Pith. sign in

REVIEW 5 major objections 6 minor 33 references

JEPA-CFM: A Joint Embedding Predictive Architecture-based Channel Foundation Model for Robust Fluid Antenna Systems

T0 review · 5 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A latent-predicting channel model beats masked-autoencoder reconstruction for fluid antennas, enabling sparse-CSI extrapolation and sub-3 m positioning.

desk verdict A plausible JEPA-plus-MAE recipe for FAS channel foundation models, but the sparse-CSI evaluation protocol is internally inconsistent and the positioning claim is overstated; the idea deserves a serious referee, not current acceptance. read the letter →

arxiv 2607.20202 v1 pith:HIHGJIIZ submitted 2026-07-22 eess.SP

Yuan Gao , Yiming Liu , Jun Jiang , Jianbo Du , Shunqing Zhang , Xiaoli Chu , Kai-Kit Wong This is my paper · ORCID
classification eess.SP
keywords channelfoundationmodeljointembeddingpredictivearchitecturefluidantennasystemsextrapolationwirelesspositioningmaskedautoencoderself-supervisedlearningsparseCSI
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 sets out to show that a channel foundation model for fluid antenna systems should learn by predicting high-level latent structures of masked channel segments rather than reconstructing raw channel coefficients. It pre-trains a transformer encoder with a three-part objective: a masked autoencoder reconstruction loss, a JEPA latent prediction loss, and a sliced isotropic Gaussian regularization that prevents representation collapse. After pre-training, the encoder is frozen and lightweight heads are attached for channel extrapolation and user positioning. The reported results indicate the approach substantially outperforms a pure masked autoencoder for extrapolation when only 5–15% of ports are known, stays accurate at low SNR, and achieves sub-3 m positioning error with only 25% known CSI.

What carries the argument

The load-bearing component is the multi-term pre-training objective L_pre = γ·L_MAE + λ_JEPA·L_JEPA + λ_SIGReg·L_SIGReg. L_JEPA is the mean squared error between predicted latent embeddings of masked channel patches and target embeddings obtained from a full-view, gradient-detached pass of the same encoder; L_MAE is the standard masked reconstruction error over raw CSI coefficients; L_SIGReg measures the distance between the empirical characteristic function of projected embeddings and that of a standard normal, over random slice directions. The encoder is a Vision Transformer that processes spatial tokens of the CSI grid, and the predictor is a small transformer that recombines visible and

What would settle it

Evaluate the same architecture at mask ratios and SNR levels far outside the pre-training distribution (e.g., 1% and 50% known CSI, or SNR below 0 dB) and also test on a different channel model. If the extrapolation error collapses or falls behind the pure MAE baseline outside the tested range, the claimed robustness is falsified. A simpler check: compare the frozen-encoder positioning head against a model fine-tuned end-to-end at each test sparsity; if fine-tuning is needed, the 'frozen' claim is weakened.

Watch

Extended reading notes

Core claim

JEPA-CFM is built on the joint embedding predictive architecture: instead of reconstructing each raw complex channel coefficient of masked antenna ports, the model learns to predict the latent embeddings of those missing regions. The target embeddings come from a full-view, gradient-detached pass through the same encoder, and a predictor network combines visible tokens with learned mask tokens to forecast them in a compact space. The pre-training objective combines this JEPA loss with a masked autoencoder reconstruction loss and a SIGReg term that pushes the projected latent distribution toward a standard Gaussian. This combination, the paper argues, prevents representation collapse and make

Load-bearing premise

The paper reports that a model pre-trained with one fixed random mask ratio (75% masked) performs well when tested at 5–25% known CSI and varying SNR without any adaptation; the central claim depends on this distributional transfer holding, which is not explicitly validated by retraining or domain adaptation experiments.

