Pith. sign in

REVIEW 4 major objections 4 minor 43 references

MultiPathFormer: Towards a Foundation Model for Multipath Wireless Propagation

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Multipath-path pretraining beats channel-tensor wireless foundation models.

desk verdict A genuinely different pretraining object for wireless foundation models, with a real confound in the headline comparison; still worth a serious referee. read the letter →

arxiv 2608.05076 v1 pith:ODPBEH2V submitted 2026-08-05 cs.LG cs.AIeess.SP

Blessed Guda , Kayley Sze , Carlee Joe-Wong This is my paper · ORCID
classification cs.LGcs.AIeess.SP
keywords wirelessfoundationmodelmultipathpropagationnext-pathpredictionretrieval-augmentedgenerationchannelestimationbeamlocalizationDeepMIMO
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

This paper argues that the right pretraining object for a wireless foundation model is the multipath propagation process itself, not the channel tensor that mixes many paths together. MultiPathFormer represents each transmitter-receiver link as an ordered sequence of continuous-valued path tokens and is pretrained to predict the next path, guided by retrieved scene geometry and a first-path codebook. If correct, this yields reusable, physically interpretable representations: on DeepMIMO scenarios it outperforms channel-based foundation models on beam prediction, localization, line-of-sight classification, and channel estimation, while transferring to unseen users and adapting to new environments. The breakthrough would be a shift from configuration-entangled channel matrices to environment-aware path-level pretraining.

What carries the argument

The central mechanism is next-path prediction over continuous, variable-length multipath sequences. Each link is sorted by descending power into path tokens; an encoder-decoder transformer conditions on transmitter-receiver geometry, Environmental RAG features (dense local objects, dense corridor objects, and sparse multi-scale summaries), and a first-path cluster codebook that supplies a coarse prior over the dominant path's delay and power, with the model learning a residual correction. This converts wireless pretraining from masked channel reconstruction into an autoregressive regression problem with explicit path-count prediction, and it lets the model preserve dominant-path structure, angular geometry, and propagation interactions before they are collapsed into a channel tensor.

What would settle it

Collect real-world channel measurements with known transmitter-receiver positions in scenes similar to the pretraining environments, extract multipath parameters, and evaluate the pretrained MultiPathFormer both zero-shot and after fine-tuning; if its downstream gains over LWM and WiFo shrink to noise or reverse, the path-level pretraining advantage would be an artifact of simulated ray tracing.

Watch

Extended reading notes

Core claim

MultiPathFormer establishes that multipath propagation can serve as the fundamental pretraining object for wireless foundation models. Each transmitter-receiver link is converted into a deterministic sequence of path tokens sorted by descending power, where each token carries delay, power, phase, arrival/departure angles, and interaction type; the model then autoregressively predicts the next path, including the total path count. To stabilize the dominant first path and inject local scene knowledge, the model combines a KMeans-based first-path cluster codebook with an Environmental RAG module that retrieves link-specific objects near the transmitter, receiver, and propagation corridor. Pretrained on 27 ray-traced environments with 23.9M path tokens, the model transfers to held-out users and, after scenario-specific fine-tuning, outperforms training the same architecture from scratch in new environments; on downstream tasks it outperforms the channel-based LWM and WiFo baselines, achieving 0.914 top-3 beam accuracy, 5.57 m mean localization error, 0.994 LoS classification accuracy, and 0.561 channel-estimation NMSE.

Load-bearing premise

The whole demonstration rests on DeepMIMO ray-tracing simulations faithfully representing real-world multipath propagation; the paper itself flags that an important next step is to test whether the gains carry over from ray-tracing to actual measurement scenarios.

Editorial extensions

If this is right

  • Pretrained path-level representations transfer to unseen users within known environments, holding delay and power errors low without additional training.
  • Scenario-specific fine-tuning from the foundation checkpoint beats training the same architecture from scratch on new environments, giving a practical adaptation recipe.
  • Explicit access to the dominant path's angle of departure directly informs beam selection, yielding 0.914 top-3 beam accuracy over channel-based baselines.
  • Line-of-sight versus non-line-of-sight structure emerges naturally from the generated multipath, enabling 0.994 classification accuracy even with a simple power-dominance rule.
  • Channel estimation can reuse predicted paths as a sparse basis, combining the generated propagation support with observed subcarriers to recover the full channel without scenario-specific fine-tuning.

