Pith. sign in

REVIEW 4 major objections 31 references

Sparse height training can build continuous, coherent 3D radio maps from 1 to 120 m with far less inference cost than diffusion.

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

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

T0 review · grok-4.5

2026-07-14 15:50 UTC pith:5PI7XCSA

load-bearing objection Solid systems paper: continuous-height 3D radio maps via flow matching + bidirectional layer fusion, with a real new dataset and clear efficiency gains; the continuous-PE claim is not cleanly isolated but the overall empirical package is still worth engaging. the 4 major comments →

arxiv 2607.09778 v1 pith:5PI7XCSA submitted 2026-07-08 eess.SP

BDFlow-3DRM: Height-Coherent 3D Radio Map Construction via Bi-Dynamical Flow Matching

Jun Yu , Meixia Tao , Shu Sun This is my paper
classification eess.SP
keywords 3D radio mapheight-aware modelinglow-altitude intelligent networksflow matchingzero-shot generalizationbi-dynamical inter-layer correlationsemantic latent space
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.

Low-altitude wireless networks need three-dimensional radio maps that tell how signal strength varies with location, including height. Most learning methods treat height layers as independent slices or fixed grids, so they fail at unseen heights and lose vertical consistency. This paper argues that the right formulation is a deterministic probability flow in a semantic latent space: continuous dual encodings of absolute receiver height and transmitter-relative height give geometric awareness, while a bi-dynamical block jointly fuses neighboring layers in both directions during generation. Trained on only about twenty receiver-height layers, the model predicts accurate maps over a continuous 1–120 m receiver range under variable transmitter heights, cuts normalized error by roughly 29% versus diffusion baselines, and reduces inference complexity by about 180 times. A new large-scale low-altitude dataset and transfer tests on realistic city geometry support the claim that height-coherent 3D maps can be built at practical scale.

Core claim

BDFlow-3DRM shows that 3D radio-map construction can be solved as a height-conditioned deterministic flow in semantic latent space: continuous dual height encodings plus bidirectional inter-layer fusion let a model trained on sparse discrete receiver heights produce accurate, vertically coherent maps over a wide continuous height range, including unseen and extrapolated heights, at far lower inference cost than diffusion.

What carries the argument

Bi-dynamical flow matching: a recursive bidirectional fusion of neighboring height-layer latents and noises, controlled by adaptive, environment- and height-dependent coefficients, so that the probability-flow ODE constructs all layers jointly while remaining invertible for decoding.

Load-bearing premise

That sinusoidal embeddings of absolute receiver height and transmitter-relative height, learned from sparse discrete layers, induce a smooth enough map from physical height geometry to electromagnetic field structure for reliable interpolation and extrapolation.

What would settle it

Train on the stated twenty receiver-height layers and measure normalized error and vertical-profile coherence on held-out intermediate heights (e.g., 8, 28, 48 m) and extrapolated heights (150–200 m) under new transmitter heights; if error rises sharply or vertical blockage patterns break, the continuous-height claim fails.

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

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

4 major / 0 minor

Summary. The paper proposes BDFlow-3DRM, a bi-dynamical flow-matching framework for 3D radio-map (RM) construction in low-altitude settings. It formulates multi-layer RM generation as a deterministic probability flow in an SVG semantic latent space, injects continuous dual height encodings (absolute Rx height and Tx-relative height), and introduces a bidirectional recursive fusion block with an adaptive eta predictor and ranking loss so neighboring height layers are constructed jointly. Empirically, on UrbanRadio3D the method reports a 28.6% NMSE reduction versus RadioDiff-3D with roughly 180× lower inference FLOPs; on the authors’ LAIN-Radio3D dataset it claims accurate zero-shot prediction over continuous Rx heights 1–120 m (and beyond) from only ~20 training Rx layers under variable Tx heights; limited zero-/few-shot transfer to the realistic LAMBDA SF scenario is also reported.

