Pith. sign in

REVIEW 3 major objections 4 minor 38 references

The paper claims that a hierarchical wireless foundation model, coupling a masked-autoencoder channel encoder with a prompt-conditioned optimization decoder, matches numerical baselines on four multi-antenna tasks while running up to 39x fa

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

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

T0 review · deepseek-v4-flash

2026-08-01 19:42 UTC pith:TRJ5CHCB

load-bearing objection Genuinely novel wireless foundation model for optimization tasks, but the headline latency claim is overstated because it excludes the shared encoder; needs major revision before acceptance. the 3 major comments →

arxiv 2607.16877 v1 pith:TRJ5CHCB submitted 2026-07-18 eess.SP cs.LG

Hierarchical Wireless Foundation Model for Multi-Task Optimization

Yangjing Wang , Ouya Wang , Shenglong Zhou , Geoffrey Ye Li This is my paper
classification eess.SP cs.LG
keywords wireless foundation modelmulti-task optimizationchannel representation learningmasked autoencoderbeamforminguser schedulingchannel estimationbeam selection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that one shared wireless foundation model can replace task-specific deep-learning or numerical pipelines for four distinct physical-layer optimization problems: beamforming, user scheduling, channel estimation, and beam selection. The key move is to split the problem in two: an upstream foundation channel encoder (FCE), pre-trained by self-supervised masked reconstruction, learns universal task-agnostic channel representations, while a downstream foundation optimization decoder (FOD) turns task identity and constraints into prompts that query those representations through geometry-aware cross-attention. On simulated urban MU-MISO-OFDM channels, the model reports 98.5% of WMMSE beamforming sum rate, near-greedy scheduling QoS, LMMSE-comparable channel estimation, and DFT-greedy-beating beam selection, at 14–39x lower inference latency than the numerical baselines, plus zero-shot generalization to unseen scenarios, constraints, and (K, N, M) configurations. If true, this supports a 'train once, deploy everywhere' paradigm for wireless physical-layer optimization.

Core claim

The central discovery is that channel representation and task decision can be cleanly separated inside a single model: the FCE encodes the full CSI tensor once, with per-user patching, block-wise masking, and a phase-aware loss (MSE plus cosine similarity), and the FOD reuses that encoding for every task by attending to it from user-anchored queries that carry task and constraint information. The paper demonstrates this on four tasks with one shared backbone: beamforming, user scheduling, channel estimation via an upstream adapter, and beam selection via a downstream head, achieving near-baseline performance while cutting inference latency sharply and generalizing to unseen propagation envir

What carries the argument

The load-bearing object is the FCE–FOD pair: an upstream masked autoencoder pre-trained with a high masking ratio and a phase-aware objective, and a downstream cross-attention decoder whose input is a composite prompt of user anchors, task embeddings, and affine constraint encodings. Geometry-aware cross-attention injects 2D rotary position embeddings keyed to frequency and spatial patch coordinates, and differentiable output heads — an INR-parameterized beamforming head, a Gumbel-Sigmoid scheduling head, a pilot-aware reconstruction adapter for estimation, and a Fourier-basis beam-selection head — map decoded features to feasible decisions.

Load-bearing premise

A single global CSI amplitude-scaling factor computed from the training corpus must stay valid for all unseen environments and configurations; if amplitude distributions shift across scenarios, the shared representation is systematically miscalibrated and the reported zero-shot gains could shrink.

What would settle it

Re-run the zero-shot evaluation after re-scaling each unseen scenario by its own q-quantile scaling factor; if performance changes materially, the single global factor is not transferable and the generalization claim rests on accidental calibration. Alternatively, compute per-scenario factors and show a monotonic degradation as they deviate from the training corpus factor.

