Pith. sign in

REVIEW 3 major objections 5 minor 40 references

BERT4beam: Large AI Model Enabled Generalized Beamforming Optimization

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A BERT-style transformer, fed with channel data as tokens, produces near-optimal beamforming across multiple wireless tasks and system scales.

desk verdict A genuinely useful transformer-based beamforming architecture whose headline claim should be 'matches SCA' rather than 'near-optimal'; worth reviewing, but needs code/data and an independent benchmark. read the letter →

arxiv 2509.11056 v2 pith:H73PORFR submitted 2025-09-14 eess.SY cs.LGcs.SY

Yuhang Li , Yang Lu , Wei Chen , Bo Ai , Zhiguo Ding This is my paper · ORCID
classification eess.SYcs.LGcs.SY
keywords beamformingoptimizationBERT4beamtokenizationMU-MISOtransformerpre-trainingfine-tuningenergyefficiency
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

The paper tries to establish that a large pre-trained transformer, built on the BERT architecture but designed for wireless channel data, can solve beamforming optimization problems for a variety of system objectives and scales. It introduces two models: BERT for single tasks such as sum-rate, min-rate, or energy-efficiency maximization, and UBERT, a multi-task version with finer-grained tokenization that can switch tasks without changing architecture. The sympathetic reader cares because today's AI beamforming methods are task-specific networks, whereas this work suggests a single foundation model could be pre-trained once and adapted cheaply to many deployments. The evidence reported is that the models stay within a few percent of the conventional SCA solver on several tasks, generalize to unseen user counts and antenna configurations, and outperform MLP, CNN, GCN, GAT, and GPT baselines.

What carries the argument

The central mechanism is the bidirectional multi-head attention (MHA) inside transformer encoder blocks operating on CSI tokens. Each CSI token concatenates the real and imaginary parts of a user's channel vector; the model stacks 12 transformer encoder blocks with 16 attention heads and an embedding dimension of 1024. In UBERT, an antenna encoding block applies cosine positional encoding and additive attention to element-wise tokens, and a trainable task-embedding token lets the model distinguish among sum-rate, min-rate, and energy-efficiency tasks. A parameter-free generalizable power adapter (GPA) projects the output onto the power constraint, which lets the model adapt to different powe

What would settle it

Take the same Rayleigh test samples and compare BERT4beam's beamforming vectors against solutions from an independent, high-accuracy method—for example, global optimization via branch-and-bound or a semidefinite relaxation with a tighter bound. If the independent solver consistently yields higher utilities than SCA, the reported 97–99% ratios would not represent true near-optimality; conversely, if the independent solver confirms SCA's values, the claim stands.

Watch

Extended reading notes

Core claim

The central claim is that beamforming can be treated as a token-level sequence learning task: each user's channel vector is tokenized into real and imaginary parts, and a bidirectional transformer encoder learns a direct mapping from the token sequence to feasible beamforming vectors that maximize a chosen system utility. The single-task BERT model shares parameters across users, so it generalizes to different numbers of users without retraining, and it can adapt to different utilities and antenna counts through light fine-tuning. The UBERT model uses element-wise tokenization, adding an antenna encoding block and a trainable task embedding token, which lets one model handle sum-rate, min-ra

Load-bearing premise

The load-bearing premise is that the SCA solver's output is a valid measure of optimality: SCA generates the training labels and also the reference values in the reported performance percentages, so if SCA is suboptimal or fails to converge on some channel realizations, the 'near-optimal' claims could be an artifact of comparing the model to a weak solver.

Editorial extensions

If this is right

  • A single pre-trained beamforming model could be deployed across cells with different user counts and antenna arrays, removing the need to retrain a network for each configuration.
  • Real-time inference could approach the quality of iterative convex optimization while running in one forward pass, which matters for dynamic 6G environments.
  • Multi-task pre-training with a task embedding could let operators switch a model between sum-rate, min-rate, and energy-efficiency objectives without changing weights or architecture.
  • Few-shot fine-tuning with 200–1,000 labeled or unlabeled samples could adapt the model to new scenarios, cutting the data collection burden.
  • The apparent advantage of bidirectional over unidirectional attention suggests that future wireless foundation models should use bidirectional encoders rather than causal decoders.