Editorial extensions

If this is right

  • A single frozen encoder can support channel extrapolation and positioning with lightweight heads, reducing per-task training data and retraining cost.
  • Channel extrapolation remains accurate at extreme sparsity (5–15% known CSI), making very low pilot overhead feasible for fluid antenna systems.
  • Performance degrades gracefully with noise, holding up even at 0 dB SNR, which matters for mobile deployments.
  • Positioning to sub-3 m accuracy from only 25% of observed ports shows channel fingerprints carry strong geometric information.

Reading between the lines

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

  • If latent prediction truly filters noise, the same encoder could also improve downstream tasks like port selection or beam prediction, where noisy raw channels are the norm; this is a natural next experiment.
  • The paper evaluates only one urban ray-traced scenario; testing on other propagation environments or on measured channels would reveal whether the representation is site-general or needs site-specific pre-training.
  • Since pre-training used a fixed 75% mask ratio, a curriculum over mask ratios could push the gains even further into the 5% known-CSI regime; the current setup may understate what the architecture can do.
  • The positioning head sees only the masked-view encoder output; feeding it the full-view or reconstructed channel might improve accuracy further, at the cost of the foundation model's two-head decoupling.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes JEPA-CFM, a joint-embedding predictive architecture-based channel foundation model for fluid antenna systems (FAS). The model pre-trains an encoder with three loss terms: a masked-autoencoder reconstruction loss, a JEPA latent-prediction loss, and a sliced isotropic Gaussian regularization (SIGReg) term, using DeepMIMO O1 28 ray-tracing data with a 64×64 FAS port grid. After pre-training, the encoder is frozen; a lightweight decoder is attached for channel extrapolation and an MLP head for user positioning. The central claims are that JEPA-CFM substantially outperforms a conventional masked-autoencoder baseline under severe CSI sparsity (5%–15% known ports) and across SNR levels (0, 10, 20 dB), and that it achieves sub-3 m positioning error with only 25% known CSI.

Significance. If the claims are substantiated, the paper offers a plausible route toward a unified, self-supervised FAS representation that is data-efficient and noise-robust, with a single frozen encoder supporting multiple downstream tasks. The use of a standard ray-tracing dataset, the comparison against a vanilla MAE baseline, and the authors' explicit admission that the noiseless positioning advantage is statistically negligible (Sec. V.B.2) are commendable. The main value of the work is the sparse-CSI extrapolation evidence in Figs. 7–8. However, the evaluation protocol has load-bearing gaps: the mask/token bookkeeping is internally inconsistent, test-time mask ratios and SNRs are not matched to the pre-training configuration, and no ablations isolate the contributions of the JEPA and SIGReg terms. The reported results therefore do not yet firmly establish the paper’s headline claims.