Reading between the lines

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

  • If these gains hold on real-world measurements, path-level pretraining could make wireless models genuinely environment-aware, letting a radio adapt to a city's geometry from map data alone.
  • The Environmental RAG design suggests that explicit geometric retrieval is a scalable way to inject site knowledge; a natural extension would be learning the retrieval function end-to-end over building footprints instead of using hand-crafted distance and height features.
  • The first-path codebook idea could generalize to anchoring other dominant structures, such as the strongest reflector, and could be updated online when a transmitter's position or environment changes.
  • A key bottleneck for deployment will be aligning simulated path parameters with paths extracted from real channel sounders, since ray-tracing and measurement differ in path discovery and noise characteristics.
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

4 major / 4 minor

Summary. The manuscript proposes MultiPathFormer, an autoregressive transformer that pretrains on multipath propagation sequences rather than channel tensors. Each transmitter-receiver link is represented as a power-sorted sequence of continuous-valued path tokens, and the model is trained with next-path prediction together with auxiliary path-count and interaction-type losses. The architecture adds an Environmental RAG mechanism and a first-path KMeans codebook. Experiments on DeepMIMO cover path-generation fidelity, transfer to held-out users and unseen scenarios, scenario-specific finetuning, and four downstream tasks (beam prediction, localization, LoS classification, channel estimation), reporting superiority over the LWM and WiFo channel-based foundation models.

Significance. If validated, path-level pretraining would be a timely and conceptually appealing reframing of wireless foundation models, since multipath parameters are the physical quantities that generate the channel tensor. The manuscript has concrete strengths: the authors release model weights, pretrain on a large and diverse set of 27 DeepMIMO scenarios with 23.9M path tokens, include a detailed appendix with hyperparameters and evaluation protocols, and provide clean ablations for the generation objective. However, the central claim that path-level pretraining yields reusable representations is not yet isolated from the auxiliary inputs in the downstream comparison, and the finetuning transfer gain is small. The significance of the paper is therefore conditional on closing this gap with further experiments.

