Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

A three-stage graph neural network jointly places pinching antennas, sets RIS phases, and designs beamformers, reaching near-optimal sum rate and energy efficiency with millisecond inference.

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 ↗

A three-stage GNN is proposed to jointly optimize pinching-antenna positions, RIS phase shifts, and beamforming for downlink sum-rate and energy-efficiency maximization, showing simulated gains and millisecond-level inference.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A careful, well-specified GNN pipeline for a genuinely new RIS+PASS problem, but the 'near-optimal' claim outruns the evidence — no joint-optimization baseline exists; send to review and ask for one. the 3 major comments →

arxiv 2511.20305 v2 pith:6LHSW2IC submitted 2025-11-25 cs.NI cs.AI

RIS-Assisted Downlink Pinching-Antenna Systems: GNN-Enabled Optimization Approaches

Changpeng He , Yang Lu , Yanqing Xu , Chong-Yung Chi , Arumugam Nallanathan This is my paper
classification cs.NI cs.AI
keywords pinching-antenna systemreconfigurable intelligent surfacegraph neural networkbeamformingsum-rate maximizationenergy efficiencyunsupervised learningmillisecond inference
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 reading

This paper proposes a way to jointly tune three sets of controls in a wireless downlink where antennas can slide along waveguides and a reconfigurable surface reflects signals: antenna positions, reflecting-element phases, and transmit beamforming weights. The authors' claim is that a single graph neural network, run in three sequential stages, can find near-optimal settings for both sum-rate and energy-efficiency objectives in under two milliseconds, whereas conventional iterative solvers take seconds. If true, it would make the combined pinching-antenna-plus-RIS architecture practical for real-time 6G links, and it would be the first effective learning-based solution for this scenario. The paper supports the claim with simulations showing the network generalizes to unseen user counts and that hybrid learning-plus-optimization strategies improve on the fully learned approach.

Core claim

The paper claims that a cascaded three-stage graph neural network, trained without labels by minimizing the inverse of the objective, maps user locations to PA positions, effective channels to RIS phases, and updated channels to beamforming vectors, with each stage's output made feasible by construction. The central discovery asserted is that this decomposition preserves most of the joint-optimization gain: in the fixed-antenna baseline the learned beamformers come within 0.6 percent in energy efficiency and 2.0 percent in sum rate of a convex solver, and the full RIS-plus-PA system outperforms PA-only and fixed-PA systems by double-digit percentages. The paper also asserts millisecond infer

What carries the argument

The load-bearing mechanism is the three-stage GNN cascade. Stage 1 (PAGNN) uses complex graph convolution layers to map user coordinates to pinching-antenna positions, with a sigmoid-plus-scaling layer that enforces the movable-region and minimum-spacing constraints. Stage 2 (RISGNN) maps effective channels, computed with an identity phase matrix, to unit-modulus RIS phases by normalization. Stage 3 (BeamGNN) uses graph attention and fully connected layers to produce hybrid zero-forcing/maximum-ratio-transmission coefficients and power allocations, then normalizes the power to meet the budget. The whole network is trained unsupervised by minimizing the reciprocal of the sum-rate or energy-ef

Load-bearing premise

The load-bearing premise is that near-optimal PA positions can be chosen from user locations alone, without knowing the random rapid fluctuations of the PA-RIS and RIS-user links, because the first network stage receives only user coordinates.

What would settle it

Run the same test configurations with a Stage 1 that receives the full instantaneous channel matrices, or with an alternating-optimization baseline that repositions PAs per channel realization, and compare sum rate and energy efficiency. If CSI-adaptive placement beats the location-only staging by more than the reported Strategy I-versus-II gaps, the claim that user locations alone determine near-optimal PA positions is false.

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

If this is right

  • Millisecond-level inference makes the joint design feasible for fast-varying channels where iterative solvers are too slow.
  • One trained model serves multiple network sizes: the GNN's permutation-equivariant message passing lets it handle unseen user counts with negligible performance loss.
  • Pinching-antenna placement and RIS phase control are complementary; the reported gains of the combined system over PA-only and fixed-PA baselines quantify the benefit of jointly moving antennas and shaping the reflection.
  • Hybrid learning-optimization dominates: using the learned antenna positions and RIS phases as inputs to a convex beamforming solver yields the best reported sum rate and energy efficiency, while the fully learned variant remains competitive and fastest.