major comments (5)
  1. [§IV.A.2–A.3, Eq. (23)–(25), Table I] The token and mask arithmetic is inconsistent and makes the architecture dimensionally ill-defined. With N_S=4096 and M_r=0.75, the formula N_o = floor(N_S(1−M_r)) gives N_o=1024. The paper then defines K_mask = K_p − N_o = 768, which is arithmetically impossible (1024−1024=0) and leaves K_p=1024 unexplained: a 64×64 port grid with 4096 ports cannot be reduced to 1024 patch tokens without a patchification step that is never described. Moreover, the JEPA predictor positional embedding E_pred,pos is fixed to shape K_mask×d_model, so the predictor is dimensioned for exactly one mask configuration. Figures 7–10 sweep 5–25% known CSI, which changes N_o (e.g., N_o≈204 at 5%) and hence K_mask; no resizing, padding, or retraining procedure is described. If the predictor is bypassed at test time, the reported advantage over the MAE baseline does not test the JEPA contribution; if it is not bypass
  2. [§V.B, Table I, Figs. 7–10] The test-time evaluation protocol is unspecified. Pre-training is performed only with a random mask ratio M_r=0.75 (Table I), and the downstream positioning fine-tuning section states that 'the mask ratio remains M_r=0.75.' Yet Figs. 7–10 report performance at 5%, 10%, 15%, 20%, and 25% known CSI, and Figs. 8 and 10 add noise at 0, 10, and 20 dB. The paper does not state how these conditions are produced at test time—whether masks are redrawn with a different ratio, whether additional tokens are discarded, whether the model is fine-tuned separately for each regime, or whether the same frozen model is evaluated under a distribution shift. As a result, the claimed generalization across sparsity levels and SNRs is not reproducible from the text.
  3. [§V.B.2, Fig. 9, Abstract, Conclusions] The positioning claim is overstated relative to the evidence. The paper’s own Sec. V.B.2 text says that 'the two curves are nearly overlapping' and that there is 'no significant difference in localization accuracy between the two methods in the absence of noise.' The figure confirms this: both methods reach roughly 2.7–3.0 m at 25% known CSI, so the abstract’s 'sub-3 m accuracy' is achieved by both baselines, not uniquely by the proposed model. The claimed advantage appears only in the noisy case (Fig. 10), but no error bars, confidence intervals, or repeated-seed statistics are provided. A formal statistical comparison (e.g., paired tests over multiple runs) or a softened claim is needed.
  4. [§IV.A.3–A.5, Eq. (43)] No ablations isolate the contributions of the three loss terms. The proposed model differs from the Pure MAE baseline by the addition of the JEPA latent-prediction branch, the SIGReg term, and (as described in Sec. IV.A.2) MoE layers and a different positional embedding. The paper attributes the gains to 'JEPA latent-prediction objective' and 'SIGReg regularization' without showing that either component alone produces the advantage. Minimal ablations—MAE alone, MAE+SIGReg, MAE+JEPA, and the full model—are necessary to support the central architectural claims. Without them, the observed improvements could stem from any of the concurrent modifications.
  5. [§V, Table I, Figs. 6–10] Reproducibility is hampered by missing information: no code or data release is indicated; no random seeds, number of runs, or error bars are reported; and several architectural hyperparameters in Table I are incomplete (encoder depth N_blocks, decoder depth N_d, predictor depth N_pred_blocks, MoE hidden dimension d_h, dropout, etc.). Figure 6 shows training and testing loss curves but does not define which loss (total or individual) is plotted or provide axis clarity. At minimum, the authors should provide mean±std over at least three independent runs and release the code or detailed configuration files.
minor comments (6)
  1. [Throughout] Numerous typos and grammatical issues: 'in-cooperated' (§IV), 'pre-trainind' (§IV.B), 'F or' (§II.A heading), 'the pre-trainind encoder', and inconsistent use of 'M_r' vs. 'M r' in Table I. A careful proofreading pass is needed.
  2. [Eq. (5)] The optimization objective uses an expression |Ĥ′ − H|² without defining the norm. Should be a squared Frobenius norm (or clarify if elementwise).
  3. [Eq. (38)] The global pooling operation is applied over 'K_p' tokens, but K_p=1024 appears to denote the number of patches in the JEPA branch, while the positioning head is described as using the full aperture. The indexing (1+K_p) is unclear and should be reconciled with the actual token count.
  4. [Fig. 6] The legend labels 'Testing loss' and 'Training loss' but the text does not specify whether this is the total loss L_pre or a component. The figure is also not referenced in Sec. V.A/B, making its exact role ambiguous.
  5. [References] Refs. [6] and [10] appear to describe the same SSNet work with different bibliographic details; the duplication should be cleaned. Ref. [12] is an arXiv preprint; if possible, cite the published version or clarify status.
  6. [Sec. V.A] The dataset is described as 30,000 independent channel realizations, but the generation from DeepMIMO O1 28 with L=5 paths is not fully explained—for instance, how user positions are sampled and why only 5 paths are retained. A brief description of the ray-tracing setup would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported gains are empirical comparisons against an external MAE baseline; self-citations are contextual, and the K_mask/test-protocol inconsistency is a reproducibility gap, not a definitional reduction.