major comments (4)
  1. [Section V-E, Table III] The downstream comparison is confounded: the full MultiPathFormer uses Environmental RAG features and a transmitter-specific first-path codebook, while LWM and WiFo receive only channel tensors. The paper's statement that simpler MultiPathFormer variants would be at least comparable relies on Table I, but Table I reports path-generation MAE and interaction scores, not beam accuracy, localization error, LoS accuracy, or channel estimation NMSE. Please report downstream metrics for the Direct and +Codebook variants, or give the baselines equivalent geometry/codebook inputs. Without this, the headline downstream superiority cannot be attributed to the path-level pretraining object rather than to privileged auxiliary inputs. For localization, receiver-dependent RAG and codebook slots are zeroed, but transmitter geometry and scene features still enter the encoder, so the localization comparison is also not free of auxiliary information.
  2. [Section III-D, Eqs. (15)-(19)] The first-path codebook is fitted to the ground-truth power and delay of training receivers for each transmitter, and at test time a receiver is assigned a prior by nearest-neighbor location among training receivers. This is a supervised spatial lookup rather than a learned generative prior, and for beam prediction and channel estimation it injects empirical first-path statistics of nearby training receivers that are not available to channel-based baselines. The reported 38.7% delay and 59.2% power MAE improvements from adding the codebook therefore do not isolate the effect of the next-path pretraining objective. Please clarify how the codebook is constructed for the four unseen zero-shot scenarios, where no training receivers exist, and state explicitly whether those zero-shot results use the codebook at all.
  3. [Table II] The adaptation benefit of the foundation checkpoint over from-scratch training is 0.065 versus 0.068 delay MAE and 2.888 versus 3.081 power MAE, with no error bars, repeated-seed statistics, or significance test reported for the second block of the table. This small margin is the main evidence for the claim that pretraining provides reusable representations in new environments. Please report variance across seeds or, alternatively, show downstream-task finetuning comparisons between foundation-initialized and from-scratch models.
  4. [Section V-E.3 and Section IV] The reported zero-shot LoS result (0.976 accuracy, 0.959 F1) is produced by a hand-built rule that classifies based on dominance of the maximum path power over the remaining paths, not by a decision rule learned or adapted from the pretrained representation. As presented, this does not demonstrate zero-shot generalization from pretraining. Please state how the dominance threshold is chosen, whether it is tuned on validation data, and compare against the same rule applied to ground-truth multipaths to calibrate the claim.
minor comments (4)
  1. [Section III-A, Eqs. (3)-(4)] There is a typo in 'L_{m,b,u} number of mulitpath', and 'multihot vector' should be 'multilabel indicator' for the interaction type. In addition, the symbol θ is used both for model parameters and for angle variables in Eq. (4), which is confusing; please rename one of them.
  2. [Section I, contribution 4] The claimed improvements in the introduction do not match Table III: the top-1 beam accuracy improvement over the MLP baseline is about 49% relative, the localization gain is about 63 m, and the LoS accuracy gain is about 1.6 percentage points. The phrase '0.751 Channel estimation NMSE' is also not a quantity that appears in Table III. Please reconcile these numbers with the reported table.
  3. [Figure 3] The y-axis labels and units are missing from the figure in the submitted text; please add explicit units for delay MAE and power MAE so the reader can interpret the low-data curves.
  4. [Appendix d] The evaluation protocol states that results are 'for each path averaged over the actual number of multipaths and then over the users', but it does not state how the error bars in Table I are computed (across scenarios, users, or seeds). Please clarify, and consider adding error bars to Table II's adaptation block as noted above.

Circularity Check

1 steps flagged · score 3.0 of 10

Partial circularity: the first-path codebook is a fitted label lookup, and the downstream comparison is confounded by privileged auxiliary inputs, though the core pretraining objective is not definitionally circular.

  1. fitted input called prediction [Section III-D, Eqs. (15)-(19); Table I ablation]
    "For each transmitter b in environment m, we fit K-Means on the first-path power and delay of all corresponding receivers in the training set: {c_k} = KMeans({(ρ_1, τ_1) : u in U_train}). ... At test time, a receiver is assigned a first-path prior through nearest-neighbor transfer within the same transmitter: u* = arg min ||x_rx - x_rx'||^2. ... by(1) = c + f_res(h1, c, σ, x)."

    The codebook's first-path delay/power prior is directly fitted to ground-truth training labels (Eq. 15), and at test time the receiver is assigned the cluster centroid of the nearest-in-location training receiver (Eqs. 17-18). The final first-path prediction is then that stored label centroid plus a residual correction (Eq. 19). Thus the reported 38.7% delay-MAE and 59.2% power-MAE reductions from the codebook are substantially a spatial lookup of training labels rather than a prediction produced by the learned autoregressive path model. This is a fitted input presented as an architectural contribution to path statistics estimation.

full rationale

The central pretraining objective—next-path prediction over ground-truth multipath sequences—is a normal self-supervised formulation and is not defined in terms of the downstream tasks. The transfer experiments (Table II) compare foundation-finetuning against from-scratch training on the same architecture, which is a fair test of initialization benefit. No load-bearing self-citation or imported uniqueness theorem is present. The clearest circularity-like issue is the first-path codebook: it is fit to the very quantity it is later credited with predicting (first-path power/delay) and assigned by location nearest-neighbor, so part of the generation-quality gain is a kNN label lookup. The downstream comparison in Section V-E is additionally confounded because the full MultiPathFormer receives environment RAG features and a transmitter-specific codebook not given to LWM/WiFo; the paper's assertion that Table I supports the simpler variants matching channel-based baselines is not supported, since Table I reports path-generation MAE, not downstream accuracy. However, that is an evidence gap/confound rather than a definitional circularity, and the localization task zeroes the codebook and RAG. The zero-shot LoS rule is a hand-built threshold on generated paths, not circular. Overall, the core claim is not equivalent to its inputs, so a score of 3 is appropriate rather than a higher score.