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

If this is right

  • A single pre-trained backbone can be adapted to new tasks with only lightweight head changes, supporting efficient deployment as the task set grows.
  • Because the FCE is computed once and shared, per-task inference cost stays nearly flat: FOD latencies of 2.67 ms and 0.95 ms versus 104 ms for WMMSE and 30.4 ms for greedy scheduling.
  • The hybrid supervised-then-unsupervised training stage improves on the supervised solution for beamforming and enforces QoS constraints for scheduling, suggesting that unsupervised utility optimization can push past algorithmic labels.
  • Zero-shot results across 27 unseen (K, N, M) configurations and 30 unseen urban scenarios indicate the representation transfers without retraining, at least within simulated urban channel data.
  • WFM-CE beats LMMSE in moderate-SNR and dense-pilot regimes without oracle channel statistics, hinting that pre-trained channel priors can substitute for explicit second-order statistics.

Where Pith is reading between the lines

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

  • If the single global scaling premise holds, the same FCE could be reused for FDD CSI feedback by treating the learned representation as a compact codebook — a direction the paper itself names.
  • The ablation results suggest geometry-aware cross-attention is load-bearing; a natural test is whether a simpler positional encoding scheme at matched parameter count recovers the same zero-shot gains.
  • The q-quantile scaling factor is never specified or validated across the 30 held-out scenarios; a concrete check would be to compute per-scenario scaling factors and see if the model's zero-shot performance tracks their deviation from the training-corpus factor.
  • The architecture invites a broader 'channel foundation model' benchmark: if the same FCE weights can be probed for other tasks beyond these four (e.g., localization or sensing), that would strengthen the task-agnostic representation claim.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a hierarchical wireless foundation model (WFM) for multi-task physical-layer optimization. An upstream foundation channel encoder (FCE) is pre-trained via masked autoencoding with amplitude and phase losses, and a downstream foundation optimization decoder (FOD) maps the resulting representation to beamforming, user scheduling, channel estimation, and beam selection through prompt-conditioned cross-attention. Training proceeds from supervised warm-up to unsupervised utility maximization, and task adaptation is confined to lightweight modules. The central claims are: competitive or near-baseline performance on four tasks, robust zero-shot generalization to unseen scenarios/configurations, and substantially lower inference latency than numerical baselines (e.g., 98.5% of WMMSE sum rate at 14–39x lower latency).

Significance. If the performance and latency claims survive scrutiny, this would be a useful step toward unified wireless physical-layer models: one shared encoder supporting heterogeneous tasks, with zero-shot generalization tested across scenarios, system dimensions, and constraints. The experimental design has real strengths: performance is compared against external numerical baselines (WMMSE, LMMSE, greedy scheduling, sequential greedy beam selection); the unsupervised losses directly maximize the physical utility rather than fitting baseline outputs, so the core comparison is not circular; and the modular adaptation paths for channel estimation and beam selection are concretely specified. The paper also provides extensive simulations on a 60-scenario DeepMIMO corpus. However, the central latency advantage is not actually measured end-to-end: the FCE forward pass is excluded from the reported WFM-FOD/WFM-BS latencies, so the advertised speedups are unsupported as stated. The paper would also be strengthened by variance reporting for the headline comparisons. These issues are fixable within the manuscript's scope, hence my recommendation.