full rationale

The paper's central claims are empirical. JEPA-CFM is pre-trained with L_pre = gamma*L_MAE + lambda_JEPA*L_JEPA + lambda_SIGReg*L_SIGReg and then compared with a Pure MAE baseline on DeepMIMO. The JEPA target in Eq. (23) is the same encoder's full-view output with detached gradients; this is a standard self-supervised target, not a hidden reuse of the evaluation metric. L_MAE in Eq. (33) and NMSE in Eq. (44) are similar by design (both are masked-region MSE), but NMSE is computed on held-out test samples, so optimizing L_MAE during training does not by itself force the reported test numbers. Self-citations [6], [9], [10], [12] supply background and taxonomy; none is used to prove the extrapolation or positioning gains. There is a genuine internal inconsistency: Section IV.A.2 defines N_o = floor(4096*0.25)=1024, while Section IV.A.3 sets K_mask = K_p - N_o = 768 (implying N_o=256), and Figs. 7-10 evaluate 5%-25% known CSI without specifying how the predictor/decoder are resized or whether the JEPA predictor is bypassed at test time. This is an omitted-protocol and reproducibility risk, not a circular reduction: the plotted errors are not algebraically forced by any fitted parameter or by a self-citation. Therefore no circularity is found.

Assumptions & free parameters 10 free parameters · 5 assumptions · 0 invented entities

The central empirical claims rest on hand-chosen architecture and loss weights, and on assumptions that the simulated DeepMIMO channel, per-sample normalization, random masking, and detached JEPA targets are representative and stable. No new physical entities are introduced.

free parameters (10)
  • gamma (MAE reconstruction loss weight) = 10000.0
    Hand-chosen in Eq. (43) to balance pixel-level reconstruction against the latent losses; central to the multi-objective pre-training.
  • lambda_JEPA (JEPA latent loss weight) = 0.2
    Hand-chosen in Eq. (43); controls the contribution of latent prediction to the pre-training objective.
  • lambda_SIGReg (SIGReg regularization weight) = 0.02
    Hand-chosen in Eq. (43); the paper claims it prevents representation collapse.
  • pre-training mask ratio M_r = 0.75
    Fixed during pre-training; all mask-ratio generalization claims at 5–25% known CSI depend on this choice.
  • encoder hidden dimension d_model = 192
    Architecture hyperparameter selected by hand in Section IV.A.
  • MoE experts E and top-K = E=4, K=2
    Hand-chosen sparse MoE configuration in Section IV.A.2.
  • predictor depth N_pred_blocks = 2
    Hand-chosen depth of the JEPA latent predictor in Section IV.A.3.
  • decoder depth N_d and decoder dim d_d = N_d=4, d_d=64
    Hand-chosen decoder configuration in Section IV.A.4.
  • SIGReg slice count N_s and frequency grid N_t/range = N_s=128, N_t=17, [-5,5]
    Hand-chosen numerical parameters for the sliced characteristic-function regularization in Section IV.A.5.
  • position normalization factor = 350.0 m
    Hand-chosen scaling in Section IV.B to stabilize positioning regression targets.
assumptions (5)
  • domain assumption The DeepMIMO O1_28 ray-tracing channel with L=5 dominant paths is a faithful proxy for real urban FAS propagation.
    All results are generated from this simulation; see Section III.A, Eq. (4), and Section V.A.
  • domain assumption Sample-wise normalization in Eq. (9) removes path-loss and scale variations without destroying the position information needed for localization.
    Positioning is performed using only the normalized CSI; Section IV.A.1.
  • domain assumption Random spatial masking during pre-training (M_r=0.75) transfers to evaluation at 5–25% known CSI and to noisy conditions.
    Section V.B and Figs. 7–10; no retraining or adaptation per test mask ratio is described.
  • domain assumption The detached full-view encoder pass provides useful target latents for JEPA prediction, and SIGReg prevents representation collapse.
    Section IV.A.3 and IV.A.5; no proof or ablation is provided for this stability mechanism.
  • standard math Frobenius normalization, L1/L2 losses, and the trapezoidal characteristic-function integral are numerically stable and differentiable as used.
    Standard numerical operations used throughout Sections IV and V.

