Pith. sign in

REVIEW 5 major objections 5 minor 7 cited by

A Wireless Foundation Model for Multi-Task Prediction

T0 review · 5 major / 5 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 single pretrained wireless time-series model, using univariate decomposition and a causal Transformer, predicts channel, angle, and traffic data, and zero-shot predicts a new delay task better than traditional full-shot baselines.

desk verdict A credible engineering result—multi-task pretraining across wireless time series transfers zero-shot to delay prediction within the same mobility model—but the out-of-task generalization claim is broader than the single adjacent task supports. read the letter →

arxiv 2507.05938 v2 pith:VZEBUQII submitted 2025-07-08 cs.AI cs.ITcs.LGmath.IT

Yucheng Sheng , Jiacheng Wang , Xingyu Zhou , Le Liang , Hao Ye , Shi Jin , Geoffrey Ye Li This is my paper · ORCID
classification cs.AIcs.ITcs.LGmath.IT
keywords tasksmodelpredictionfoundationlayermulti-tasknetworksproposed
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

Mobile networks need to predict things like signal quality, user direction, and data traffic. The authors build one large neural network, called a foundation model, that handles all three. The network looks at each measurement channel separately, breaks the history into small patches, and uses a causal Transformer to forecast the next few values. A special encoding tells the model whether the data was sampled every half millisecond, every 50 milliseconds, or every hour, because the patterns differ across these time scales.

The model is trained on synthetic channel data from standard 3GPP scenarios, synthetic angle data for four mobility modes, and real mobile traffic data from Milan. After training, the authors test it on an unseen channel scenario and on an entirely new task: predicting the signal propagation delay to a user. Without any retraining, the multi-task model predicts the delay better than RNN, CLSTM, and Informer networks that were trained directly on delay data. Adding more pretraining tasks improves this zero-shot performance.

The claims are plausible but come with caveats. The synthetic datasets are not released, no code is provided, and all results lack error bars or multiple-seed runs. The new task is structurally similar to the angle-prediction task, so calling it a genuinely new type of task overstates the breadth. Still, the paper gives a clear recipe for building domain-specific foundation models in wireless, with evidence that multi-task pretraining transfers.

Extended reading notes

Core claim

After trained on large-scale datasets, the proposed foundation model demonstrates strong generalization to unseen scenarios and achieves zero-shot performance on new tasks that surpass traditional full-shot baselines. (Abstract). Specifically, in Section IV-E, its zero-shot performance on the unseen time-delay prediction task significantly outperforms traditional baselines such as RNN, CLRNet, and Informer. If correct, a single pretrained wireless time-series model, without task-specific fine-tuning, can predict an unseen physical quantity (propagation delay) better than dedicated full-shot models, and the gap grows with the number of pretraining tasks.

Load-bearing premise

The premise that the new task, time-delay prediction, shares the same temporal structure as the pretraining tasks, so that zero-shot transfer works. The paper states in Section IV-E: Structurally, this task resembles previous angle prediction tasks, such as the one described in (11). This is not a restatement of the central claim; it is a data-similarity assumption on which the headline zero-shot result depends. If the target task followed fundamentally different dynamics, such as non-kinematic discontinuities, different noise structure, or a different sampling scale, the multi-task pretraining would not transfer, and the claimed zero-shot superiority over full-shot baselines would vanish. The assumption is untested outside this one delay task.

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 / 5 minor

Summary. The paper proposes a unified causal-Transformer foundation model for wireless time-series prediction, combining univariate decomposition, instance normalization, patching, granularity encoding, and patch masking to handle heterogeneous tasks and arbitrary context lengths. The model is pretrained on synthetic massive-MIMO channel datasets, synthetic ISAC angle datasets, and a real Telecom Italia traffic dataset, then evaluated on channel, angle, and traffic prediction, on downstream spectral-efficiency tasks, and on zero-shot generalization to a held-out channel scenario (D11) and to an unseen time-delay prediction task. The central claim is that the pretrained model achieves state-of-the-art prediction performance on known tasks and that its zero-shot performance on the unseen delay task surpasses traditional full-shot baselines such as RNN, CLRNet, and Informer.

