to the reported conservative bias and reliability improvement. It is therefore unclear whether the 86.89% score arises primarily from the Transformer architecture, the loss function, or dataset idiosyncrasies.
minor comments (1)
  1. [Abstract] The exact mathematical definition and computation of the 'reliability score' used for the 86.89% versus 31.65% comparison is not stated in the abstract or early sections, hindering reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. The comments highlight important aspects for strengthening the evaluation of our results and methodology. We address each major comment below and outline the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: [Experimental Results] Experimental Results section: the manuscript reports a reliability score of 86.89% but provides no information on dataset size, train-test split ratios, temporal or spatial hold-out strategy, or cross-validation procedure. Without these details the central performance claim cannot be evaluated for robustness or potential overfitting to the specific commercial traces.

    Authors: We agree that these experimental details are necessary to allow proper assessment of result robustness. The current manuscript provides only high-level information on the commercial 0.5 ms slot-level dataset. In the revised version we will add a dedicated subsection describing the dataset size, the train-test split ratios, the temporal hold-out strategy employed to respect the time-series nature of the traces and avoid leakage, and the cross-validation procedure used. This addition will enable readers to evaluate potential overfitting and the generalizability of the reported 86.89% reliability score. revision: yes

  2. Referee: [Methodology] Methodology section: no ablation study or quantitative analysis isolates the contribution of the Asymmetric Safety Loss to the reported conservative bias and reliability improvement. It is therefore unclear whether the 86.89% score arises primarily from the Transformer architecture, the loss function, or dataset idiosyncrasies.

    Authors: The referee correctly notes the absence of an explicit ablation study isolating the Asymmetric Safety Loss. While the overall performance gains relative to throughput-optimized baselines provide indirect support for the loss design, a direct quantitative comparison would strengthen the claims. In the revised manuscript we will include an ablation analysis that trains the same Transformer architecture with and without the Asymmetric Safety Loss, reporting the resulting changes in conservative bias and reliability score. This will clarify the specific contribution of the loss function. revision: yes

Circularity Check

0 steps flagged

No significant circularity: empirical ML model with independent evaluation

full rationale

The paper describes a standard supervised learning pipeline: a Transformer is trained on a commercial 0.5 ms slot-level dataset using an Asymmetric Safety Loss to predict per-MCS success probabilities, then evaluated on held-out traces for a reliability score. No equations, uniqueness theorems, or self-citations are invoked to derive the reported 86.89 % figure from the training inputs themselves. The performance metric is computed from model outputs on separate data and does not reduce to a fitted parameter or renamed input by construction. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the representativeness of the proprietary commercial dataset and on the effectiveness of the custom loss function; no explicit free parameters, axioms, or invented entities are stated in the provided abstract.

pith-pipeline@v0.9.0 · 5772 in / 1099 out tokens · 34281 ms · 2026-05-19T20:00:01.580303+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages · 1 internal anchor

  1. [1]

    Street smart in 5g: Vehicular applications, communication, and computing,

    A. Y . Alhilal, B. Finley, T. Braud, D. Su, and P. Hui, “Street smart in 5g: Vehicular applications, communication, and computing,”IEEE Access, vol. 10, pp. 105 631–105 656, 2022

    work page 2022

  2. [2]

    Multimedia service trials based on 5g nr mbs,

    L. Xiangjun, P. Changyong, and Z. Qingjun, “Multimedia service trials based on 5g nr mbs,” in2024 8th International Conference on Information, Control, and Communication Technologies (ICCT), 2024, pp. 1–4

    work page 2024

  3. [3]

    A survey on multicast broadcast services in 5g and beyond,

    R. Kamran, P. Jha, S. Kiran, A. Karandikar, P. Chaporkar, A. Saha, and A. Chakraborty, “A survey on multicast broadcast services in 5g and beyond,” in2022 National Conference on Communications (NCC), 2022, pp. 344–349

    work page 2022

  4. [4]

    5g mbs system capability verification in commercial scenario,

    Z. Qingjun, S. Jian, L. Xiangjun, and P. Changyong, “5g mbs system capability verification in commercial scenario,” in2022 International Conference on Electrical Engineering and Photonics (EExPolytech), 2022, pp. 29–32

    work page 2022

  5. [5]

    Private 5g mimo for cable tv ip broadcasting,

    H. Ito, H. Ohno, H. Kitano, and S. Matsumoto, “Private 5g mimo for cable tv ip broadcasting,” in2024 International Conference on Engineering and Emerging Technologies (ICEET), 2024, pp. 1–6. [6]5G; NR; Radio Resource Control (RRC); Protocol specification, 3GPP Standard TS 38.331 version 18.5.1 Release 18, Apr. 2025

    work page 2024

  6. [6]

    3gpp rel-17 extensions for 5g media delivery,

    A. Rico-Alvari ˜no, I. Bouazizi, M. Griot, P. Kadiri, L. Liu, and T. Stockhammer, “3gpp rel-17 extensions for 5g media delivery,”IEEE Transactions on Broadcasting, vol. 68, no. 2, pp. 422–438, 2022. [8]5G; NR; Packet Data Convergence Protocol (PDCP) specification (3GPP TS 38.323 version 18.5.0 Release 18), 3GPP Standard TS 38.323 version 18.5.0 Release 18...

    work page 2022

  7. [7]

    Toward a better quality metric for the video community,

    Z. Li, K. Swanson, C. Bampis, L. Krasula, and A. Aaron, “Toward a better quality metric for the video community,” Dec 2020. [Online]. Available: https://netflixtechblog.com/toward-a-better-quality-metric- for-the-video-community-7ed94e752a30

    work page 2020

  8. [8]