Reading between the lines

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

  • An implication not tested in the paper: the same tokenize-transformer-finetune recipe could be applied to other continuous wireless optimization problems, such as power allocation, scheduling, or waveform design, if those problems can be mapped to token sequences.
  • The reported 1% gain over SCA in energy efficiency may mean the trained model occasionally finds better local solutions than the solver, but only if the SCA labels are imperfect; the paper stops short of claiming a better optimum.
  • The generalization results were obtained on Rayleigh fading channels; a testable extension would be to run the same models on spatially correlated or non-Rayleigh channels to see whether the attention mechanism or the dataset statistics drive the generalization.
  • Because SCA supplies both the training labels and the evaluation metric, an independent benchmark against a different solver or tighter bound would reveal how much of the 'near-optimal' gap is real and how much is inherited from the reference.
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 / 5 minor

Summary. The paper proposes BERT4beam, a BERT-style transformer framework for MU-MISO beamforming optimization. The problem is cast as token-level sequence learning: CSI is tokenized per user (BERT) or per antenna-user element (UBERT), fed through transformer encoder blocks, and mapped to feasible beamforming vectors via a power adapter. Two approaches are presented: a single-task BERT that can be fine-tuned to new utilities/antenna configurations, and a multi-task UBERT with element-wise tokenization, antenna encoding, task embedding, and a multi-task loss with uniform task sampling. Experiments compare against SCA, MLP, CNN, GCN, GAT, and GPT across sum-rate, min-rate, and energy-efficiency objectives, with varying power budgets, user numbers, antenna counts, CSI errors, and few-shot fine-tuning. The main empirical claim is that BERT and UBERT achieve near-optimal performance, within 1–3% of SCA, and outperform existing AI baselines.

Significance. If the near-optimality claim were properly supported, this would be a useful contribution: it demonstrates a single transformer architecture handling multiple beamforming objectives and system scales, with ablations showing the importance of the proposed tokenization, task embedding, position encoding, and uniform-sampling multi-task training. The parameter-free GPA is a clean design for scale generalization, and the UBERT architecture's scale-invariant input/output is appealing. The paper also compares against several baselines and includes generalization and fine-tuning experiments. However, the central quantitative claim is currently calibrated only against the SCA solver that generated the training labels, and the experiments lack error bars and seed variation. These are load-bearing weaknesses, not mere presentation issues.