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

The central results rest on the validity of DeepMIMO ray-tracing as a stand-in for real propagation, on the sufficiency of the path representation in Eq. (4), and on a set of hand-chosen hyperparameters (cluster count, retrieval radii, loss weights). No new physical entities are introduced.

free parameters (4)
  • KMeans cluster count K = 25
    Set in Appendix (b); controls the granularity of the first-path prior and directly affects the residual prediction in Eq. (19).
  • RAG retrieval radii and top-K counts = radii 25/50/100 m, K_local=5, K_corr=5, B=8 bins
    Appendix (b); these hand-chosen numbers define the environment context fed to the model and the reported gains depend on them.
  • Loss weights lambda_len and lambda_int = 0.01
    Appendix (c); balance path, length, and interaction losses, set by hand without a reported sensitivity analysis.
  • Maximum sequence length = 25
    Appendix (a); caps the number of predicted paths per link and affects the termination behavior of generation.
assumptions (4)
  • domain assumption The multipath tuple in Eq. (4) is a complete and sufficient description of the propagation channel for the pretraining objective.
    Used to define the prediction target and the downstream tasks; if path parameters were insufficient, pretraining would miss channel information.
  • domain assumption DeepMIMO ray-tracing outputs faithfully reproduce real-world multipath propagation.
    All experiments use simulated data; the authors themselves flag real-world validation as future work in Section VI.
  • domain assumption OpenStreetMap-derived 3D object geometry is accurate enough for the RAG context.
    Section III-C assumes objects 'can easily be obtained' from OSM and are usable without modeling their errors.
  • domain assumption Ordering paths by descending power is a valid autoregressive ordering for generation.
    Section III-A; the paper notes sorting by delay yields similar performance, so the sequential structure is not uniquely determined by physics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MultiPathFormer: Towards a Foundation Model for Multipath Wireless Propagation." pith.science (2026). https://pith.science/paper/ODPBEH2V

@misc{pith2026260805076,
  author       = {Pith},
  title        = {Pith review of: MultiPathFormer: Towards a Foundation Model for Multipath Wireless Propagation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ODPBEH2V}},
  note         = {Machine review of arXiv:2608.05076}
}
read the original abstract

Recent advances in machine learning have enabled training of wireless foundation models, which aim to support tasks such as channel estimation, beam prediction, and localization based on wireless signals. Existing wireless foundation models typically pretrain on channel tensors using masked reconstruction over subcarriers, antennas, or time but ignore the physical characteristics of wireless propagation. In this work, we propose to instead use multipath propagation as the fundamental pretraining object. We present MultiPathFormer, an autoregressive foundation model that represents each transmitter-receiver link as an ordered sequence of continuous-valued path tokens and pretrains with next-path prediction. We introduce an Environmental RAG (retrieval-augmented generation) mechanism and a first-path codebook on top of the transformer backbone, leveraging environment knowledge to improve path statistics estimation like delay and power by up to 59%. MultiPathFormer pretrained on 27 environments transfers to unseen users and, after scenario-specific fine-tuning, outperforms training the corresponding models from scratch in new environments. Across downstream tasks, it outperforms SOTA channel-based foundation models, achieving 5.57 m mean localization error, 0.914 top-3 beam accuracy, 0.994 line-of-sight classification accuracy, and 0.561 channel estimation NMSE. These results show that path-level pretraining can learn reusable representations of wireless propagation.

Figures

Figures reproduced from arXiv: 2608.05076 by the authors.