Where Pith is reading between the lines

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

  • Inference: Because Stage 1 sees only user locations, the architecture implicitly bets that small-scale fading does not change where antennas should sit. A CSI-aware Stage 1 is the obvious ablation that would test this bet; the paper does not report it.
  • Inference: The same staged decomposition (geometry first, channel-dependent phases second, beamforming last) could transfer to other movable-antenna systems, such as fluid antennas or UAV-mounted arrays, where placement decisions are currently made without instantaneous fading information.
  • Inference: The reported saturation of RIS gains near L=26 elements suggests a design rule: doubling or quadrupling the surface beyond that point buys little, so deployment budgets might be better spent on more waveguides or PAs.
  • Inference: The paper compares against MLP and fixed baselines but not against a joint alternating-optimization baseline that adapts PA positions to each channel realization; adding that comparison would bound the cost of the staged decomposition.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper studies a downlink multi-user RIS-assisted pinching-antenna system (PASS) and formulates joint sum-rate (SR) and energy-efficiency (EE) maximization problems over PA positions, RIS phase shifts, and beamforming vectors. It proposes a three-stage unsupervised GNN: Stage 1 (PAGNN) maps user locations to PA positions, Stage 2 (RISGNN) maps effective channels to RIS phases, and Stage 3 (BeamGNN) uses an HZM parameterization to produce beamforming vectors. Three implementation strategies are presented: fully learning-based, GNN followed by SCA-based beamforming, and a two-stage GNN trained with a closed-form beamforming expression plus SCA. Feasibility of PA spacing, unit-modulus phases, and power constraints is enforced by custom activation and normalization layers. Numerical results report EE/SR gains over MLP, generalization across different user numbers, and sub-2 ms inference times.

Significance. The problem is timely and the proposed architecture is a natural fit for the coupled optimization variables. The main strengths are the explicit feasibility-guarantee layers, the unsupervised training that avoids labeled optimal solutions, the parameter count being independent of the number of users, and the systematic comparison among the three strategies and system configurations. If the near-optimality claim were established, the paper would make a useful contribution to the RIS/PASS literature. However, as detailed below, the current evidence does not support the near-optimality claim for the full joint problem.

major comments (3)
  1. [Sections V.A and VI.B, Table III] The central claim that the three-stage GNN yields a "near-optimal" solution to P1/P2 is not supported by the numerical evaluation. The only convex baseline, CVX (Section VI.A.2), optimizes beamforming vectors with pre-given PA positions and RIS phase shifts. The comparison in Table III therefore certifies BeamGNN for fixed PA/RIS, but not PAGNN or RISGNN. The statement in Section VI.B that the GNN "rapidly learns near-optimal PA positions and RIS phase shift matrix" has no corresponding benchmark. To make this claim load-bearing, the authors should add a joint-optimization baseline, e.g., alternating optimization over PA positions, RIS phases, and beamforming, or an exhaustive/grid-search upper bound for small M, N, L; alternatively, the near-optimal wording should be softened to a heuristic claim.
  2. [Sections IV.A and IV.B, Eqs. (7), (10), (14)] Stage 1 (PAGNN) takes only user locations as input, while the objective in (14) is evaluated on instantaneous channel realizations that contain Rician/NLoS components in the PA-RIS link (7) and RIS-user link (10); with κ = 3 dB these random components are non-negligible. The paper offers no argument that the optimal PA positions are independent of small-scale fading. If the optimal positions vary with the instantaneous realization, the staged architecture is structurally suboptimal because Stage 1 cannot adapt. The authors should provide at least one of: (i) a derivation that the PA position affects only the LoS/large-scale terms, (ii) a comparison against a CSI-adaptive PA-placement baseline, or (iii) an ablation at κ = ∞ (pure LoS) showing that small-scale fading does not affect the learned placement.
  3. [Section IV.D.2, Eq. (27)] The HZM learning restricts each beamforming direction to a convex combination of the ZF and MRT directions. The favorable comparison against CVX in the Fixed PA-only configuration shows that this restriction is acceptable when PA positions and RIS phases are fixed. However, in the full joint setting the beamforming restriction may interact adversely with the PAGNN/RISGNN outputs, and no evidence is given that the HZM family is rich enough for near-optimal joint performance. I suggest an additional comparison in which BeamGNN is replaced by an unconstrained complex-output layer, or by CVX while keeping PAGNN/RISGNN unchanged, to isolate the effect of the HZM parameterization.
minor comments (5)
  1. [Section IV.D] The opening sentence says "The input and output of RISGNN are ... baseband beamforming vectors"; this should refer to BeamGNN.
  2. [Eq. (8)] The text states that φ_{n,m} ∈ [0,2π) denotes the cosine of the AoD, but a cosine value lies in [-1,1]. Please use a distinct symbol or correct the description.
  3. [Table II and Eq. (8)] The symbol Δ is used both for the RIS element separation in Eq. (8) and for a minimum inter-PA spacing in Table II. Use distinct symbols to avoid ambiguity.
  4. [Section VI.A.2 and Table III] CVX is not a separate row in Table III. Clarify that for the Fixed PA-only configuration, Strategy II (and Strategy III) coincides with the CVX-based beamforming baseline, and that CVX does not optimize PA positions or RIS phases.
  5. [Section VII] The phrase "the first practical solution" is a strong claim. Since no comparison with existing PASS-specific learning methods (e.g., refs. [28], [29]) is provided, consider qualifying the statement or adding a positional discussion of why those methods are not applicable to the RIS-assisted setting.