major comments (3)
  1. [§VIII-E, Table II; Abstract; §IX] The reported latencies for WFM-FOD and WFM-BS exclude the FCE, but the FCE forward pass is mandatory in the architecture (Fig. 1, §IV). The exact additional latency for downstream tasks is not reported; the 1.87 ms WFM-CE row includes the decoder and is therefore only an upper bound. Still, the current table cannot establish the claimed '14–39x' speedup or the '2.3x faster than Seq Greedy' statement. For example, even a fraction of the FCE encoder cost may exceed the entire Seq Greedy latency (2.15 ms) when added to WFM-BS's 0.95 ms. Please report true end-to-end latency for each task, and separately report the amortized per-task latency when multiple tasks share one FCE pass. This is a load-bearing issue because latency reduction is a headline contribution.
  2. [§VIII-C–D, Figs. 6, 8–10, Tables II–III] The principal performance comparisons are reported as point estimates with no variance. Figs. 6, 8–10 show no error bars or shaded bands, and Tables II–III give single numbers for each method. The text in §VIII-D claims 'minor shifts in both median and dispersion' in Fig. 10, but no dispersion is actually quantified. Since many claims are of the form 'comparable to WMMSE' or 'on par with WLLM' (e.g., BF 98.5% vs 100%, US rate 93.2% vs 94.3%), it is impossible to assess whether the differences are meaningful. Please report means and standard deviations over seeds (and/or bootstrapped confidence intervals over CSI samples) for the key figures and tables.
  3. [§V-A] The global quantile scaling factor is defined as the reciprocal of the q-quantile of absolute CSI amplitudes over the training corpus, but q is never specified. This scaling is an upstream premise for the entire pipeline, including pre-training and all downstream heads. The zero-shot generalization claims in §VIII-D assume that the same scaling remains appropriate for the 30 held-out urban scenarios and all unseen (K,N,M) configurations. Because the corpus is deliberately heterogeneous in RMS delay spread and LoS probability (Fig. 4), the amplitude distribution may also vary. Please report q and add a sensitivity analysis (e.g., varying q, or validating that the scaling transfers across the cross-scenario set).
minor comments (4)
  1. [§V-A, §VI-D, §VII] Several hyperparameters with direct influence on the reported results are not given numerical values: q (quantile scaling), β (pre-training phase-loss weight), λ (channel-estimation phase weight), ρ and λ_max (augmented Lagrangian), and η (beam-selection residual scale). The masking ratio and patch size are stated; the others should be reported for reproducibility.
  2. [§VIII-A, Fig. 6] The training ranges for SNR and pilot ratio are not specified. The text describes low-SNR and sparse-pilot regimes as 'OOD' but never quantifies the boundary of the training distribution. Please state the sampling ranges used during channel-estimation adaptation so that 'in-distribution' and 'OOD' are well defined.
  3. [§VIII-E, Table II] The numerical baselines (WMMSE, Greedy, Seq Greedy) are reported with '–' for parameters/FLOPs; please specify the hardware and software configuration used for latency measurements, and confirm whether all methods were benchmarked on the same machine and batch settings. This is useful context for interpreting the speedups.
  4. [§VIII-B, Table II] For channel estimation, Table II shows WFM-CE at -13.3 dB NMSE while the task-specific Transformer achieves -14.0 dB. The text says they are 'comparable' within the training range; please clarify whether the table corresponds to in-domain or cross-scenario settings, and avoid overstating the CE comparison in the abstract/conclusion.

Circularity Check

0 steps flagged

No significant circularity: the WFM's core claims are validated against external numerical baselines and unsupervised physical-utility objectives, not against its own fitted inputs.

full rationale

I walked the derivation chain: (1) FCE pre-training reconstructs masked CSI with the loss in Eq. (20); no downstream benchmark or baseline enters the pre-training objective, and the quantile scaling is a data-dependent preprocessing choice, not a fit to any claimed result. (2) The FOD uses supervised warm-up with WMMSE/greedy decompositions as targets (Eqs. (31)-(32)), but the final optimization uses unsupervised losses that directly maximize sum rate or enforce QoS (Eqs. (33)-(34), (40)); the reported numbers are normalized against external WMMSE/Greedy/Seq-Greedy baselines, so the near-baseline performance is not forced by construction. (3) Channel-estimation adaptation uses a pilot-aware masked-reconstruction formulation evaluated against LS/LMMSE; beam-selection adaptation uses a single-user gain prior plus a learned residual trained by the unsupervised sum-rate loss, not by fitting Seq Greedy outputs. None of these steps equate the prediction with the input or rename a fitted parameter as a prediction. (4) The authors cite their own prior works ([3], [5], [7], [8], [11]), but these are background/motivational citations and do not supply a load-bearing uniqueness theorem, ansatz, or result. The Table II note excluding FCE latency from the WFM-FOD/WFM-BS rows is a legitimate quantitative-accounting concern that may affect the latency claim, but it is not a circular-derivation defect: it does not make a prediction equal to its input by definition. I therefore find no circularity and score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The model depends on many hand-chosen values. The most consequential are the corpus-derived amplitude quantile scale (§V-A), loss weights β/λ, augmented-Lagrangian ρ/λ_max, the beam-selection residual scale η, and the masking ratio/patch size (75%, (4,2)). None of these are reported with sensitivity analysis. The architecture also relies on standard results (Eq. 25) and domain assumptions (perfect CSI, transferable scaling, RoPE interpolation) that are unproven for the OOD settings claimed.