Figure 1
Figure 1. a) An example physical environment and the corresponding 3D model through which wireless signals may propagate. b) User clusters obtained from KMeans on first-path delay and power (right), showing that the clusters organize users into spatially coherent groups (left) under a fixed transmitter. the model simultaneously predicts the path parameters defined in Equation (4). To terminate generation, rather than define a… view at source ↗
Figure 2
Figure 2. Overview of the proposed corridor-aware MultiPathFormer. The [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. MultiPathFormer retains low MAE with fewer training users. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative generation example for a receiver in a sample city. Direct captures the broad trend but drifts on early dominant paths. The codebook [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Object-token attention visualization. We expose each retrieved scene [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 5
Figure 5. Figure 5: Spatial comparison of predicted AoA azimuth for one transmitter. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 24 canonical work pages

  1. [28]

    Wico-mg: Wireless channel foundation model for multipath generation via synesthesia of machines,

    Z. Han, L. Bai, X. Cai, and X. Cheng, “Wico-mg: Wireless channel foundation model for multipath generation via synesthesia of machines,”arXiv preprint arXiv:2511.15026, 2025. [Online]. Available: https://arxiv.org/abs/2511.15026

    work page arXiv 2025

  2. [1]

    On the opportunities and risks of foundation models,

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

    arXiv 2021

  3. [2]

    Language mod- els are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askellet al., “Language mod- els are few-shot learners,”Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    1901

  4. [3]

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

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    arXiv 2010

  5. [4]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,”arXiv preprint arXiv:2103.00020, 2021. [Online]. Available: https://arxiv.org/abs/2103.00020

    arXiv 2021

  6. [5]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control,

    B. Zitkovich, T. Yu, S. Xuet al., “Rt-2: Vision-language-action models transfer web knowledge to robotic control,” inProceedings of The 7th Conference on Robot Learning, ser. Proceedings of Machine Learning Research, J. Tan, M. Toussaint, and K. Darvish, Eds., vol

    work page

  7. [6]

    A generalist agent,

    S. Reed, K. Zolna, E. Parisotto, S. G. Colmenarejo, A. Novikov, G. Barth-Maron, M. Gimenez, Y . Sulsky, J. Kay, J. T. Springenberg, T. Eccles, J. Bruce, A. Razavi, A. Edwards, N. Heess, Y . Chen, R. Hadsell, O. Vinyals, M. Bordbar, and N. de Freitas, “A generalist agent,” 2022. [Online]. Available: https://arxiv.org/abs/2205.06175

    arXiv 2022

  8. [7]

    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,” inPro- ceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), 2019, pp. 4171–4186

    2019

Show all 43 references
  1. [8]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    2017

  2. [9]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Biet al., “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  3. [10]

    Wireless ai evolution: From statistical learners to electromagnetic-guided foundation models,

    J. Xiao, J. Wang, K. Cao, X. Li, Z. Chen, and C. Yuen, “Wireless ai evolution: From statistical learners to electromagnetic-guided foundation models,”IEEE Communications Magazine, 2026

    2026

  4. [11]

    Millimeter wave mobile communications for 5g cellular: It will work!

    T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y . Azar, K. Wang, G. N. Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave mobile communications for 5g cellular: It will work!”IEEE Access, vol. 1, pp. 335–349, 2013

    2013

  5. [12]

    Wide- band millimeter-wave propagation measurements and channel models for future wireless communication system design,

    T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun, “Wide- band millimeter-wave propagation measurements and channel models for future wireless communication system design,”IEEE Transactions on Communications, vol. 63, no. 9, pp. 3029–3056, 2015

    2015

  6. [13]

    Study on channel model for frequencies from 0.5 to 100 ghz (release 14),

    B. 3GPP, “Study on channel model for frequencies from 0.5 to 100 ghz (release 14),”3GPP TR 38.901 V14. 0.0, 2017

    2017

  7. [14]

    Scalable pre-trained large channel models for wireless communication systems,

    J. Guo, J. Xue, J. Zhang, Z. Qiao, and Z. Xu, “Scalable pre-trained large channel models for wireless communication systems,”IEEE Communi- cations Magazine, 2026

    2026

  8. [15]