Significance. If the height-generalization and efficiency claims hold under broader validation, the work is practically relevant for environment-aware low-altitude networking, where Tx/Rx heights vary continuously and dense RT or measurement campaigns are costly. Strengths that should be credited include: (i) release of the large-scale LAIN-Radio3D dataset (Rx to 120 m, variable Tx heights, 5.07M slices); (ii) multi-dataset evaluation spanning simplified and realistic urban scenes; (iii) ablations of inter-layer correlation (Table III), vertical-profile coherence (Fig. 4), and eta-predictor behavior (Fig. 5); and (iv) a clear efficiency comparison showing high-quality generation with 1–5 ODE steps (Table VI). The combination of continuous height conditioning with bidirectional layer coupling is a useful design point relative to slice-wise or full-volume diffusion baselines.

major comments (4)
  1. The load-bearing continuous-height claim (Abstract; §III-A; §IV-G; Tables IV–V) attributes accurate interpolation/extrapolation over 1–120 m (and 150/200 m) from ~20 discrete training Rx layers to continuous dual sinusoidal PE of absolute Rx and Tx-relative height (Eqs. 4–5). However, the reported zero-shot numbers use the full stack (dynamic D=5 sampling, SVG latent, bi-dynamical fusion with adaptive η, ranking loss). Ablations in Table III and Fig. 4 only toggle correlation on UrbanRadio3D’s 1–20 m band; there is no ablation that removes or replaces PE (e.g., discrete height indices, absolute-only PE, or no height embedding) on the LAIN zero-shot suite. Without that isolation, it is not established that continuous PE—not joint multi-layer fusion plus dense RT supervision—is what carries wide-range generalization. Please add this ablation on LAIN (and report the same metrics as Tables I
  2. Table II / §IV-C: the 28.6% NMSE gain versus RadioDiff-3D mixes architecture differences with training-protocol differences. Baselines are trained on complete Rx heights 1–20 m, while BDFlow-3DRM is trained on a sparse subset (1,4,7,9,11,13,15,17,19,20 m) and still evaluated on the full range. That protocol is a strength for the sparse-height story, but it confounds a pure method comparison. Please also report BDFlow-3DRM trained under the same full-height protocol as the baselines (and, ideally, baselines under the sparse protocol) so the architecture contribution can be separated from the sparse-training generalization contribution.
  3. §IV-G3 / Fig. 9: zero-shot transfer to realistic LAMBDA SF is substantially weaker (NMSE 0.0246, Power-MAE 4.19 dB) than in-domain LAIN results, and strong performance requires 15–20 fine-tuning samples. Given that almost all primary evidence is ray-traced (WinProp/SionnaRT), the Abstract’s “practical potential for large-scale 3D RM construction” and “robust zero-shot … transferability to realistic urban environments” should be qualified, and the paper should state more clearly what is and is not demonstrated without measured data or fine-tuning.
  4. §IV-H / Table VI: the ~180-fold complexity reduction is measured against RadioDiff-3D with a 1000-step reverse chain. Modern diffusion samplers often use far fewer steps; without a matched low-NFE diffusion baseline (or reporting quality vs. NFE for both families), the efficiency claim overstates the practical gap. Please either re-benchmark diffusion baselines at comparable NFE/latency or frame the comparison strictly as “vs. the authors’ 1000-step RadioDiff-3D configuration.”

Circularity Check

0 steps flagged

No circularity: standard flow-matching training and held-out RT evaluation; continuous-height claims are empirical, not definitional.

full rationale

BDFlow-3DRM formulates 3D RM construction as conditional flow matching in an SVG latent space (Eqs. 1–2, 14–16), with dual sinusoidal height encodings (Eqs. 4–5), bi-dynamical fusion (Eqs. 9–13), and an eta predictor regularized by ranking loss (Alg. 1, L_total = L_FM + L_rank). The velocity target is the difference between fused noise and fused latents under fixed coefficients at each t; this is ordinary supervised regression of a constructed path, not a redefinition of the evaluation metrics. All reported NMSE/PSNR/SSIM/Power-MAE (Tables II–VI) are computed against independent ray-traced ground truth (WinProp/SionnaRT/LAMBDA) on held-out height configurations and environments. Sparse-height training followed by zero-shot evaluation on unseen Rx/Tx heights is an empirical generalization test, not a fit-then-predict circularity: the continuous PE map is an inductive bias whose success is measured externally, not forced by construction. No load-bearing uniqueness theorem, self-citation chain, or ansatz smuggled as derivation appears; citations (flow matching, dynamical diffusion, DINOv3, SVG) supply standard components. Ablation incompleteness on the PE mechanism is a validity concern, not circularity. The derivation chain is self-contained and non-circular.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 3 invented entities

The central performance claims rest on standard generative-model and radio-propagation assumptions, several hand-chosen training hyperparameters, and architectural modules invented for this paper. No new physical particles or forces are postulated; the invented entities are method components whose only evidence is the paper’s own ablations and metrics.