Circularity Check

0 steps flagged

No significant circularity; training/evaluation on the same objective is a standard learning loop, and self-citations are background only.

full rationale

The derivation chain is not circular. The GNN is trained by unsupervised minimization of the reciprocal of the very SR/EE objectives (40)-(41) and then evaluated on the same objectives on held-out samples from the same simulator; this is a conventional self-consistent validation loop for a learned optimizer, not a definitional reduction of the output to the input. PAGNN's use of only user locations to set PA positions is a structural optimality limitation (the objective (14) contains instantaneous Rician/NLoS terms through (7) and (10)), but no equation defines the objective in terms of the GNN outputs, so this is an unvalidated optimality assumption, not circularity. The HZM beamforming parameterization (26)-(28) is adopted from [38] by citation and is not claimed to be derived from P1/P2. The paper explicitly acknowledges "the lack of optimality guarantees for the three-stage GNN" in the Introduction, consistent with treating near-optimality as an empirical claim rather than a derived identity. Multiple related-work citations overlap with the authors (e.g., [18], [27], [30], [32], [33], [36], [37]) but none is load-bearing; no uniqueness theorem is imported, and no fitted parameter is renamed as a prediction. The CVX comparison in Table III only certifies BeamGNN against a beamforming-only baseline, which weakens the near-optimality claim for PAGNN and RISGNN but does not make any step circular.

Axiom & Free-Parameter Ledger

2 free parameters · 7 axioms · 0 invented entities

The central claim rests on channel-modeling and architectural assumptions listed above. The only 'free parameters' are hand-chosen network and training hyperparameters, which are not justified by sensitivity analysis, plus the equal-power restriction in Strategy III. No new physical entities are introduced.

free parameters (2)
  • Network hyperparameters (G1..G6, hidden dim, learning rate, batch size, epochs) = G1=3; G2=3; G3=3; G4=5; G5=3; G6=5; hidden=1024; lr=1e-3; batch=1024; epochs=100
    Chosen by hand (Table I and Section VI.A.4); no sensitivity analysis is provided, so the central performance claim depends on unoptimized hyperparameters.
  • HZM equal-power coefficients (Strategy III) = p_k = lambda_k = P_max/K
    Section V.C sets equal power and dual variables for the closed-form beamforming (42); this restricts the solution space and is not justified beyond simplicity.
axioms (7)
  • domain assumption PA-user channel is LoS free-space; PA-RIS and RIS-user channels are Rician with parameters alpha=2.8, beta0=-20 dB, kappa=3 dB.
    Used in Eqs. (4)-(10). If the real environment has NLoS direct links or different statistics, the optimized solutions may not generalize.
  • domain assumption AoD from a PA to all RIS elements is identical because RIS size is negligible relative to transmission distance.
    Footnote 2 simplifies l_LoS in Eq. (8); it ignores angle variations across RIS elements.
  • domain assumption Waveguide is lossless; the pinching beamforming matrix G contains only phase delays.
    Eqs. (1)-(3) ignore waveguide attenuation and coupling; real waveguides introduce losses that would change the objective.
  • domain assumption Perfect CSI is available for computing effective channels in Eqs. (22) and (25).
    Stages 2 and 3 require composite channel knowledge; no channel estimation error model is included.
  • ad hoc to paper Optimal PA positions can be learned from user locations alone, without instantaneous CSI.
    Stage 1 input is only the user-location set (Section IV.A/IV.B). The paper provides no justification that PA placement is independent of small-scale fading.
  • ad hoc to paper The HZM parameterization w_k = sqrt(p_k) w(alpha_k) with alpha_k in [0,1] and zero-forcing/MRT directions covers near-optimal beamformers.
    Section IV.D.2 restricts beamformers to a convex combination of ZF and MRT; no optimality gap is analyzed.
  • ad hoc to paper Unsupervised training minimizing the reciprocal objectives (40)/(41) converges to a good solution.
    Section IV.F; no convergence or generalization bound is given.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of RIS-Assisted Downlink Pinching-Antenna Systems: GNN-Enabled Optimization Approaches." pith.science (2026). https://pith.science/paper/6LHSW2IC