how reviews work

0 comments
Cite this review

Pith. "Pith review of JEPA-CFM: A Joint Embedding Predictive Architecture-based Channel Foundation Model for Robust Fluid Antenna Systems." pith.science (2026). https://pith.science/paper/HIHGJIIZ

@misc{pith2026260720202,
  author       = {Pith},
  title        = {Pith review of: JEPA-CFM: A Joint Embedding Predictive Architecture-based Channel Foundation Model for Robust Fluid Antenna Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HIHGJIIZ}},
  note         = {Machine review of arXiv:2607.20202}
}
read the original abstract

Fluid antenna systems (FAS) have emerged as a promising technology for sixth-generation (6G) wireless networks. By allowing antenna elements to move freely within a compact region, FAS can exploit rich spatial diversity without additional hardware. However, acquiring real-time channel state information (CSI), extrapolating channel values to unmeasured antenna ports, and determining accurate user positions remain major obstacles. These challenges stem mainly from strong spatial correlations within the limited aperture and the scarcity of observable data. To overcome these limitations, this paper introduces joint embedding predictive architecture (JEPA)-based channel foundation model (CFM) specifically designed for FAS. The model adopts JEPA to learn versatile representations by extracting high-level latent embeddings of masked or unobserved channel segments. Unlike conventional approaches that attempt pixel-by-pixel reconstruction of raw CSI coefficients, JEPA-CFM focuses on predicting abstract structures in a compact feature space. The pre-training objective combines three complementary loss terms: the standard masked autoencoder reconstruction loss, the JEPA latent prediction loss, and a sliced isotropic Gaussian regularization (SIGReg) term. Together, these components prevent representation collapse and significantly enhance robustness under severe spatial correlation and highly sparse observations. After pre-training, the encoder is frozen, and lightweight task-specific heads are attached: a decoder for channel extrapolation and a global average pooling layer followed by a multi-layer perceptron regression head for wireless positioning. Extensive simulations in the realistic DeepMIMO urban scenario demonstrate that JEPA-CFM substantially outperforms the conventional masked autoencoder baseline in channel extrapolation and wireless positioning.

Figures

Figures reproduced from arXiv: 2607.20202 by the authors.

Figure 1
Figure 1. The architecture of the proposed model [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. The architecture of the encoder. To capture the high-dimensional non-linear channel correla￾tions across the visible ports, the augmented latent matrix X′ p is routed through a series of specialized encoder blocks. In order to enhance feature extraction and boost resilience against deep fading noise, each constituent block is designed as an in￾terleaved micro-topology combining multi-head self-attention (MSA) module… view at source ↗
Figure 3
Figure 3. The architecture of the JEPA latent predictor and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The pipeline of using pre-trained encoder and JEPA [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 6
Figure 6. Figure 6: The training and testing loss of the proposed JEPA [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 9
Figure 9. Figure 9: The testing performance of the proposed model and [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 8
Figure 8. Figure 8: The robustness the proposed model and MAE-based [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 5 linked inside Pith

  1. [6]

    SSNet: Flexible and robust channel extrapolation for fluid antenna systems enabled by an self-supervised learning framework,

    Y . Gao, Y . Liu, R. Yu, S. Liu, Y . Jin, S. Zhang, S. Xu, and X. Chu, “SSNet: Flexible and robust channel extrapolation for fluid antenna systems enabled by an self-supervised learning framework,”IEEE Journal on Selected Areas in Communica- tions, pp. 1–1, Oct. 2025

    2025

  2. [10]

    SSNet: Flexible and robust channel extrapolation for fluid antenna systems enabled by an self-supervised learning framework,

    Y . Gao, Y . Liu, R. Yu, S. Liu, Y . Jin, S. Zhang, S. Xu, and X. Chu, “SSNet: Flexible and robust channel extrapolation for fluid antenna systems enabled by an self-supervised learning framework,”IEEE Journal on Selected Areas in Communica- tions, vol. 44, pp. 1276–1289, Jan. 2026