    Large ai models for wireless physical layer,

    J. Guo, Y . Cui, S. Jin, and J. Zhang, “Large ai models for wireless physical layer,”IEEE Communications Magazine, 2026

    2026

  9. [16]

    Wifo: Wireless foundation model for channel prediction,

    B. Liu, S. Gao, X. Liu, X. Cheng, and L. Yang, “Wifo: Wireless foundation model for channel prediction,”Science China Information Sciences, vol. 68, no. 6, p. 162302, 2025

    2025

  10. [17]

    Channelgpt: A large model to generate digital twin channel for 6g environment intelligence,

    L. Yu, L. Shi, J. Zhang, J. Wang, Z. Zhang, Y . Zhang, and G. Liu, “Channelgpt: A large model to generate digital twin channel for 6g environment intelligence,”arXiv preprint arXiv:2410.13379, 2024

    2024 arXiv

  11. [18]

    Large wireless model (lwm): A foundation model for wireless channels,

    S. Alikhani, G. Charan, and A. Alkhateeb, “Large wireless model (lwm): A foundation model for wireless channels,”arXiv preprint arXiv:2411.08872, 2024. [Online]. Available: https://arxiv.org/abs/2411. 08872

    2024 arXiv

  12. [19]

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

    T. Yang, P. Zhang, M. Zheng, Y . Shi, L. Jing, J. Huang, and N. Li, “Wirelessgpt: A generative pre-trained multi-task learning framework for wireless communication,”IEEE Network, 2025

    2025

  13. [20]

    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, 2025

    2025 arXiv

  14. [21]

    Available: https://www.openstreetmap.org/#map=5/38.01/-95

    [Online]. Available: https://www.openstreetmap.org/#map=5/38.01/-95. 84

  15. [22]

    Masked autoencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked autoencoders are scalable vision learners,” 2021. [Online]. Available: https://arxiv.org/abs/2111.06377

    2021 arXiv

  16. [23]

    Robust channel representation for wireless: A multi-task masked contrastive approach,

    B. Guler, G. Geraci, and H. Jafarkhani, “Robust channel representation for wireless: A multi-task masked contrastive approach,” inNeurIPS 2025 Workshop: AI and ML for Next-Generation Wireless Communications and Networking, 2025. [Online]. Available: https://openreview.net/foru...

    2025

  17. [24]

    Spa-mae: A physics- guided csi foundation model for wireless physical layer,

    C. Chen, W. Jin, H. He, X. Sun, and S. Jin, “Spa-mae: A physics- guided csi foundation model for wireless physical layer,”arXiv preprint arXiv:2605.19849, 2026

    2026 arXiv

  18. [25]

    A wireless foundation model for multi-task prediction,

    Y . Sheng, J. Wang, X. Zhou, L. Liang, H. Ye, S. Jin, and G. Y . Li, “A wireless foundation model for multi-task prediction,”arXiv preprint arXiv:2507.05938, 2025

    2025 arXiv

  19. [26]

    Towards a wireless physical- layer foundation model: Challenges and strategies,

    J. Fontaine, A. Shahid, and E. De Poorter, “Towards a wireless physical- layer foundation model: Challenges and strategies,” in2024 IEEE Inter- national Conference on Communications Workshops (ICC Workshops). IEEE, 2024, pp. 1–7

    2024

  20. [27]

    Multi-modal data-enhanced foundation models for prediction and control in wireless networks: A survey,

    H. Zhang, M. Farzanullah, M. Ghassemi, A. B. Sediq, A. Afana, and M. Erol-Kantarci, “Multi-modal data-enhanced foundation models for prediction and control in wireless networks: A survey,”IEEE Commu- nications Surveys & Tutorials, vol. 28, pp. 4359–4393, 2025

    2025

  21. [29]

    Deep learning for mmwave beam and blockage prediction using sub-6 ghz channels,

    M. Alrabeiah and A. Alkhateeb, “Deep learning for mmwave beam and blockage prediction using sub-6 ghz channels,”IEEE Transactions on Communications, vol. 68, no. 9, pp. 5504–5518, 2020

    2020

  22. [30]