@misc{pith2026251120305,
  author       = {Pith},
  title        = {Pith review of: RIS-Assisted Downlink Pinching-Antenna Systems: GNN-Enabled Optimization Approaches},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6LHSW2IC}},
  note         = {Machine review of arXiv:2511.20305}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper investigates a reconfigurable intelligent surface (RIS)-assisted multi-waveguide pinching-antenna (PA) system (PASS) for multi-user downlink information transmission, motivated by the unknown impact of the integration of emerging PASS and RIS on wireless communications. First, we formulate sum rate (SR) and energy efficiency (EE) maximization problems in a unified framework, subject to constraints on the movable region of PAs, total power budget, and tunable phase of RIS elements. Then, by leveraging a graph-structured topology of the RIS-assisted PASS, a novel three-stage graph neural network (GNN) is proposed, which learns PA positions based on user locations, and RIS phase shifts according to composite channel conditions at the first two stages, respectively, and finally determines beamforming vectors. Specifically, the proposed GNN is achieved through unsupervised training, together with three implementation strategies for its integration with convex optimization, thus offering trade-offs between inference time and solution optimality. Extensive numerical results are provided to validate the effectiveness of the proposed GNN, and to support its unique attributes of viable generalization capability, good performance reliability, and real-time applicability. Moreover, the impact of key parameters on RIS-assisted PASS is illustrated and analyzed.

Figures

Figures reproduced from arXiv: 2511.20305 by Arumugam Nallanathan, Changpeng He, Chong-Yung Chi, Yang Lu, Yanqing Xu.

Figure 1
Figure 1. Figure 1: Illustration of RIS-assisted PASS for downlink multi-user communi [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The structure of the proposed three-stage GNN maps the given user locations [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of three strategies. We directly minimize the reciprocal of each objective func￾tion. For Problem P1, the loss function over a mini-batch of size T is defined as LT (Θ) = 1 T X T t=1 1 PK k=1 R (t) k ({{wk}, {x P n,m}, Φ|Θ}) . (40) For Problem P2, the loss function is defined as LT (Θ) = 1 T X T t=1 PK k=1 ∥w (t) k ∥ 2 + PC PK k=1 R (t) k ({{wk}, {x P n,m}, Φ|Θ}) . (41) V. IMPLEMENTATION STRAT… view at source ↗
Figure 4
Figure 4. Figure 4: Impact of number of waveguides on EE and SR, with [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Impact of number of reflecting elements on performance with [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. Spectral- and Energy-efficient Multi-BS Multi-RIS Pinching-antenna Systems: A GNN-based Approach

    eess.SP 2026-05 unverdicted novelty 6.0

    A three-stage GNN jointly optimizes PA placement, RIS phases, beamforming and associations to maximize sum rate and energy efficiency in multi-BS multi-RIS pinching-antenna systems.

  2. A Unified Fully Reconfigurable Architecture for Wireless Powered Communication Networks

    eess.SY 2026-07 conditional novelty 5.0

    A unified WPCN architecture integrates PASS, FAS, MA, and RIS for end-to-end spatial reconfigurability across wireless energy transfer and information transmission.