major comments (3)
  1. [§VI.A.6, §VI.A.2, Eq. (17), Eq. (31)] The evaluation metric (31) normalizes the model utility by U_cvx, and Section VI.A.6 explicitly states that “SCA also serves as the method to generate U_cvx in (31).” The same SCA solver produces the W_cvx labels used in the pre-training loss (17). Hence the reported percentages measure agreement with SCA, not closeness to the true optimum of the nonconvex problem (3). This directly undermines the abstract's “near-optimal performance” claim. The fact that Table I reports BERT EE = 101.24% and Table III reports BERT EE up to 100.56% shows the model can exceed this local solver, but that does not establish optimality. The authors must either compare against an independent optimality upper bound (e.g., exhaustive search or a global-solver bound on small instances) or rephrase all claims as “close to SCA.”
  2. [§VI.A.3, Tables I–VIII] No error bars, confidence intervals, or multiple-seed results are reported anywhere in the numerical section. The key differences that support the claims—BERT within 1–2% of SCA, UBERT slightly better or worse than BERT, ablation gaps—are single runs. For stochastic neural-network training, 1–3% gaps can be within random variation. The authors should rerun each configuration with at least five seeds and report mean ± std, and ideally a significance test for the claimed improvements over baselines. This is essential because the paper's main quantitative conclusions are based on these small margins.
  3. [§VI.A.2, §III.A] The dataset is generated “using an optimization algorithm with a convergence accuracy of 10^-4.” The problem (3), especially the fractional EE objective and the min-rate utility, is nonconvex, and SCA converges only to a stationary point. A tolerance of 10^-4 on the solver iterates does not bound the gap to global optimality. Consequently, the labels and the normalization baseline may themselves be inconsistent or suboptimal across the 80,000 samples. The paper should validate the SCA solutions on at least one small configuration against a global optimality bound, or discuss the expected suboptimality of SCA for these problems, before using SCA as the sole oracle for both training and evaluation.
minor comments (5)
  1. [§VI.A.2 / §III.A] The introduction says “Section VI concludes the paper,” but the conclusion is actually Section VII. Please fix the cross-reference.
  2. [Eq. (10)] After the first TEB sub-block, T_fir should be in R^{K×F}, not R^{K×N_T}, since the residual connection preserves the embedding dimension F. This appears to be a typo.
  3. [Eq. (28)] The output layer of UBERT uses “GAP” in Eq. (28), while the paper consistently defines “GPA” for the generalizable power adapter. Please unify the notation.
  4. [§VI.A.6 / §VI.B.1] The text says “GPT... can only rely on preceding context, making it suitable for NLP token prediction tasks but not for beamforming design.” This is a reasonable intuition, but the paper does not analyze why bidirectional attention is necessary for beamforming beyond the empirical result. A brief discussion would strengthen the motivation.
  5. [§V.C / Algorithm 1] Algorithm 1 uses T for the multi-task set, while T also denotes the CSI token matrix elsewhere. Rename one of them to avoid confusion.

Circularity Check

1 steps flagged · score 5.0 of 10

SCA is both the label generator and the optimality benchmark; 'near-optimal' is an interpretation of the SCA-relative metric.

  1. fitted input called prediction [Section VI.A.2, VI.A.6, Eq. (17), Eq. (31), Section VI.B.1, Abstract]
    "Using an optimization algorithm with a convergence accuracy of 10^-4, 80,000 labeled samples are generated for each power level... SCA also serves as the method to generate U_cvx in (31). ... The BERT has a performance loss of no more than 2% compared to SCA across three beamforming design tasks, and even outperforms SCA by 1% in the EE task."

    The training loss in Eq. (17) uses W_cvx (SCA-generated beamforming) as a supervised target, and the UBERT loss in Eq. (30) likewise uses W_Ucvx from traditional optimization. The evaluation metric in Eq. (31) normalizes the neural-network utility by U_cvx, and Section VI.A.6 explicitly states that SCA generates U_cvx. Thus the reported 'performance' measures closeness to SCA, the same solver that produced the training labels. The abstract's 'near-optimal performance' and Section VI.B.1's 'performance loss no more than 2% compared to SCA' are therefore statements about agreement with the label generator, not about distance to the true optimum of the nonconvex problem (3). Since SCA is a local method with no optimality certificate, the central 'near-optimal' claim reduces to 'the model mimi

full rationale

The paper's architectural contributions and its comparisons against other AI models (GAT, CNN, MLP, etc.) are independent and not circular. However, the headline claim of 'near-optimal performance' is calibrated entirely against the SCA solver that also generated the supervised training labels. Section VI.A.2 states that 80,000 labeled samples per power level were generated with a convergence accuracy of 10^-4; Section VI.A.6 states that SCA also generates U_cvx in the performance metric (31); and Eq. (17) trains BERT to match W_cvx from the same source. Consequently, the reported percentages (e.g., within 2% of SCA, or 101% EE) are ratios to SCA, not to a certified global optimum. This makes the 'near-optimal' language a renaming of 'close to SCA,' which is the fitted input. There is no indication of load-bearing self-citations, imported uniqueness theorems, or ansatz smuggling. The score reflects this partial circularity in the central optimality claim, while acknowledging that the model's generalization across scales and tasks and its superiority over other AI baselines have independent content.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. The task embedding token and antenna encoding are learned components of the neural network, not independent postulates with falsifiable handles outside the model.