Significance. If the claims hold, the paper makes a useful engineering contribution toward a multi-task wireless foundation model: the unified univariate decomposition and granularity encoding are sensible design choices, the patch-masking trick plausibly enables flexible context lengths, and the reported gains over task-specific baselines on channel, angle, and traffic prediction are nontrivial. The downstream spectral-efficiency results in Tables II and III give practical validation. However, the broader significance is tempered by the fact that the headline out-of-task generalization rests on a single synthetic task that shares its mobility simulator with the angle prediction task, and by the absence of released data, code, and statistical uncertainty quantification. The claims are believable but not yet established at the level the abstract asserts.

major comments (5)
  1. [Section IV-E, Fig. 8] The zero-shot time-delay result is the load-bearing evidence for the abstract's claim that the model achieves 'zero-shot performance on new tasks that surpass traditional full-shot baselines.' Yet the paper explicitly states, 'Structurally, this task resembles previous angle prediction tasks,' and no independent generative model or dataset statistics for the delay task are provided. Since angle and delay in Eq. (8) are different projections of the same simulated user trajectory, the strong result may reflect transfer of smooth trajectory statistics rather than task-general wireless dynamics. The paper provides no ablation isolating the contribution of angle pretraining (e.g., angle-only versus channel-only pretraining) and no second out-of-task target with genuinely different dynamics, noise structure, or sampling scale. Without such evidence, the out-of-task generalization claim is supported by only one adjacent example.
  2. [Section IV-A and all experimental figures] No error bars, multiple-seed runs, or statistical significance tests are reported anywhere, despite the text repeatedly using 'significantly outperforms' (e.g., Section IV-B, Section IV-E). Differences as small as 0.0057 bps/Hz in Table II and a few tenths of a dB in Fig. 3 are presented without uncertainty estimates. This is particularly problematic for the headline zero-shot claim in Fig. 8, where a single NMSE curve without variance is used to argue superiority over three baselines. Please report mean and standard deviation over at least five independent training runs and use appropriate significance tests for the key comparisons.
  3. [Section IV-A, Table I] The dataset D11 is labeled an 'unseen scenario,' but it is a combination of properties already present in pretraining: RMa appears in D5 and D10, LOS appears in D1, D3, D6, and D8, and a 0.5 ms interval appears in D1 through D5. Thus the D11 evaluation is interpolation over scenario attributes rather than a test on a genuinely novel channel regime. This weakens the cross-scenario generalization claim in Section IV-B. Please either qualify the claim as generalization to an unseen combination of known scenario attributes or add a dataset with a truly novel channel model, carrier frequency, or bandwidth.
  4. [Section III-A and Section IV-A] Several hyperparameters essential for reproducibility are missing: the patch length L_p, the embedding dimension d_m, the number of attention heads, the feed-forward dimension, dropout rates, the number of training steps, the optimizer and learning-rate schedule, the total parameter count, and the exact instance-normalization and padding procedure for arbitrary input lengths. In addition, the synthetic CSI, angle, and delay datasets are not released, and no details are given for how the baseline models (RNN, LSTM, Transformer, CLRNet, Informer, STGCN, ASTGNN) were tuned. Without these, the experiments cannot be reproduced or compared fairly. Please add an appendix with complete hyperparameter and training configurations and release code and data.
  5. [Section IV-E] The comparison in Fig. 8 is between zero-shot multi-task pretraining and full-shot single-task baselines, but the paper does not include a full-shot multi-task control that also includes the delay task during pretraining. Such a control would separate the effect of multi-task pretraining from the effect of the delay task appearing in the training set. Additionally, the delay dataset generation is not described at all; readers cannot determine whether the delay task shares the same mobility simulator, speeds, sampling intervals, and noise model as the angle task, which is the key data-similarity assumption behind the zero-shot transfer.
minor comments (5)
  1. [Throughout, e.g., Fig. 2] There are several typos: 'Casual Transformer' should be 'Causal Transformer' in Fig. 2, and 'normalizated' should be 'normalized' in Section III-A. Please proofread the text.
  2. [Table IV] For channel prediction at 0.5 ms, removing the patch module improves NMSE from 0.00823 to 0.00803, yet the text says removing the patch has the greatest impact on traffic prediction. This observation should be acknowledged or explained, since it is the opposite of a degradation for that configuration.
  3. [Section III-A] The description of arbitrary input-length support says the model handles any history length L, but the actual procedure prepends zeros before patching and then applies instance normalization; the effect of zero-padding on the computed mean and variance used for denormalization is not discussed. Please clarify how padded zeros are treated in instance normalization and in the loss.
  4. [Section IV-A] The reported dataset sizes (e.g., 12,960,000 samples for CSI) are stated without defining what constitutes a sample; please specify whether a sample is one context–prediction pair, one sequence, or one time step.
  5. [Section II-A] The preprocessing 'log10(XCP) and log(ˆXCP)' mixes a base-10 logarithm on the input with a natural logarithm on the output; please use consistent notation and specify whether the transformation is applied elementwise to magnitudes only.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the zero-shot delay result is an empirical transfer test on a task excluded from training, not a reduction to fitted inputs or self-citation.