  3. Pinching Antenna Systems (PASS): Enabling Reconfigurable and Controllable Wireless Channels -- A Comprehensive Survey

    cs.IT 2026-04 unverdicted novelty 2.0

    The paper provides a comprehensive review and categorization of pinching antenna systems (PASS) for objectives including network coverage, data rate, secure transmission, sensing, integrated sensing and communication,...

Reference graph

Works this paper leans on

42 extracted references · 2 linked inside Pith · cited by 3 Pith papers

  1. [1]

    On the road to 6G: visions, requirements, key technologies, and testbeds,

    C.-X. Wang et al., “On the road to 6G: visions, requirements, key technologies, and testbeds,”IEEE Commun. Surv. Tutorials, vol. 25, no. 2, pp. 905-974, Second quarter 2023

    2023

  2. [2]

    Smart radio environments empowered by reconfigurable intelligent surfaces: How it works, state of research, and the road ahead,

    M. Di Renzo, A. Zappone, M. Debbah, M.-S. Alouini, C. Yuen, J. De Rosny, and S. Tretyakov, “Smart radio environments empowered by reconfigurable intelligent surfaces: How it works, state of research, and the road ahead,”IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2450–2525, Nov. 2020

    2020

  3. [3]

    Intelligent reflecting surface-aided wireless communications: A tutorial,

    Q. Wu, S. Zhang, B. Zheng, C. You, and R. Zhang, “Intelligent reflecting surface-aided wireless communications: A tutorial,”IEEE Trans. Commun., vol. 69, no. 5, pp. 3313–3351, 2021

    2021

  4. [4]

    Pinching- antenna systems (PASS): A tutorial,

    Y . Liu, H. Jiang, X. Xu, Z. Wang, J. Guo, C. Ouyang, X. Mu, Z. Ding, A. Nallanathan, G. K. Karagiannidis, and R. Schober, “Pinching- antenna systems (PASS): A tutorial,” arXiv preprint arXiv:2508.07572, Aug. 2024

    arXiv 2024

  5. [5]

    Flexible-antenna systems: A pinching-antenna perspective,

    Z. Ding, R. Schober, and H. V . Poor, “Flexible-antenna systems: A pinching-antenna perspective,”IEEE Trans. Commun., early access, 2025

    2025

  6. [6]

    Pinching antenna systems versus reconfigurable intelligent surfaces in mmWave,

    M. Samy, H. Al-Hraishawi, M. Alsenwi, A. B. M. Adam, S. Chatzinotas, and B. Otteresten, “Pinching antenna systems versus reconfigurable intelligent surfaces in mmWave,” arXiv preprint arXiv:2506.05102, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    Wireless communications through reconfigurable intelligent surfaces,

    E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, M. Alouini, and R. Zhang, “Wireless communications through reconfigurable intelligent surfaces,”IEEE Trans. Wireless Commun., vol. 18, no. 12, pp. 6506– 6518, 2019

    2019

  8. [8]

    Reconfigurable intelligent surfaces: Principles and opportunities,

    Y . Liu, X. Liu, X. Mu, T. Hou, J. Xu, M. Di Renzo, and N. Al-Dhahir, “Reconfigurable intelligent surfaces: Principles and opportunities,”IEEE Commun. Surv. Tutorials, vol. 23, no. 3, pp. 1546–1577, 2021

    2021

  9. [9]

    Reconfigurable intelligent surfaces for energy efficiency in wireless communication,

    C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and C. Yuen, “Reconfigurable intelligent surfaces for energy efficiency in wireless communication,”IEEE Trans. Wireless Commun., vol. 18, no. 8, pp. 4157–4170, 2019

    2019

  10. [10]

    Intelligent reflecting surface assisted secrecy communication: Is artificial noise helpful or not?