    Deep learning coordinated beamforming for highly-mobile millimeter wave systems,

    A. Alkhateeb, S. Alex, P. Varkey, Y . Li, Q. Qu, and D. Tujkovic, “Deep learning coordinated beamforming for highly-mobile millimeter wave systems,”IEEE Access, vol. 6, pp. 37 328–37 348, 2018

    2018

  23. [31]

    Auto-csinet: Scenario-customized automatic neural network architecture generation for massive mimo csi feedback,

    X. Li, J. Guo, C.-K. Wen, and S. Jin, “Auto-csinet: Scenario-customized automatic neural network architecture generation for massive mimo csi feedback,” 2023. [Online]. Available: https://arxiv.org/abs/2311.15950

    2023 arXiv

  24. [32]

    Deep learning-based channel estimation,

    M. Soltani, V . Pourahmadi, A. Mirzaei, and H. Sheikhzadeh, “Deep learning-based channel estimation,” 2019. [Online]. Available: https://arxiv.org/abs/1810.05893

    2019 arXiv

  25. [33]

    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 Communications Letters, vol. 7, no. 5, pp. 852–855, 2018

    2018

  26. [34]

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

    A. Alkhateeb, “Deepmimo: A generic deep learning dataset for millimeter wave and massive MIMO applications,”CoRR, vol. abs/1902.06435, 2019. [Online]. Available: http://arxiv.org/abs/1902. 06435

    1902 arXiv

  27. [35]

    Trocr: Transformer-based optical character recognition with pre-trained models,

    M. Li, T. Lv, J. Chen, L. Cui, Y . Lu, D. Florencio, C. Zhang, Z. Li, and F. Wei, “Trocr: Transformer-based optical character recognition with pre-trained models,” inProceedings of the AAAI conference on artificial intelligence, vol. 37, no. 11, 2023, pp. 13 094–13 102

    2023

  28. [36]

    3d urban outdoor wifi 7 network planning and analysis using ray-tracing and machine learning: Transformer-based surrogate modeling for high-resolution digital twin,

    E.-C. Tr ˆınc, C. Ancut,i, A. Vesa, C. Simu, V .-A. Nit,˘a, and C. Stolojescu- Cris ¸an, “3d urban outdoor wifi 7 network planning and analysis using ray-tracing and machine learning: Transformer-based surrogate modeling for high-resolution digital twin,”Sensors, vol. 26, no. 7,

  29. [37]

    Wireless communications–principles and practice, (the book end)

    T. S. Rappaport, “Wireless communications–principles and practice, (the book end).”Microwave Journal, vol. 45, no. 12, pp. 128–129, 2002

    2002

  30. [38]

    Sionna: An open-source library for next-generation physical layer research,

    J. Hoydis, S. Cammerer, F. A. Aoudia, A. Vem, N. Binder, G. Marcus, and A. Keller, “Sionna: An open-source library for next-generation physical layer research,” 2023. [Online]. Available: https://arxiv.org/abs/2203.11854

    2023 arXiv

  31. [39]

    Wireless em propagation software

    RemCom, “Wireless em propagation software.” [Online]. Available: https://www.remcom.com/wireless-insite-propagation-software

  32. [40]

    Omniverse digital twin — aerial omniverse digital twin

    “Omniverse digital twin — aerial omniverse digital twin.” [Online]. Available: https://docs.nvidia.com/aerial/aerial-dt/index.html

  33. [41]

    Deep learning for fading channel prediction,

    W. Jiang and H. D. Schotten, “Deep learning for fading channel prediction,”IEEE Open Journal of the Communications Society, vol. 1, pp. 320–332, 2020

    2020

  34. [229]

    2165–2183

    PMLR, 06–09 Nov 2023, pp. 2165–2183. [Online]. Available: https://proceedings.mlr.press/v229/zitkovich23a.html

    2023

  35. [2026]

    Available: https://www.mdpi.com/1424-8220/26/7/2223

    [Online]. Available: https://www.mdpi.com/1424-8220/26/7/2223

Pith tools

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