full rationale

The paper's central claims are empirical. The foundation model is pretrained on channel, angle, and traffic tasks, and the headline zero-shot result (Section IV-E, Fig. 8) is evaluated on a time-delay task that is explicitly excluded from training; no delay value is used to set model parameters or design constants. Granularity encoding thresholds in Eq. (22) are hand-chosen design choices based on sampling interval rather than fitted to the delay targets, and the ablation in Table IV only measures their effect on the training tasks. The cited components ([26] univariate decomposition, [27] instance normalization, [31] patch masking) are standard external methods, and the one self-citation ([21]) is not load-bearing. The sentence 'Structurally, this task resembles previous angle prediction tasks' is a transferability assumption, not a circular definition, and Fig. 8 tests, rather than assumes, the transfer. If the assumption were false the zero-shot gains would fail, but that is a correctness risk about task similarity, not circularity. No derivation step in the paper reduces to its own input.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical, so the ledger records hand-chosen design parameters and domain assumptions rather than fitted physical constants. The granularity thresholds in Eq. (22), the unspecified patch length, and the omitted model width and head counts are all load-bearing choices. The domain assumptions include univariate decomposition preserving performance, synthetic datasets representing real wireless environments, and the delay task sharing dynamics with the pretraining tasks. No new physical entities are introduced.

free parameters (3)
  • Granularity thresholds and resolution categories = high: <1 ms; medium: 1 ms to <1 h; low: >=1 h
    Hand-chosen cutoffs in Eq. (22) categorize sampling intervals into three levels. The ablation in Table IV shows that removing granularity encoding degrades NMSE substantially, so these thresholds are load-bearing design choices, not derived from data.
  • Patch length L_p and maximum context lengths = L_p not stated; max history 32/32/64 for channel/angle/traffic
    Patching length determines N_p and the patch masking schedule. The paper never specifies L_p, yet the arbitrary-history-length capability in Fig. 9 depends on it.
  • Model architecture hyperparameters = 16 layers; d_m, number of heads, FFN size unspecified
    The central performance claims depend on the Transformer width and depth. These values are not reported, limiting reproduction and making the compute budget hard to verify.
assumptions (4)
  • domain assumption Univariate decomposition of multivariate time series preserves or improves forecasting performance for the target tasks.
    The model processes each of the M channels independently in Section III-A, Eq. (17), citing [26]. This discards cross-channel and spatial correlations, such as antenna correlations in MIMO CSI and spatial correlations in traffic grids, and assumes per-channel temporal modeling is sufficient.
  • domain assumption QuaDRiGa-generated channel data and the synthetic ISAC angle data are representative of real-world wireless environments.
    D1-D11 are generated with QuaDRiGa under 3GPP scenarios in Section IV-A, and the angle datasets are synthetic mobility simulations. The foundation model generalization claims rest on these synthetic datasets capturing real channel and angle dynamics.
  • domain assumption The time-delay prediction task shares the same dynamics as the pretraining tasks, enabling zero-shot transfer.
    Section IV-E states that the new task structurally resembles previous angle prediction tasks. This similarity assumption is load-bearing for the headline zero-shot claim.
  • domain assumption The mmWave LOS channel model in Eqs. (9)-(10) and the path loss model alpha_k,t = alpha0(d/d0)^(-zeta) with fixed parameters describe the ISAC system.
    These models are used to generate the angle prediction data and the downstream beamforming evaluation in Section II-B.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Wireless Foundation Model for Multi-Task Prediction." pith.science (2026). https://pith.science/paper/VZEBUQII

@misc{pith2026250705938,
  author       = {Pith},
  title        = {Pith review of: A Wireless Foundation Model for Multi-Task Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VZEBUQII}},
  note         = {Machine review of arXiv:2507.05938}
}
read the original abstract

With the growing complexity and dynamics of the mobile communication networks, accurately predicting key system parameters, such as channel state information (CSI), user location, and network traffic, has become essential for a wide range of physical (PHY)-layer and medium access control (MAC)-layer tasks. Although traditional deep learning (DL)-based methods have been widely applied to such prediction tasks, they often struggle to generalize across different scenarios and tasks. In response, we propose a unified foundation model for multi-task prediction in wireless networks that supports diverse prediction intervals. The proposed model enforces univariate decomposition to unify heterogeneous tasks, encodes granularity for interval awareness, and uses a causal Transformer backbone for accurate predictions. Additionally, we introduce a patch masking strategy during training to support arbitrary input lengths. After trained on large-scale datasets, the proposed foundation model demonstrates strong generalization to unseen scenarios and achieves zero-shot performance on new tasks that surpass traditional full-shot baselines.