    X. Guan, Q. Wu, and R. Zhang, “Intelligent reflecting surface assisted secrecy communication: Is artificial noise helpful or not?”IEEE Wireless Commun. Lett., vol. 9, no. 6, pp. 778–782, 2020

    2020

  11. [11]

    Robust and secure wireless communi- cations via intelligent reflecting surfaces,

    X. Yu, D. Xu, and R. Schober, “Robust and secure wireless communi- cations via intelligent reflecting surfaces,”IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2637–2652, 2020

    2020

  12. [12]

    Multicell MIMO communications relying on intelligent reflecting surfaces,

    C. Panet al., “Multicell MIMO communications relying on intelligent reflecting surfaces,”IEEE Trans. Wireless Commun., vol. 19, no. 8, pp. 5218–5233, 2020

    2020

  13. [13]

    Machine learning for large-scale optimization in 6G wireless networks,

    Y . Shi et al., “Machine learning for large-scale optimization in 6G wireless networks,”IEEE Commun. Surv. Tutorials, vol. 25, no. 4, pp. 2088-2132, Fourth quarter 2023

    2088

  14. [14]

    Deep learning in mobile and wireless networking: A survey,

    C. Zhang, P. Patras, and H. Haddadi, “Deep learning in mobile and wireless networking: A survey,”IEEE Commun. Surv. Tutorials, vol. 21, no. 3, pp. 2224–2287, 2019

    2019

  15. [15]

    Deep reinforcement learning-based mode selection and resource management for green fog radio access networks,

    Y . Sun, M. Peng, and S. Mao, “Deep reinforcement learning-based mode selection and resource management for green fog radio access networks,” IEEE Internet Things J., vol. 6, no. 2, pp. 1960–1971, 2019

    1960

  16. [16]

    Deep learning enabled multicast beamforming with mov- able antenna array,

    J.-M. Kang, “Deep learning enabled multicast beamforming with mov- able antenna array,”IEEE Wireless Commun. Lett., vol. 13, no. 7, pp. 1848-1852, Jul. 2024

    2024

  17. [17]

    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

  18. [18]

    Graph neural networks for wireless networks: Graph representation, architecture and evaluation,

    Y . Lu, Y . Li, R. Zhang, W. Chen, B. Ai, and D. Niyato, “Graph neural networks for wireless networks: Graph representation, architecture and evaluation,”IEEE Wireless Commun., vol. 32, no. 1, pp. 150–156, Feb. 2025

    2025

  19. [19]

    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

  20. [20]

    Zhang, T

    Y . Zhang, T. Li and Y . Zhang, ”GNN-SML: Graphic Neural Network- Based Spectrum Misuser Localization,” IEEE INFOCOM 2025 - IEEE Conference on Computer Communications, London, United Kingdom, 2025, pp. 1-10

    2025

  21. [21]

    Deep learning for energy efficient beamforming in MU-MISO networks: A GAT-based approach,

    Y . Li, Y . Lu, R. Zhang, B. Ai, and Z. Zhong, “Deep learning for energy efficient beamforming in MU-MISO networks: A GAT-based approach,” IEEE Wireless Commun. Lett., vol. 12, no. 7, pp. 1264–1268, 2023

    2023

  22. [22]

    Reconfigurable intelligent surface as- sisted multiuser MISO systems exploiting deep reinforcement learning,

    C. Huang, R. Mo, and Y . Yuen, “Reconfigurable intelligent surface as- sisted multiuser MISO systems exploiting deep reinforcement learning,” IEEE J. Sel. Areas Commun., vol. 38, no. 8, pp. 1839–1850, 2020

    2020

  23. [23]

    Learning to reflect and to beamform for intelligent reflecting surface with implicit channel estimation,

    T. Jiang, H. V . Cheng, and W. Yu, “Learning to reflect and to beamform for intelligent reflecting surface with implicit channel estimation,”IEEE J. Sel. Areas Commun., vol. 39, no. 7, pp. 1931–1945, 2021

    1931

  24. [24]

    Beam training for pinching-antenna systems (PASS),

    S. Lv, Y . Liu, and Z. Ding, “Beam training for pinching-antenna systems (PASS),”IEEE Trans. Wireless Commun., early access, 2025. 13

    2025

  25. [25]

    On the performance of uplink pinching antenna systems (PASS),

    T. Hou, Y . Liu, and A. Nallanathan, “On the performance of uplink pinching antenna systems (PASS),”IEEE Trans. Commun., early access, 2025

    2025

  26. [26]

    Pinching- antenna systems (PASS)-enabled secure wireless communications,

    G. Zhu, X. Mu, L. Guo, S. Xu, Y . Liu, and N. Al-Dhahir, “Pinching- antenna systems (PASS)-enabled secure wireless communications,” IEEE Trans. Commun., early access, 2025

    2025

  27. [27]