free parameters (6)
  • Global quantile scaling factor q (reciprocal of q-quantile of absolute amplitudes) = not stated (q unspecified)
    Defined in §V-A as reciprocal of q-quantile of absolute amplitudes aggregated over training corpus; chosen by hand to normalize CSI, and is load-bearing for unseen-scenario generalization.
  • β (pre-training phase-loss weight) = not stated
    In Eq (20), balances MSE and cosine-similarity; chosen by hand; not reported.
  • λ (channel-estimation phase-alignment weight) = not stated
    In Eq (37), controls relative weight of phase-alignment loss; not reported.
  • ρ (penalty coefficient) and λ_max for augmented Lagrangian = not stated
    In Eq (34)-(35), used for unsupervised user-scheduling; chosen by hand; no values given.
  • η (beam-selection residual scale) = not stated
    In Eq (38), scales learnable residual correction relative to the DFT gain prior; chosen by hand; not reported.
  • Masking ratio and patch size = mask=75%, patch=(4,2)
    Architecture choices in §VIII-A that determine the pre-training task's difficulty; chosen by hand with no sensitivity analysis.
axioms (5)
  • standard math Closed-form optimal beamforming structure Eq (25) [Björnson et al.]
    Used in §VI-C1 to reduce beamforming to predicting regularization and power matrices; standard result from convex optimization of downlink beamforming.
  • domain assumption MU-MISO-OFDM TDD system with perfect CSI at the BS for optimization tasks
    All optimization formulations in §III assume perfect CSI for beamforming, scheduling, and beam selection; channel estimation is treated as a separate upstream task.
  • ad hoc to paper A single task-agnostic channel representation from masked reconstruction is sufficient for four heterogeneous optimization tasks
    The central architectural premise of the FCE-FOD split (§IV); not derived from theory, supported only by the reported experiments.
  • ad hoc to paper Corpus amplitude quantile factor transfers to unseen scenarios
    In §V-A, global scaling is fixed from training corpus; no analysis of how large-scale fading distribution shift affects held-out scenarios.
  • domain assumption 2D RoPE with position interpolation extrapolates to unseen grid sizes
    Footnote 1 in §VI-B; relies on linear position interpolation from [31] to handle coordinates outside the training range; no separate validation.

pith-pipeline@v1.3.0-alltime-deepseek · 103 in / 13551 out tokens · 121088 ms · 2026-08-01T19:42:00.365836+00:00 · methodology

0 comments
read the original abstract

The increasing complexity of next-generation wireless networks has driven the integration of artificial intelligence (AI) into wireless communications. However, most existing studies focus on developing task-specific deep learning techniques for single scenarios, which limits their ability to generalize across diverse tasks, channel conditions, and system configurations. To address this generalization bottleneck, we propose a hierarchical wireless foundation model (WFM) for multi-task optimization. The proposed WFM couples an upstream foundation channel encoder (FCE) with a downstream foundation optimization decoder (FOD) via geometry-aware cross-attention. Specifically, the FCE extracts task-agnostic channel representations via self-supervised masked reconstruction while the FOD generates multi-task optimization decisions through differentiable output heads. Moreover, a hybrid supervised-to-unsupervised training strategy is employed to overcome the performance ceiling of purely supervised learning, and the modular architecture of the WFM enables efficient adaptation to unseen communication tasks with minimal parameter overhead. Simulation results show that the proposed WFM learns high-fidelity channel representations and achieves competitive multi-task optimization performance while substantially reducing optimization inference latency relative to numerical baselines. Furthermore, it exhibits robust generalization to unseen propagation environments, varying constraint parameters, and heterogeneous system configurations.