Figures

Figures reproduced from arXiv: 2507.05938 by the authors.

Figure 1
Figure 1. Three representative prediction tasks within wireless [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The architecture of the proposed wireless foundation model for multi-task prediction. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The NMSE performance of the proposed method compared with other baselines in the channel prediction task. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The zero-shot generalization performance in the D11 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: The NMSE performance of our proposed foundation models compared with other baselines in the angle prediction task [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: The zero-shot generalization performance of our proposed foundation models compared with other baselines in the [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: The NMSE performance of the foundation models [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 9
Figure 9. Figure 9: The NMSE performance of the foundation models for [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: The NMSE performance of the foundation model [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 7 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MultiPathFormer: Towards a Foundation Model for Multipath Wireless Propagation

    cs.LG 2026-08 conditional novelty 6.0 of 10

    Pretraining a transformer on ordered multipath sequences transfers to downstream wireless tasks like beam prediction and localization in simulated environments.

  2. Generalizable and Computational Efficient Channel Extrapolation for 6G: A Configurable AI-Driven Framework Built from a Modular Perspective

    eess.SP 2026-08 conditional novelty 6.0 of 10

    A three-stage modular AI framework, pretrain, cluster experts, and learn routing, improves channel extrapolation accuracy and cuts FLOPs in simulated 6G scenarios.

  3. Cross-System Neural Precoder: Exploiting Structural Consistency for Fast Adaptation

    eess.SP 2026-07 conditional novelty 6.0 of 10

    AO subproblems for digital, analog, and RIS variables share reusable update structure, so a lightweight unfolded network adapts across precoding systems with far less data than a GNN baseline.

  4. Hierarchical Wireless Foundation Model for Multi-Task Optimization

    eess.SP 2026-07 conditional novelty 6.0 of 10

    A hierarchical wireless foundation model with a shared channel encoder and prompt-conditioned decoder solves beamforming, scheduling, channel estimation, and beam selection with competitive performance and large laten...

  5. CFM-Bench: A Unified Multi-Domain, Multi-Task Benchmark for Channel Foundation Models

    cs.AI 2026-07 conditional novelty 6.0 of 10

    CFM-Bench provides six curated radio-channel datasets with leakage-resistant splits and six task groups for standardizing evaluation of channel foundation models.

  6. EMind: A Foundation Model for Multi-task Electromagnetic Signals Understanding

    eess.SP 2025-08 conditional novelty 5.0 of 10

    EMind reports that one masked-autoencoder transformer, pretrained on 81 million heterogeneous IQ samples, transfers to seven electromagnetic signal tasks with strong accuracy, but post-hoc checkpoint selection and mis...

  7. Towards channel foundation models (CFMs): Motivations, methodologies and opportunities

    eess.SP 2025-07 conditional novelty 4.0 of 10

    A survey and position paper proposing channel foundation models, with experiments on two pretrained CSI models showing gains over a vanilla ViT baseline.

Reference graph

Works this paper leans on

35 extracted references · 25 canonical work pages · cited by 7 Pith papers

  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]

    Neural network-based fading channel prediction: A comprehensive overview,

    W. Jiang and H. D. Schotten, “Neural network-based fading channel prediction: A comprehensive overview,” IEEE Access, vol. 7, pp. 118 112–118 124, Aug. 2019

    work page 2019

  3. [3]

    Deep CLSTM for predictive beamforming in integrated sensing and communication-enabled vehicular networks,

    C. Liu, X. Liu, S. Li, W. Yuan, and D. W. K. Ng, “Deep CLSTM for predictive beamforming in integrated sensing and communication-enabled vehicular networks,” J. Commun. Inf. Netw., vol. 7, no. 3, pp. 269–277, Sep. 2022

    work page 2022

  4. [4]

    LSTM network: A deep learning approach for short-term traffic fore- cast,

    W. Chen, Z. Zhao, J. Liu, P. C. Y . Chen, and X. Wu, “LSTM network: A deep learning approach for short-term traffic fore- cast,” IET Intell. Transp. Syst. , vol. 11, no. 2, pp. 66–75, Jan. 2017

    work page 2017

  5. [5]