    2026

  3. [1]

    A contemporary survey on fluid antenna systems: Fundamentals and network- ing perspectives,

    H. Hong, K.-K. Wong, C.-B. Chae, H. Xu, X. Guo, F. R. Ghadi, Y . Chen, Y . Xu, B. Liu, K.-F. Tonget al., “A contemporary survey on fluid antenna systems: Fundamentals and network- ing perspectives,”IEEE Transactions on Network Science and Engineering, 2025

    2025

  4. [2]

    Fluid antennas: Pioneering a new era in next-generation wireless communications,

    D. Bepari, S. Mondal, A. Bhowal, K. Singh, H. Shin, K.- K. Wong, and D. W. K. Ng, “Fluid antennas: Pioneering a new era in next-generation wireless communications,”IEEE Communications Surveys & Tutorials, vol. 28, pp. 5596–5631, 2026

    2026

  5. [3]

    Be water, my antennas: Riding on radio wave fluctuation in nature for spatial multiplexing using programmable meta-fluid antenna,

    B. Liu, K.-F. Tong, K.-K. Wong, C.-B. Chae, and H. Wong, “Be water, my antennas: Riding on radio wave fluctuation in nature for spatial multiplexing using programmable meta-fluid antenna,”arXiv preprint arXiv:2502.04693, 2025

    arXiv 2025

  6. [4]

    A tutorial on fluid antenna system for 6g networks: Encompassing communication theory, optimization methods and hardware designs,

    W. K. New, K.-K. Wong, H. Xu, C. Wang, F. R. Ghadi, J. Zhang, J. Rao, R. Murch, P. Ram ´ırez-Espinosa, D. Morales- Jimenezet al., “A tutorial on fluid antenna system for 6g networks: Encompassing communication theory, optimization methods and hardware designs,”IEEE Communications Surveys & Tutorials, vol. 27, no. 4, pp. 2325–2377, 2024

    2024

  7. [5]

    Fluid antenna systems,

    K.-K. Wong, A. Shojaeifard, K.-F. Tong, and Y . Zhang, “Fluid antenna systems,”IEEE Transactions on Wireless Communica- tions, vol. 20, no. 3, pp. 1950–1962, 2020

    1950

  8. [7]

    Accurate and fast channel estimation for fluid antenna systems with diffusion models,

    E. Tang, W. Guo, H. He, S. Song, J. Zhang, and K. B. Letaief, “Accurate and fast channel estimation for fluid antenna systems with diffusion models,”arXiv preprint arXiv:2505.04930, 2025

    arXiv 2025

Show all 33 references
  1. [8]

    Large language model empowered design of fluid antenna systems: Challenges, frameworks, and case studies for 6g,

    C. Wang, K.-K. Wong, Z. Li, L. Jin, and C.-B. Chae, “Large language model empowered design of fluid antenna systems: Challenges, frameworks, and case studies for 6g,”IEEE Wireless Communications, 2025

    2025

  2. [9]

    Ai-driven channel state information (csi) extrapolation for 6g: Current situations, challenges and future research,

    Y . Gao, Z. Lu, X. Wu, W. Yu, S. Liu, J. Du, Y . Jin, S. Zhang, X. Chu, and S. Xu, “Ai-driven channel state information (csi) extrapolation for 6g: Current situations, challenges and future research,”IEEE Communications Surveys & Tutorials, 2026

    2026

  3. [11]

    Learning-induced channel extrapo- lation for fluid antenna systems using asymmetric graph masked autoencoder,

    H. Zhang, J. Wang, C. Wang, C.-C. Wang, K.-K. Wong, B. Wang, and C.-B. Chae, “Learning-induced channel extrapo- lation for fluid antenna systems using asymmetric graph masked autoencoder,”IEEE Wireless Communications Letters, vol. 13, no. 6, pp. 1665–1669, 2024

    2024

  4. [12]