Figures

Figures reproduced from arXiv: 2607.16877 by Geoffrey Ye Li, Ouya Wang, Shenglong Zhou, Yangjing Wang.

Figure 1
Figure 1. Figure 1: Overall workflow of the proposed hierarchical WFM. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Network architecture of the proposed FCE. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Network architecture of the proposed FOD. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the multi-scenario CSI corpus. The upper panel maps the geographic distribution of the urban scenarios, while the lower panel shows [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training convergence curves for (a) FCE pre-training and (b) channel [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Channel estimation performance of WFM-CE against baseline methods: (a) NMSE versus SNR and (b) NMSE versus pilot ratio. Solid and dashed [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Training convergence curves for (a) FOD beamforming, (b) user scheduling rate, (c) user scheduling QoS satisfaction, and (d) beam selection fine [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance of WFM-FOD against baseline methods: (a) beamforming sum rate versus [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Beam selection performance of WFM-BS against baseline methods [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Generalizability of the WFM against corresponding baselines across four tasks: (a) channel estimation, (b) beamforming, (c) user scheduling, and [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

38 extracted references · 6 linked inside Pith

  1. [1]

    A vision of 6G wireless systems: Applications, trends, technologies, and open research problems,

    W. Saad, M. Bennis, and M. Chen, “A vision of 6G wireless systems: Applications, trends, technologies, and open research problems,”IEEE Netw., vol. 34, no. 3, pp. 134–142, May 2020

    2020

  2. [2]

    Deep learning for wireless communications: An emerging interdisciplinary paradigm,

    L. Dai, R. Jiao, F. Adachi, H. V . Poor, and L. Hanzo, “Deep learning for wireless communications: An emerging interdisciplinary paradigm,” IEEE Wireless Commun., vol. 27, no. 4, pp. 133–139, Aug. 2020

    2020

  3. [3]

    Power of deep learning for channel estimation and signal detection in OFDM systems,

    H. Ye, G. Y . Li, and B.-H. Juang, “Power of deep learning for channel estimation and signal detection in OFDM systems,”IEEE Wireless Commun. Lett., vol. 7, no. 1, pp. 114–117, Feb. 2018

    2018

  4. [4]

    Deep learning in physical layer communications,

    Z. Qin, H. Ye, G. Y . Li, and B.-H. F. Juang, “Deep learning in physical layer communications,”IEEE Wireless Commun., vol. 26, no. 2, pp. 93–99, Apr. 2019

    2019

  5. [5]

    Learn to adapt to new environments from past experience and few pilot blocks,

    O. Wang, J. Gao, and G. Y . Li, “Learn to adapt to new environments from past experience and few pilot blocks,”IEEE Trans. Cogn. Commun. Netw., vol. 9, no. 2, pp. 373–385, Apr. 2023

    2023

  6. [6]

    Transfer learning for signal detection in wireless networks,

    N. Van Huynh and G. Y . Li, “Transfer learning for signal detection in wireless networks,”IEEE Wireless Commun. Lett., vol. 11, no. 11, pp. 2325–2329, Nov. 2022

    2022

  7. [7]

    Frameworks on few-shot learning with applications in wireless communication,

    O. Wang, S. Zhou, and G. Y . Li, “Frameworks on few-shot learning with applications in wireless communication,”IEEE Trans. Signal Process., vol. 73, pp. 3857–3871, 2025

    2025

  8. [8]

    Fast adaptation for deep learning-based wireless communications,

    O. Wanget al., “Fast adaptation for deep learning-based wireless communications,”IEEE Commun. Mag., vol. 63, no. 10, pp. 158–164, Oct. 2025. 15 WFM-CE Trans LS LMMSE −20 −10 0 10 NMSE (dB) (a) Channel Estimation WFM-FOD WLLM Trans WMMSE 1000 2000 3000 4000Sum Rate (bps/Hz) (b) Beamforming WFM-FOD WLLM Trans Greedy 1000 2000 3000 4000Sum Rate (bps/Hz) (c) ...