    5g multicast-broadcast for group communication: Why it matters and how it works,

    E. Stare, F. Munier, and D. P. Van, “5g multicast-broadcast for group communication: Why it matters and how it works,” Dec

    work page

  9. [9]

    Available: https://www.ericsson.com/en/blog/2022/12/ multicast-broadcast-group-communication

    [Online]. Available: https://www.ericsson.com/en/blog/2022/12/ multicast-broadcast-group-communication

    work page 2022

  10. [10]

    A channel quality indicator (cqi) prediction scheme using feed forward neural network (ff-nn) technique for mu- mimo lte system,

    M. Q. Abdulhasan, M. I. Salman, C. K. Ng, N. K. Noordin, S. J. Hashim, and F. B. Hashim, “A channel quality indicator (cqi) prediction scheme using feed forward neural network (ff-nn) technique for mu- mimo lte system,” in2014 IEEE 2nd International Symposium on Telecommunication Technologies (ISTT), 2014, pp. 17–22

    work page 2014

  11. [11]

    A machine learning approach for snr prediction in 5g systems,

    K. Saija, S. Nethi, S. Chaudhuri, and R. M. Karthik, “A machine learning approach for snr prediction in 5g systems,” in2019 IEEE International Conference on Advanced Networks and Telecommunica- tions Systems (ANTS), 2019, pp. 1–6

    work page 2019

  12. [12]

    Throughput predic- tion using machine learning in lte and 5g networks,

    D. Minovski, N. ¨Ogren, K. Mitra, and C. ˚Ahlund, “Throughput predic- tion using machine learning in lte and 5g networks,”IEEE Transactions on Mobile Computing, vol. 22, no. 3, pp. 1825–1840, 2023

    work page 2023

  13. [13]

    Deep learning-based throughput prediction in 5g cellular networks,

    I. Batool, M. M. Fouda, and Z. M. Fadlullah, “Deep learning-based throughput prediction in 5g cellular networks,” in2024 International Conference on Smart Applications, Communications and Networking (SmartNets), 2024, pp. 1–6

    work page 2024

  14. [14]

    Uplinknet: Practical commercial 5g standalone (sa) uplink throughput prediction,

    K. Arunruangsirilert and J. Katto, “Uplinknet: Practical commercial 5g standalone (sa) uplink throughput prediction,” in2024 IEEE Interna- tional Conference on Visual Communications and Image Processing (VCIP), 2024, pp. 1–5

    work page 2024

  15. [15]

    Machine learning-based methods for mcs prediction in 5g networks,

    L. Tsipi, M. Karavolos, G. Papaioannou, M. V olakaki, and D. V ouyioukas, “Machine learning-based methods for mcs prediction in 5g networks,”Telecommunication Systems: Modelling, Analysis, Design and Management, vol. 86, no. 4, pp. 705–728, August 2024. [Online]. Available: https://ideas.repec.org/a/spr/telsys/v86y2024i4d10. 1007 s11235-024-01158-x.html

    work page 2024

  16. [16]

    Beyond throughput, the next generation: a 5g dataset with channel and context metrics,

    D. Raca, D. Leahy, C. J. Sreenan, and J. J. Quinlan, “Beyond throughput, the next generation: a 5g dataset with channel and context metrics,” inProceedings of the 11th ACM Multimedia Systems Conference, ser. MMSys ’20. New York, NY , USA: Association for Computing Machinery, 2020, p. 303–308. [Online]. Available: https://doi.org/10.1145/3339825.3394938

    work page doi:10.1145/3339825.3394938 2020

  17. [17]

    Joint link adaptation and scheduling for 5g ultra-reliable low-latency communications,

    G. Pocovi, K. I. Pedersen, and P. Mogensen, “Joint link adaptation and scheduling for 5g ultra-reliable low-latency communications,”IEEE Access, vol. 6, pp. 28 912–28 922, 2018

    work page 2018

  18. [18]

    Adaptive modulation and coding tech- nology in 5g system,

    Y . Wang, W. Liu, and L. Fang, “Adaptive modulation and coding tech- nology in 5g system,” in2020 International Wireless Communications and Mobile Computing (IWCMC), 2020, pp. 159–164

    work page 2020

  19. [19]

    Self-attention-based uplink radio resource prediction in 5g dual connectivity,

    J. Jung, S. Lee, J. Shin, and Y . Kim, “Self-attention-based uplink radio resource prediction in 5g dual connectivity,”IEEE Internet of Things Journal, vol. 10, no. 22, pp. 19 925–19 936, 2023

    work page 2023

  20. [20]

    Performance eval- uations of c-band 5g nr fr1 (sub-6 ghz) uplink mimo on urban train,

    K. Arunruangsirilert, P. Wongprasert, and J. Katto, “Performance eval- uations of c-band 5g nr fr1 (sub-6 ghz) uplink mimo on urban train,” in2023 IEEE Wireless Communications and Networking Conference (WCNC), 2023, pp. 1–6

    work page 2023

  21. [21]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,”

    work page

  22. [22]

    Attention Is All You Need

    [Online]. Available: https://arxiv.org/abs/1706.03762

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    Alpaca: An asymmetric loss prediction algorithm for channel adaptation based on a convolutional- recurrent neural network in urllc systems,

    K. Glinskiy, A. A. Kureev, and E. Khorov, “Alpaca: An asymmetric loss prediction algorithm for channel adaptation based on a convolutional- recurrent neural network in urllc systems,”IEEE Access, vol. 12, pp. 329–338, 2024

    work page 2024

  24. [24]

    On the cost of achieving downlink ultra-reliable low-latency communications in 5g networks,

    G. Pocovi, T. Kolding, and K. I. Pedersen, “On the cost of achieving downlink ultra-reliable low-latency communications in 5g networks,” IEEE Access, vol. 10, pp. 29 506–29 513, 2022

    work page 2022