free parameters (6)
  • Number of sampled height layers D
    Fixed to D=5 during training for efficiency/performance balance (§IV-A); not derived from first principles.
  • Available training height pool size N
    Set to N=20 discrete Rx layers for main experiments (§IV-A); sparse subset choices (e.g., 1,4,7,...,20 m) are design choices that affect the generalization claim.
  • eta_init prior and residual gate λ
    η_init=0.5 and negative-initialized learnable λ stabilize early training of fusion coefficients (Eq. 19); hand-chosen priors.
  • ODE integration step count
    Best reported metrics use 5 Euler steps; 1–50 steps explored (Table VI). Inference quality/complexity tradeoff depends on this choice.
  • SVG latent dimensions and DINOv3 backbone
    384-dim semantic + 24-dim residual features at 16×16 resolution with frozen DINOv3-s16p (§IV-A); architectural hyperparameters that condition all generation quality.
  • Learning rates and batch size
    AdamW, batch 16, LR 1e-4 then cosine to 5e-5; standard but free training knobs.
axioms (5)
  • standard math Conditional flow matching with linear path xt=(1-t)x0+tε yields a learnable velocity field whose ODE integration generates samples from the data distribution.
    Section II-A imports the Lipman et al. flow-matching framework as the generative backbone.
  • domain assumption Ray-tracing / dominant-path simulations (WinProp DPM, SionnaRT) provide sufficiently accurate ground-truth radio maps for training and metric evaluation.
    All primary labels in UrbanRadio3D and LAIN-Radio3D are simulated (§IV-A); real-world correctness of the method inherits this assumption.
  • domain assumption Neighboring height layers of a 3D radio map exhibit bidirectional vertical correlations that can be usefully fused during generation.
    Motivation for the bi-dynamical block in §III-C; ablation Table III supports usefulness but does not prove universality.
  • ad hoc to paper Recursive bidirectional fusion of latents and noises remains a tractable Gaussian path for flow matching under fixed fusion coefficients.
    Asserted in §III-C1 after Eqs. 10–14; invertibility of reverse bi-dynamical block is claimed by construction without a formal proof appendix.
  • ad hoc to paper Sinusoidal encodings of absolute Rx height and Tx-relative height suffice for continuous geometric awareness and zero-shot height interpolation/extrapolation.
    Core of dual height encoding in §III-A; load-bearing for the 1–120 m continuous-range claim.
invented entities (3)
  • Bi-dynamical inter-layer correlation block (bidirectional Dyn/Mid-Dyn fusion) no independent evidence
    purpose: Couple neighboring height-layer latents and noises so 3D RMs are generated jointly with vertical coherence.
    Extension of unidirectional dynamical diffusion to bidirectional height fusion (Eqs. 8–14, Fig. 2); evidence is internal ablations only.
  • Adaptive eta predictor with ranking loss no independent evidence
    purpose: Predict height- and environment-dependent fusion strengths η_up/down and regularize them by discrepancy ranking.
    Introduced in §III-C3 and Algorithm 1; no external validation outside this paper’s metrics.
  • LAIN-Radio3D dataset independent evidence
    purpose: Provide large-scale 3D radio maps with Rx heights to 120 m and variable Tx heights for low-altitude scenarios.
    New simulated corpus (5.07M slices) released by the authors; useful artifact but still RT-based labels.

pith-pipeline@v1.1.0-grok45 · 23755 in / 3936 out tokens · 38294 ms · 2026-07-14T15:50:45.385411+00:00 · methodology

0 comments
read the original abstract

Three-dimensional (3D) radio map (RM) is a key enabler for environment-aware communications in low-altitude wireless scenarios by providing site-specific channel priors indexed by spatial locations. However, existing 3D RM construction methods lack effective modeling of the height dimension, which limits their generalization to unseen height configurations and degrades construction coherence across height layers. In this paper, we propose BDFlow-3DRM, a bi-dynamical flow matching framework for 3D RM construction. Specifically, the 3D RM construction problem is formulated as a deterministic probability flow in a semantic latent space. BDFlow-3DRM learns continuous height-aware representations from flexible transceiver height inputs, enhancing geometric awareness. Meanwhile, its bi-dynamical design explicitly models bidirectional dependencies across neighboring height layers, so that RMs at different height layers can be constructed jointly. Extensive experiments on multiple datasets, covering diverse simplified and realistic urban scenarios, validate the effectiveness of BDFlow-3DRM. Compared with diffusion-based baselines, it reduces the normalized mean square error (NMSE) by 28.6% and attains a 180-fold reduction in inference complexity. More importantly, with only 20 training receiver-height layers, BDFlow-3DRM maintains accurate prediction over a wide continuous receiver-height range from 1 to 120 m under variable transmitter heights, highlighting its practical potential for large-scale 3D RM construction.