free parameters (2)
  • Loss balancing weights lambda1, lambda2 in Eq. (17)
    Chosen by hand; values not reported in the paper. They balance cosine similarity to SCA labels against the task utility, directly shaping the training objective.
  • Model architecture hyperparameters (embedding dim, number of TEBs, attention heads) = 1024, 12, 16
    Standard transformer hyperparameters chosen by hand in Section VI.A.4; the paper does not ablate or justify them beyond convention.
assumptions (4)
  • domain assumption Rayleigh fading with i.i.d. CN(0,1) channels is representative of deployment scenarios
    Section VI.A.2 generates all channels this way; the generalization claims are only validated within this distribution.
  • domain assumption SCA with convergence accuracy 10^-4 produces near-globally-optimal solutions that can serve as both training labels and evaluation ground truth
    Section VI.A.2 and Section VI.A.6; the near-optimal-performance claim is measured against SCA, so the entire evaluation depends on SCA's optimality.
  • ad hoc to paper Bidirectional attention captures the interference structure needed for beamforming across tasks
    Remark 1 and Remark 2 assert that the MHA mechanism learns underlying representations generalizable across utilities and antenna configurations; this is the conceptual justification for the framework and is not proven.
  • standard math The GPA power normalization preserves the beam direction adequately
    Eq. (15) scales the beamforming matrix exactly to the power budget, so direction is preserved; this is a standard normalization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BERT4beam: Large AI Model Enabled Generalized Beamforming Optimization." pith.science (2026). https://pith.science/paper/H73PORFR

@misc{pith2026250911056,
  author       = {Pith},
  title        = {Pith review of: BERT4beam: Large AI Model Enabled Generalized Beamforming Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H73PORFR}},
  note         = {Machine review of arXiv:2509.11056}
}
read the original abstract

Artificial intelligence (AI) is anticipated to emerge as a pivotal enabler for the forthcoming sixth-generation (6G) wireless communication systems. However, current research efforts regarding large AI models for wireless communications primarily focus on fine-tuning pre-trained large language models (LLMs) for specific tasks. This paper investigates the large-scale AI model designed for beamforming optimization to adapt and generalize to diverse tasks defined by system utilities and scales. We propose a novel framework based on bidirectional encoder representations from transformers (BERT), termed BERT4beam. We aim to formulate the beamforming optimization problem as a token-level sequence learning task, perform tokenization of the channel state information, construct the BERT model, and conduct task-specific pre-training and fine-tuning strategies. Based on the framework, we propose two BERT-based approaches for single-task and multi-task beamforming optimization, respectively. Both approaches are generalizable for varying user scales. Moreover, the former can adapt to varying system utilities and antenna configurations by re-configuring the input and output module of the BERT model, while the latter, termed UBERT, can directly generalize to diverse tasks, due to a finer-grained tokenization strategy. Extensive simulation results demonstrate that the two proposed approaches can achieve near-optimal performance and outperform existing AI models across various beamforming optimization tasks, showcasing strong adaptability and generalizability.

Figures

Figures reproduced from arXiv: 2509.11056 by the authors.

Figure 1
Figure 1. The architecture of BERT (with the input being CSI token matrix and the output being feasible beamforming [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The architecture of the UBERT (with the input being antenna token matrix and the output being feasible [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Performance across different system utilities. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance under different CSI error level. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Fine-tuning performance under limited data. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Effectiveness of uniform sampling in multi-task [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 8 linked inside Pith

  1. [1]

    6G wireless networks: Vision, requirements, architecture, and key technologies,

    Z. Zhang, Y . Xiao, Z. Ma, M. Xiao, Z. Ding, X. Lei, G. K. Karagiannidis, and P. Fan, “6G wireless networks: Vision, requirements, architecture, and key technologies,”IEEE Veh. Technol. Mag., vol. 14, no. 3, pp. 28–41, 2019

    2019

  2. [2]