    2025

  9. [9]

    Large language models for wireless communications: From adaptation to autonomy,

    L. Lianget al., “Large language models for wireless communications: From adaptation to autonomy,”IEEE Commun. Mag., vol. 64, no. 5, pp. 140–147, May 2026

    2026

  10. [10]

    Large AI models for wireless physical layer,

    J. Guo, Y . Cui, S. Jin, and J. Zhang, “Large AI models for wireless physical layer,”IEEE Commun. Mag., vol. 64, no. 5, pp. 148–155, May 2026

    2026

  11. [11]

    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 Wireless Commun. Lett., vol. 14, no. 5, pp. 1406–1410, May 2025

    2025

  12. [12]

    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

  13. [13]

    Large language model enabled multi-task physical layer network,

    T. Zheng and L. Dai, “Large language model enabled multi-task physical layer network,”IEEE Trans. Commun., vol. 74, pp. 307–321, 2026

    2026

  14. [14]

    MUSE-FM: Multi-task environment-aware foundation model for wireless communications,

    T. Zheng, J. Guo, L. Dai, S. Jin, and J. Zhang, “MUSE-FM: Multi-task environment-aware foundation model for wireless communications,” arXiv preprint arXiv:2509.01967, 2025

    arXiv 2025

  15. [15]

    WiFo: Wireless foundation model for channel prediction,

    B. Liu, S. Gao, X. Liu, X. Cheng, and L. Yang, “WiFo: Wireless foundation model for channel prediction,”Sci. China Inf. Sci., vol. 68, no. 6, p. 162302, May 2025

    2025

  16. [16]

    WiFo-CF: Wireless foundation model for CSI feedback,

    X. Liu, S. Gao, B. Liu, X. Cheng, and L. Yang, “WiFo-CF: Wireless foundation model for CSI feedback,”arXiv preprint arXiv:2508.04068, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    A wireless foundation model for multi-task prediction,

    Y . Shenget al., “A wireless foundation model for multi-task prediction,” arXiv preprint arXiv:2507.05938, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Large wireless model: A foundation model for wireless channels,

    S. Alikhani, G. Charan, and A. Alkhateeb, “Large wireless model: A foundation model for wireless channels,”arXiv preprint arXiv:2411.08872, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    WirelessGPT: A generative pre-trained multi-task learning framework for wireless communication,

    T. Yanget al., “WirelessGPT: A generative pre-trained multi-task learning framework for wireless communication,”IEEE Netw., vol. 39, no. 5, pp. 58–65, Sep. 2025

    2025

  20. [20]

    Signal compression for wireless communication and sensing: A general ap- proach utilizing pretrained wireless foundation models,

    L. Jing, T. Yang, H. Zhang, Y . Shi, C. Zhang, and B. Zhang, “Signal compression for wireless communication and sensing: A general ap- proach utilizing pretrained wireless foundation models,”IEEE Trans. Mobile Comput., early access, 2026

    2026

  21. [21]

    An iteratively weighted MMSE approach to distributed sum-utility maximization for a MIMO interfering broadcast channel,

    Q. Shi, M. Razaviyayn, Z.-Q. Luo, and C. He, “An iteratively weighted MMSE approach to distributed sum-utility maximization for a MIMO interfering broadcast channel,”IEEE Trans. Signal Process., vol. 59, no. 9, pp. 4331–4340, Sep. 2011

    2011

  22. [22]

    On the optimality of multiantenna broad- cast scheduling using zero-forcing beamforming,

    T. Yoo and A. Goldsmith, “On the optimality of multiantenna broad- cast scheduling using zero-forcing beamforming,”IEEE J. Sel. Areas Commun., vol. 24, no. 3, pp. 528–541, Mar. 2006