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

    J. Jiang, Y . Gao, X. Wu, and S. Xu, “Towards channel foundation models (CFMs): Motivations, methodologies and opportunities,”arXiv preprint arXiv:2507.13637, Aug. 2025

    2025 arXiv

  5. [13]

    LeJEPA: Provable and scalable self-supervised learning without the heuristics,

    R. Balestriero and Y . LeCun, “LeJEPA: Provable and scalable self-supervised learning without the heuristics,”arXiv preprint arXiv:2511.08544, 2025

    2025 arXiv

  6. [14]

    A path towards autonomous machine intelli- gence,

    Y . LeCunet al., “A path towards autonomous machine intelli- gence,”Open Review, vol. 62, no. 1, pp. 1–62, 2022

    2022

  7. [15]

    WirelessJEPA: A Multi-Antenna Foundation Model using Spatio-temporal Wire- less Latent Predictions,

    V . Chu, O. Mashaal, and H. Abou-Zeid, “WirelessJEPA: A Multi-Antenna Foundation Model using Spatio-temporal Wire- less Latent Predictions,”arXiv preprint 2601.20190, Jan. 2026

    2026

  8. [16]

    JEPA-MSAC: A Joint-Embedding Predictive Ar- chitecture for Multimodal Sensing-Assisted Communications,

    C. Zheng, J. He, G. Cai, N. Li, M. Bennis, H. Wymeersch, and M. Debbah, “JEPA-MSAC: A Joint-Embedding Predictive Ar- chitecture for Multimodal Sensing-Assisted Communications,” arXiv preprint arXiv:2603.29796, Mar. 2026

    2026

  9. [17]

    Learned-approximate message passing under karhunen-lo `eve modeling for fluid antenna systems,

    Y . Wu, Z. Zhang, H. Jiang, K.-K. Wong, and C.-B. Chae, “Learned-approximate message passing under karhunen-lo `eve modeling for fluid antenna systems,”IEEE Wireless Communi- cations Letters, 2026

    2026

  10. [18]

    Context-aware deep learning for robust chan- nel extrapolation in fluid antenna systems,

    Y . Jin, R. Yu, Y . Gao, S. Liu, X. Chu, K.-K. Wong, and C.-B. Chae, “Context-aware deep learning for robust chan- nel extrapolation in fluid antenna systems,”arXiv preprint arXiv:2507.04435, 2025

    2025

  11. [19]

    Port selection for fluid antenna systems,

    Z. Chai, K.-K. Wong, K.-F. Tong, Y . Chen, and Y . Zhang, “Port selection for fluid antenna systems,”IEEE Communications Letters, vol. 26, no. 5, pp. 1180–1184, 2022

    2022

  12. [20]

    Transmit and receive antenna port selection for channel capacity maximization in fluid-mimo sys- tems,

    C. N. Efrem and I. Krikidis, “Transmit and receive antenna port selection for channel capacity maximization in fluid-mimo sys- tems,”IEEE Wireless Communications Letters, vol. 13, no. 11, pp. 3202–3206, 2024

    2024

  13. [21]

    Switching-cost-aware deep reinforcement learning for dynamic port selection in fluid antenna systems,

    J. Liu, K. Liang, L. Zhao, G. Zheng, H. Gu, K.-K. Wong, and C.-B. Chae, “Switching-cost-aware deep reinforcement learning for dynamic port selection in fluid antenna systems,”IEEE Communications Letters, vol. 30, pp. 1548–1552, 2026

    2026

  14. [22]

    Toward practical fluid antenna systems: Co-optimizing hardware and software for port selection and beamforming,

    S. Xu, K.-K. Wong, Y . Du, H. Hong, C.-B. Chae, B. Liu, and K.- F. Tong, “Toward practical fluid antenna systems: Co-optimizing hardware and software for port selection and beamforming,” 13 IEEE Transactions on Wireless Communications, 2025

    2025

  15. [23]

    Port- llm: A port prediction method for fluid antenna based on large language models,