    A survey of recent advances in optimization methods for wireless communications,

    Y . Liu, T. Chang, M. Hong, Z. Wu, A. ManCho So, E. A. Jorswieck, and W. Yu, “A survey of recent advances in optimization methods for wireless communications,”IEEE J. Sel. Areas Commun., vol. 42, no. 11, pp. 2992–3031, 2024

    2024

  3. [3]

    Graph neural networks for wireless communications: From theory to practice,

    Y . Shen, J. Zhang, S. H. Song, and K. B. Letaief, “Graph neural networks for wireless communications: From theory to practice,”IEEE Trans. Wireless Commun., vol. 22, no. 5, pp. 3554–3569, 2023

    2023

  4. [4]

    The roadmap to 6G: AI empowered wireless networks,

    K. B. Letaief, W. Chen, Y . Shi, J. Zhang, and Y .-J. A. Zhang, “The roadmap to 6G: AI empowered wireless networks,”IEEE Commun. Mag., vol. 57, no. 8, pp. 84–90, 2019

    2019

  5. [5]

    Large multi-modal models (LMMs) as universal foundation models for AI-native wireless systems,

    S. Xu, C. Kurisummoottil Thomas, O. Hashash, N. Muralidhar, W. Saad, and N. Ramakrishnan, “Large multi-modal models (LMMs) as universal foundation models for AI-native wireless systems,”IEEE Network, vol. 38, no. 5, pp. 10–20, 2024

    2024

  6. [6]

    Enhancing deep reinforcement learning: A tutorial on generative diffusion models in network optimization,

    H. Du, R. Zhang, Y . Liu, J. Wang, Y . Lin, Z. Li, D. Niyato, J. Kang, Z. Xiong, S. Cui, B. Ai, H. Zhou, and D. I. Kim, “Enhancing deep reinforcement learning: A tutorial on generative diffusion models in network optimization,”IEEE Commun. Surv. Tutorials, vol. 26, no. 4, pp. 2611–2646, 2024

    2024

  7. [7]

    Interactive AI with retrieval-augmented generation for next generation networking,

    R. Zhang, H. Du, Y . Liu, D. Niyato, J. Kang, S. Sun, X. Shen, and H. V . Poor, “Interactive AI with retrieval-augmented generation for next generation networking,”IEEE Network, vol. 38, no. 6, pp. 414–424, 2024

    2024

  8. [8]

    Learning to optimize: Training deep neural networks for interference management,

    H. Sun, X. Chen, Q. Shi, M. Hong, X. Fu, and N. D. Sidiropoulos, “Learning to optimize: Training deep neural networks for interference management,”IEEE Trans. Signal Process., vol. 66, no. 20, pp. 5438– 5453, 2018

    2018

Show all 40 references
  1. [9]

    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,” inProc. Neural Inf. Process. Syst., vol. 30, 2017, pp. 5998–6008

    2017

  2. [10]

    BERT: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understanding,” inProc. Conf. North Amer. Chapter Assoc. Comput. Linguistics Hum. Lang. Technol., 2019, pp. 4171–4186

    2019

  3. [11]

    Im- proving language understanding by generative pre-training,

    A. Radford, K. Narasimhan, T. Salimans, and I. Sutskever, “Im- proving language understanding by generative pre-training,”OpenAI Blog, vol. 1, no. 8, 2018, available at https://openai.com/research/ language-unsupervised

    2018

  4. [12]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample, “Llama: Open and efficient foundation language models,” 2023. [Online]. Available: https: //arxiv.org/abs/2...