    Addressing the curse of mo- bility in massive mimo with prony-based angular-delay domain channel predictions,

    H. Yin, H. Wang, Y . Liu et al. , “Addressing the curse of mo- bility in massive mimo with prony-based angular-delay domain channel predictions,” IEEE J. Sel. Areas Commun. , vol. 38, pp. 2903–2917, Dec. 2020

    work page 2020

  6. [6]

    Performance analysis of channel extrapolation in FDD massive MIMO systems,

    F. Rottenberg, T. Choi, P. Luo et al., “Performance analysis of channel extrapolation in FDD massive MIMO systems,” IEEE Trans. Wirel. Commun., vol. 19, pp. 2728–2741, Apr. 2020

    work page 2020

  7. [7]

    Deep learning for fading channel prediction,

    W. Jiang and H. D. Schotten, “Deep learning for fading channel prediction,” IEEE Open J. Commun. Soc. , vol. 1, pp. 320–332, Mar. 2020

    work page 2020

  8. [8]

    Accurate channel prediction based on transformer: Making mobility negligible,

    H. Jiang, M. Cui, D. W. K. Ng, and L. Dai, “Accurate channel prediction based on transformer: Making mobility negligible,” IEEE J. Sel. Areas Commun. , vol. 40, no. 9, pp. 2717–2732, Sep. 2022

    work page 2022

Show all 35 references
  1. [9]

    Cram´er-rao bound optimization for joint radar-communication beamforming,

    F. Liu, Y .-F. Liu, A. Li, C. Masouros, and Y . C. Eldar, “Cram´er-rao bound optimization for joint radar-communication beamforming,” IEEE Trans. Signal Process. , vol. 70, pp. 240– 253, Dec. 2021

    2021

  2. [10]

    Radar-assisted predic- tive beamforming for vehicular links: communication served by sensing,

    F. Liu, W. J. Yuan, C. Masouros et al., “Radar-assisted predic- tive beamforming for vehicular links: communication served by sensing,” IEEE Trans. Wireless Commun. , vol. 19, no. 11, pp. 7704–7719, Nov. 2020

    2020

  3. [11]

    Bayesian predictive beamforming for vehicular networks: a low-overhead joint radar-communication approach,

    W. J. Yuan, F. Liu, C. Masouros et al. , “Bayesian predictive beamforming for vehicular networks: a low-overhead joint radar-communication approach,” IEEE Trans. Wireless Com- mun., vol. 20, no. 3, pp. 1442–1456, Mar. 2020

    2020

  4. [12]

    Spatial–temporal graph neural network traffic prediction based load balancing with reinforcement learning in cellular networks,

    S. Liu, M. He, Z. Wu, P. Lu, and W. Gu, “Spatial–temporal graph neural network traffic prediction based load balancing with reinforcement learning in cellular networks,” Inf. Fusion, vol. 103, p. 102079, Mar. 2024

    2024

  5. [13]

    A data-driven base station sleeping strategy based on traffic prediction,

    J. Lin et al., “A data-driven base station sleeping strategy based on traffic prediction,” IEEE Trans. Netw. Sci. Eng., vol. 8, no. 4, pp. 2755–2766, Sep. 2021

    2021

  6. [14]

    A survey of anticipatory mobile networking: Context-based classification, prediction methodologies, and op- timization techniques,

    N. Bui, M. Cesana, S. A. Hosseini, Q. Liao, I. Malanchini, and J. Widmer, “A survey of anticipatory mobile networking: Context-based classification, prediction methodologies, and op- timization techniques,” IEEE Commun. Surveys Tuts. , vol. 19, no. 3, pp. 1790–1821, 3rd Quart. 2017

    2017

  7. [15]

    A survey of time series forecasting from stochastic method to soft computing,

    P. Hendikawati, “A survey of time series forecasting from stochastic method to soft computing,” J. Phys.: Conf. Ser., vol. 1613, no. 1, 2020

    2020

  8. [16]

    A study of deep learn- ing networks on mobile traffic forecasting,

    C.-W. Huang, C.-T. Chiang, and Q. Li, “A study of deep learn- ing networks on mobile traffic forecasting,” in Proc. PIMRC , 2017, pp. 1–6

    2017

  9. [17]

    Spatio-temporal graph convolu- tional networks: A deep learning framework for traffic fore- casting,