    Y . Zhang, H. Yin, W. Li, E. Bj ¨ornson, and M. Debbah, “Port- llm: A port prediction method for fluid antenna based on large language models,”IEEE Transactions on Communica- tions, 2025

    2025

  16. [24]

    Secure ISAC with fluid antenna systems: Joint precoding and port selection,

    A. Salem, H. Xu, K.-K. Wong, C.-B. Chae, and Y . Zhang, “Secure ISAC with fluid antenna systems: Joint precoding and port selection,”arXiv preprint arXiv:2509.26572, 2025

    2025

  17. [25]

    Joint port selection and beamforming design for fluid antenna assisted integrated data and energy transfer,

    L. Zhang, H. Yang, Y . Zhao, and J. Hu, “Joint port selection and beamforming design for fluid antenna assisted integrated data and energy transfer,”IEEE Wireless Communications Letters, vol. 13, no. 7, pp. 1833–1837, 2024

    2024

  18. [26]

    Fundamental limits of localization with fluid antenna systems: A fisher information analysis,

    A. Salem, K.-K. Wong, H. Shin, and Y . Zhang, “Fundamental limits of localization with fluid antenna systems: A fisher information analysis,”arXiv preprint arXiv:2512.13941, 2025

    2025

  19. [27]

    Towards intelligent antenna positioning: Leveraging DRL for FAS-aided ISAC systems,

    S. Yang, J. Yao, J. Tang, T. Wu, M. Elkashlan, C. Yuen, M. Debbah, H. Shin, and M. Valenti, “Towards intelligent antenna positioning: Leveraging DRL for FAS-aided ISAC systems,”IEEE Internet of Things Journal, 2025

    2025

  20. [28]

    RSSI Positioning with Fluid Antenna Systems,

    W. Liu, Z. Rong, X. Liu, Y . Jiang, and X. Zhu, “RSSI Positioning with Fluid Antenna Systems,”arXiv preprint arXiv:2503.00980, 2025

    2025 arXiv

  21. [29]

    Indoor fluid antenna systems enabled by layout-specific modeling and group relative policy optimization,

    T. Zhang, Q. Li, S. Wang, W. Ni, J. Zhang, R. Wang, K.- K. Wong, and C.-B. Chae, “Indoor fluid antenna systems enabled by layout-specific modeling and group relative policy optimization,”IEEE Transactions on Wireless Communications, 2025

    2025

  22. [30]

    3GPP, “3rd generation partnership project; technical specifica- tion group radio access network; study on artificial intelligence (AI)/machine learning (ML) for NR air interface (Release 19),” 3GPP, Tech. Rep. TS 38.843 V19.0.0, Sep 2025

    2025

  23. [31]

    AI/ML for mobile networks: Current status in Rel. 19 and challenges ahead,

    Y . Gao, X. Wu, J. Jiang, B. Hu, J. Du, Q. Ye, S. Zhang, F. R. Yu, and S. Xu, “AI/ML for mobile networks: Current status in Rel. 19 and challenges ahead,”arXiv preprint arXiv:2603.14317, Mar. 2026

    2026

  24. [32]

    Enabling 6G through multi-domain channel ex- trapolation: Opportunities and challenges of generative artificial intelligence,

    Y . Gao, Z. Lu, Y . Wu, Y . Jin, S. Zhang, X. Chu, S. Xu, and C.-X. Wang, “Enabling 6G through multi-domain channel ex- trapolation: Opportunities and challenges of generative artificial intelligence,”IEEE Communications Magazine, vol. 64, no. 1, pp. 222–228, Jan. 2026

    2026

  25. [33]

    Sidelink positioning: Standardization advancements, challenges and opportunities,

    Y . Gao, G. Pan, Z. Zhong, Z. Jinm, Y . Hu, Y . Jin, , and S. Xu, “Sidelink positioning: Standardization advancements, challenges and opportunities,”IEEE Communications Maga- zine, vol. 64, no. 4, pp. 128–134, Feb. 2026

    2026

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.