    2023 arXiv

  5. [13]

    Toward intelligent communications: Large model empowered semantic communications,

    H. Xie, Z. Qin, X. Tao, and Z. Han, “Toward intelligent communications: Large model empowered semantic communications,”IEEE Commun. Mag., vol. 63, no. 1, pp. 69–75, 2025

    2025

  6. [14]

    Large AI model-based semantic communications,

    F. Jiang, Y . Peng, L. Dong, K. Wang, K. Yang, C. Pan, and X. You, “Large AI model-based semantic communications,”IEEE Wireless Com- mun., vol. 31, no. 3, pp. 68–75, 2024

    2024

  7. [15]

    Large language models empowered autonomous edge AI for connected intelligence,

    Y . Shen, J. Shao, X. Zhang, Z. Lin, H. Pan, D. Li, J. Zhang, and K. B. Letaief, “Large language models empowered autonomous edge AI for connected intelligence,”IEEE Commun. Mag., vol. 62, no. 10, pp. 140– 146, 2024

    2024

  8. [16]

    Toward democratized generative AI in next-generation mobile edge networks,

    R. Zhang, J. He, X. Luo, D. Niyato, J. Kang, Z. Xiong, Y . Li, and B. Sikdar, “Toward democratized generative AI in next-generation mobile edge networks,”IEEE Network, pp. 1–1, 2025

    2025

  9. [17]

    Leveraging large language models for integrated satellite-aerial-terrestrial networks: Recent advances and future directions,

    S. Javaid, R. A. Khalil, N. Saeed, B. He, and M.-S. Alouini, “Leveraging large language models for integrated satellite-aerial-terrestrial networks: Recent advances and future directions,”IEEE Open J. Commun. Soc., vol. 6, pp. 399–432, 2025

    2025

  10. [18]

    AI-native interconnect framework for integration of large language model technologies in 6G systems,

    S. Tarkoma, R. Morabito, and J. Sauvola, “AI-native interconnect framework for integration of large language model technologies in 6G systems,” 2023. [Online]. Available: https://arxiv.org/abs/2311.05842

    2023 arXiv

  11. [19]

    Generative AI agents with large language model for satellite networks via a mixture of experts transmission,

    R. Zhang, H. Du, Y . Liu, D. Niyato, J. Kang, Z. Xiong, A. Jamalipour, and D. In Kim, “Generative AI agents with large language model for satellite networks via a mixture of experts transmission,”IEEE J. Sel. Areas Commun., vol. 42, no. 12, pp. 3581–3596, 2024

    2024

  12. [20]

    Learning to optimize: Training deep neural networks for wireless resource management,

    H. Sun, X. Chen, Q. Shi, M. Hong, X. Fu, and N. D. Sidiropoulos, “Learning to optimize: Training deep neural networks for wireless resource management,” inProc. IEEE Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC), 2017, pp. 1–6

    2017

  13. [21]

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

    H. He, C. K. Wen, S. Jin, and G. Y . Li, “Deep learning-based channel estimation for beamspace mmWave massive MIMO systems,”IEEE Wireless Commun. Lett., vol. 7, no. 5, pp. 852–855, 2018

    2018

  14. [22]

    Joint link scheduling and resource allocation for hierarchical asynchronous deep mutual learning system,

    T. Wang, S. Liu, J. Yuan, X. Chen, C. Wu, and R. Yin, “Joint link scheduling and resource allocation for hierarchical asynchronous deep mutual learning system,” inProc. IEEE Global Commun. Conf. (GLOBECOM), 2024, pp. 2629–2634

    2024

  15. [23]

    Deep learning methods for universal MISO beamforming,

    J. Kim, H. Lee, S.-E. Hong, and S.-H. Park, “Deep learning methods for universal MISO beamforming,”IEEE Wireless Commun. Lett., vol. 9, no. 11, pp. 1894–1898, 2020

    2020

  16. [24]

    Fast beamforming design via deep learning,

    H. Huang, Y . Peng, J. Yang, W. Xia, and G. Gui, “Fast beamforming design via deep learning,”IEEE Trans. Veh. Technol., vol. 69, no. 1, pp. 1065–1069, 2020

    2020

  17. [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, 2020

    2020

  18. [26]

    Learning robust beamforming for MISO downlink systems,

    J. Kim, H. Lee, and S.-H. Park, “Learning robust beamforming for MISO downlink systems,”IEEE Commun. Lett., vol. 25, no. 6, pp. 1916–1920, 2021