    Graph neural network enabled pinching antennas,

    X. Xie, Y . Lu, and Z. Ding, “Graph neural network enabled pinching antennas,”IEEE Wireless Commun. Lett., vol. 14, no. 9, pp. 2982–2986, Sept. 2025

    2025

  28. [28]

    CaMPASS-Net: A deep learning framework on capacity maximization for MIMO pinching antenna systems in IoT,

    J.-M. Kang, S. Yun, and I.-M. Kim, “CaMPASS-Net: A deep learning framework on capacity maximization for MIMO pinching antenna systems in IoT,”IEEE Internet Things J., early access, 2025

    2025

  29. [29]

    A graph neural network for learning beamforming in pinching antenna systems (PASS),

    J. Guo, Y . Liu, and A. Nallanathan, “A graph neural network for learning beamforming in pinching antenna systems (PASS),”IEEE Wireless Commun. Lett., early access, 2025

    2025

  30. [30]

    Agentic graph neural networks for wireless communications and networking towards edge general intelligence: A survey,

    Y . Lu, S. Zhang, C. Liu, R. Zhang, B. Ai, D. Niyato, W. Ni, X. Wang, and A. Jamalipour, “Agentic graph neural networks for wireless communications and networking towards edge general intelligence: A survey,” arXiv preprint arXiv:2508.08620, Aug. 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Optimal wireless resource allocation with random edge graph neural networks,

    M. Eisen and A. Ribeiro, “Optimal wireless resource allocation with random edge graph neural networks,”IEEE Trans. Signal Process., vol. 68, pp. 2977–2991, 2020

    2020

  32. [32]

    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

  33. [33]

    GNN-enabled max-min fair beamforming,

    Y . Li, Y . Lu, B. Ai, Z. Zhong, D. Niyato, and Z. Ding, “GNN-enabled max-min fair beamforming,”IEEE Trans. Veh. Technol., vol. 73, no. 8, pp. 12184–12188, 2024

    2024

  34. [34]

    A GNN-based supervised learning framework for resource allocation in wireless IoT networks,

    T. Chen, M. You, G. Zheng, and S. Lambotharan, “A GNN-based supervised learning framework for resource allocation in wireless IoT networks,”IEEE Internet Things J., vol. 9, no. 3, pp. 1712–1722, 2022

    2022

  35. [35]

    A deep learning framework for physical-layer secure beamforming,

    Z. Songet al., “A deep learning framework for physical-layer secure beamforming,”IEEE Trans. Veh. Technol., vol. 73, no. 7, pp. 9847– 9860, 2024

    2024

  36. [36]

    GNN- enabled optimization of placement and transmission design for UA V communications,

    Q. Wang, Y . Lu, W. Chen, B. Ai, Z. Zhong, and D. Niyato, “GNN- enabled optimization of placement and transmission design for UA V communications,”IEEE Trans. Veh. Technol., vol. 73, no. 11, pp. 16789– 16802, 2024

    2024

  37. [37]

    Graph neural network enabled fluid antenna systems: A two-stage approach,

    C. He, Y . Lu, W. Chen, B. Ai, K.-K. Wong, and D. Niyato, “Graph neural network enabled fluid antenna systems: A two-stage approach,” IEEE Trans. Veh. Technol., vol. 74, no. 10, pp. 16625–16629, Oct. 2025

    2025

  38. [38]

    Hybrid MRT and ZF learning for energy-efficient transmission in multi-RIS-assisted networks,

    W. Guo et al., “Hybrid MRT and ZF learning for energy-efficient transmission in multi-RIS-assisted networks,”IEEE Trans. Veh. Technol., vol. 73, no. 8, pp. 12247-12251, Aug. 2024

    2024

  39. [39]

    Deep complex networks,

    C. Trabelsi, O. Bilaniuk, Y . Zhang, D. Serdyuk, S. Subramanian, J. F. Santos, S. Mehri, N. Rostamzadeh, Y . Bengio, and C. J Pal, “Deep complex networks,” inProc. ICLR, 2018

    2018

  40. [40]

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

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

    2014

  41. [41]

    Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification,

    K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification,” in Proc. ICCV, pp. 1026-1034, 2015

    2015

  42. [42]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inProc. ICLR, pp. 1–15, Feb. 2015

    2015

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.