    B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolu- tional networks: A deep learning framework for traffic fore- casting,” arXiv preprint arXiv:1709.04875 , 2017

    2017 arXiv

  10. [18]

    Attention based spatial-temporal graph convo- lutional networks for traffic flow forecasting,

    S. Guo et al. , “Attention based spatial-temporal graph convo- lutional networks for traffic flow forecasting,” in Proc. AAAI , vol. 33, no. 01, 2019, pp. 922–929

    2019

  11. [19]

    Language models are unsupervised multitask learners,

    A. Radford et al., “Language models are unsupervised multitask learners,” OpenAI blog, vol. 1.8, p. 9, 2019

    2019

  12. [20]

    Llama 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. R. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosaleet al., “Llama 2: Open foundation and fine-tuned chat models,” arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  13. [21]

    Beam prediction based on large language models,

    Y . Sheng, K. Huang, L. Liang, P. Liu, S. Jin, and G. Y . Li, “Beam prediction based on large language models,” IEEE Wirel. Commun. Lett., vol. 14, no. 5, pp. 1406–1410, May 2025

    2025

  14. [22]

    LLM4CP: Adapting large language models for channel prediction,

    B. Liu, X. Liu, S. Gao, X. Cheng, and L. Yang, “LLM4CP: Adapting large language models for channel prediction,” J. Commun. Inf. Netw., vol. 9, no. 2, pp. 113–125, Jun. 2024

    2024

  15. [23]

    Large language model enabled multi- task physical layer network,

    T. Zheng and L. Dai, “Large language model enabled multi- task physical layer network,” arXiv preprint arXiv:2412.20772, 2024

    2024 arXiv

  16. [24]

    On the opportunities and risks of foundation models,

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill et al. , “On the opportunities and risks of foundation models,” arXiv preprint arXiv:2108.07258 , 2021

    2021 arXiv

  17. [25]

    A survey of millimeter wave communications (mmWave) for 5G: opportu- nities and challenges,

    Y . Niu, Y . Li, D. Jin, L. Su, and A. V . Vasilakos, “A survey of millimeter wave communications (mmWave) for 5G: opportu- nities and challenges,” Wireless Netw., vol. 21, pp. 2657–2676, Apr. 2015

    2015

  18. [26]

    A time series is worth 64 words: Long-term forecasting with transformers,

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: Long-term forecasting with transformers,” arXiv preprint arXiv:2211.14730 , 2022

    2022 arXiv

  19. [27]

    Instance normaliza- tion: The missing ingredient for fast stylization,

    D. Ulyanov, A. Vedaldi, and V . Lempitsky, “Instance normaliza- tion: The missing ingredient for fast stylization,” arXiv preprint arXiv:1607.08022, 2016

    2016 arXiv

  20. [28]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. CVPR, 2016, pp. 770–778

    2016

  21. [29]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Proc. NeurIPS, vol. 30, 2017

    2017

  22. [30]

    Are transformers effective for time series forecasting?

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, “Are transformers effective for time series forecasting?” in Proc. AAAI, vol. 37, no. 9, 2023, pp. 11 121–11 128

    2023

  23. [31]

    A decoder-only foundation model for time-series forecasting,

    A. Das, W. Kong, R. Sen, and Y . Zhou, “A decoder-only foundation model for time-series forecasting,” in Proc. ICML, 2024

    2024

  24. [32]

    QuaDRiGa: A 3-D multi-cell channel model with time evo- lution for enabling virtual field trials,

    S. Jaeckel, L. Raschkowski, K. B ¨orner, and L. Thiele, “QuaDRiGa: A 3-D multi-cell channel model with time evo- lution for enabling virtual field trials,” IEEE Trans. Antennas Propag., vol. 62, no. 6, pp. 3242–3256, Jun. 2014

    2014

  25. [33]

    A multi- source dataset of urban life in the city of milan and the province of trentino,

    G. Barlacchi, M. De Nadai, R. Larcher, Casella et al., “A multi- source dataset of urban life in the city of milan and the province of trentino,” Sci. Data, vol. 2, no. 1, pp. 1–15, Oct. 2015

    2015

  26. [34]

    Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network,

    A. Sherstinsky, “Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network,” Phys. D, Nonlinear Phenom. , vol. 404, p. 132306, Mar. 2020

    2020

  27. [35]

    Informer: Beyond efficient transformer for long sequence time-series forecasting,

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” in Proc. AAAI, vol. 35, no. 12, 2021, pp. 11 106–11 115

    2021

Pith tools

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