    1916

  19. [27]

    Graph neural networks for scalable radio resource management: Architecture design and theoretical analysis,

    Y . Shen, Y . Shi, J. Zhang, and K. B. Letaief, “Graph neural networks for scalable radio resource management: Architecture design and theoretical analysis,”IEEE J. Sel. Areas Commun., vol. 39, no. 1, pp. 101–115, 2021

    2021

  20. [28]

    Graph neural network based beamforming in D2D wireless networks,

    T. Chen, M. You, G. Zheng, and S. Lambotharan, “Graph neural network based beamforming in D2D wireless networks,” inProc. WSA 25th Int. ITG Workshop Smart Antennas, 2021, pp. 1–5

    2021

  21. [29]

    GNN-based beamforming for sum-rate maximization in MU-MISO networks,

    Y . Li, Y . Lu, B. Ai, O. A. Dobre, Z. Ding, and D. Niyato, “GNN-based beamforming for sum-rate maximization in MU-MISO networks,”IEEE Trans. Wireless Commun., vol. 23, no. 8, pp. 9251–9264, 2024

    2024

  22. [30]

    Scalable power control/beamforming in heterogeneous wireless networks with graph neural networks,

    X. Zhang, H. Zhao, J. Xiong, X. Liu, L. Zhou, and J. Wei, “Scalable power control/beamforming in heterogeneous wireless networks with graph neural networks,” inProc. IEEE Global Commun. Conf. (GLOBE- COM), 2021, pp. 01–06

    2021

  23. [31]

    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,” 2025. [Online]. Available: https://arxiv.org/abs/2408.08707

    2025 arXiv

  24. [32]

    LLM4WM: Adapting LLM for wireless multi-tasking,

    X. Liu, S. Gao, B. Liu, X. Cheng, and L. Yang, “LLM4WM: Adapting LLM for wireless multi-tasking,”IEEE Trans. Mach. Learn. Commun. Networking, vol. 3, pp. 835–847, 2025

    2025

  25. [33]

    BeamLLM: Vision-empowered mmWave beam prediction with large language models,

    C. Zheng, J. He, G. Cai, Z. Yu, and C. G. Kang, “BeamLLM: Vision-empowered mmWave beam prediction with large language models,” 2025. [Online]. Available: https://arxiv.org/abs/2503.10432

    2025 arXiv

  26. [34]

    Large language model enabled multi- task physical layer network,

    T. Zheng and L. Dai, “Large language model enabled multi- task physical layer network,” 2025. [Online]. Available: https: //arxiv.org/abs/2412.20772

    2025 arXiv

  27. [35]

    LLM-empowered resource allocation in wireless communications systems,

    W. Lee and J. Park, “LLM-empowered resource allocation in wireless communications systems,” 2024. [Online]. Available: https: //arxiv.org/abs/2408.02944

    2024 arXiv

  28. [36]

    Prompt-enabled large AI models for CSI feedback,

    J. Guo, Y . Cui, C.-K. Wen, and S. Jin, “Prompt-enabled large AI models for CSI feedback,” 2025. [Online]. Available: https: //arxiv.org/abs/2501.10629

    2025 arXiv

  29. [37]

    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. Net- works, vol. 9, no. 2, pp. 113–125, 2024

    2024

  30. [38]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” 2016. [Online]. Available: https://arxiv.org/abs/1607.06450

    2016 arXiv

  31. [39]

    Fast and accurate deep network learning by exponential linear units (ELUs),

    D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and accurate deep network learning by exponential linear units (ELUs),” inProc. Int. Conf. Learn. Represent., 2016, pp. 1–14

    2016

  32. [40]

    Resource allocation for multi-user downlink MISO OFDMA-URLLC systems,

    W. R. Ghanem, V . Jamali, Y . Sun, and R. Schober, “Resource allocation for multi-user downlink MISO OFDMA-URLLC systems,”IEEE Trans. Commun., vol. 68, no. 11, pp. 7184–7200, 2020

    2020

Pith tools

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