    2006

  23. [23]

    On channel estimation in OFDM systems,

    J.-J. van de Beek, O. Edfors, M. Sandell, S. K. Wilson, and P. O. B¨orjesson, “On channel estimation in OFDM systems,” inProc. IEEE V eh. Technol. Conf. (VTC), vol. 2, 1995, pp. 815–819

    1995

  24. [24]

    Channel estimation and hybrid precoding for millimeter wave cellular systems,

    A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, “Channel estimation and hybrid precoding for millimeter wave cellular systems,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831–846, Oct. 2014

    2014

  25. [25]

    A deep learning framework for optimization of MISO downlink beam- forming,

    W. Xia, G. Zheng, Y . Zhu, J. Zhang, J. Wang, and A. P. Petropulu, “A deep learning framework for optimization of MISO downlink beam- forming,”IEEE Trans. Commun., vol. 68, no. 3, pp. 1866–1880, Mar. 2020

    2020

  26. [26]

    Masked autoencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked autoencoders are scalable vision learners,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 16000–16009

    2022

  27. [27]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,” inProc. Int. Conf. Learn. Represent. (ICLR), 2021

    2021

  28. [28]

    BEiT: BERT pre-training of image Transformers,

    H. Bao, L. Dong, S. Piao, and F. Wei, “BEiT: BERT pre-training of image Transformers,” inProc. Int. Conf. Learn. Represent. (ICLR), 2022

    2022

  29. [29]

    RoFormer: En- hanced Transformer with rotary position embedding,

    J. Su, M. Ahmed, Y . Lu, S. Pan, W. Bo, and Y . Liu, “RoFormer: En- hanced Transformer with rotary position embedding,”Neurocomputing, vol. 568, p. 127063, Feb. 2024

    2024

  30. [30]

    Attention is all you need,

    A. Vaswaniet al., “Attention is all you need,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2017

    2017

  31. [31]

    Extending context window of large language models via positional interpolation,

    S. Chen, S. Wong, L. Chen, and Y . Tian, “Extending context window of large language models via positional interpolation,”arXiv preprint arXiv:2306.15595, 2023

    Pith/arXiv arXiv 2023

  32. [32]

    Optimal multiuser trans- mit beamforming: A difficult problem with a simple solution structure,

    E. Bj ¨ornson, M. Bengtsson, and B. Ottersten, “Optimal multiuser trans- mit beamforming: A difficult problem with a simple solution structure,” IEEE Signal Process. Mag., vol. 31, no. 4, pp. 142–148, Jul. 2014

    2014

  33. [33]

    Fourier features let networks learn high frequency functions in low dimensional domains,

    M. Tanciket al., “Fourier features let networks learn high frequency functions in low dimensional domains,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2020, pp. 7537–7547

    2020

  34. [34]

    Categorical reparameterization with Gumbel-Softmax,

    E. Jang, S. Gu, and B. Poole, “Categorical reparameterization with Gumbel-Softmax,” inProc. Int. Conf. Learn. Represent. (ICLR), 2017

    2017

  35. [35]

    Estimating or propagating gradients through stochastic neurons for conditional computation,

    Y . Bengio, N. L ´eonard, and A. Courville, “Estimating or propagating gradients through stochastic neurons for conditional computation,”arXiv preprint arXiv:1308.3432, 2013

    Pith/arXiv arXiv 2013

  36. [36]

    Rethinking the inception architecture for computer vision,

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the inception architecture for computer vision,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 2818–2826

    2016

  37. [37]

    The Hungarian method for the assignment problem,

    H. W. Kuhn, “The Hungarian method for the assignment problem,”Nav. Res. Logist. Q., vol. 2, no. 1–2, pp. 83–97, 1955

    1955

  38. [38]

    DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,

    A. Alkhateeb, “DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,”arXiv preprint arXiv:1902.06435, 2019

    Pith/arXiv arXiv 1902