Figures

Figures reproduced from arXiv: 2607.09778 by Jun Yu, Meixia Tao, Shu Sun.

Figure 1
Figure 1. Figure 1: The framework formulates 3D RM construction as a [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: Overview of the proposed BDFlow-3DRM. step, where N denotes the total number of discrete height layers available in datasets, the model randomly selects a fixed subset containing D height layers from the complete height pool as the training input. This sampling strategy exposes the model to diverse combinations of height configu￾rations, thereby encouraging the learning of signal distribution characteristi… view at source ↗
Figure 2
Figure 2. Figure 2: Network architecture for the Bi-Dynamical Block [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual comparison of 3D RM construction results at different heights [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Vertical profile comparison for a representative 3D RM slice at [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Heatmap of the predicted fusion coefficient [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of zero-shot generalization on unseen Rx heights across the urban scenario. The annotation [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of zero-shot generalization on unseen Tx heights. From left to right, the paired columns represent zero-shot inference for Tx deployed at [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Few-shot generalization performance of BDFlow-3DRM in the SF [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of the SF evaluation scenario: (a) the fine-grained 3D [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

31 extracted references · 5 linked inside Pith

  1. [1]

    Engineering radio maps for wireless resource management,

    S. Biet al., “Engineering radio maps for wireless resource management,” IEEE Wireless Communications, vol. 26, no. 2, pp. 133–141, 2019

    2019

  2. [2]

    Accessing from the sky: A tutorial on uav communications for 5G and beyond,

    Y . Zeng, Q. Wu, and R. Zhang, “Accessing from the sky: A tutorial on uav communications for 5G and beyond,”Proceedings of the IEEE, vol. 107, no. 12, pp. 2327–2375, 2019

    2019

  3. [3]

    Radio map estimation: A data-driven approach to spectrum cartography,

    D. Romero and S.-J. Kim, “Radio map estimation: A data-driven approach to spectrum cartography,”IEEE Signal Processing Magazine, vol. 39, no. 6, pp. 53–72, 2022

    2022

  4. [4]

    Toward environment-aware 6G communications via channel knowledge map,

    Y . Zeng and X. Xu, “Toward environment-aware 6G communications via channel knowledge map,”IEEE Wireless Communications, vol. 28, no. 3, pp. 84–91, 2021

    2021

  5. [5]

    Csi-tuples-based 3-d channel fingerprints construction assisted by multimodal learning,

    C. Xieet al., “Csi-tuples-based 3-d channel fingerprints construction assisted by multimodal learning,”IEEE Transactions on Wireless Com- munications, vol. 25, pp. 17 369–17 383, 2026

    2026

  6. [6]

    The design and applications of high-performance ray- tracing simulation platform for 5G and beyond wireless communica- tions: A tutorial,

    D. Heet al., “The design and applications of high-performance ray- tracing simulation platform for 5G and beyond wireless communica- tions: A tutorial,”IEEE Commun. Surveys Tuts., vol. 21, no. 1, pp. 10–27, Aug. 2019

    2019

  7. [7]

    RadioUNet: Fast radio map estimation with convolutional neural networks,

    R. Levie, C ¸ . Yapar, G. Kutyniok, and G. Caire, “RadioUNet: Fast radio map estimation with convolutional neural networks,”IEEE Trans. Wireless Commun., vol. 20, no. 6, pp. 4001–4015, Jun. 2021

    2021

  8. [8]

    RME-GAN: A learning framework for radio map estimation based on conditional generative adversarial network,

    S. Zhanget al., “RME-GAN: A learning framework for radio map estimation based on conditional generative adversarial network,”IEEE Internet Things J., vol. 10, no. 20, pp. 18 016–18 027, Oct. 2023

    2023

  9. [9]

    RadioDiff: An effective generative diffusion model for sampling-free dynamic radio map construction,

    X. Wanget al., “RadioDiff: An effective generative diffusion model for sampling-free dynamic radio map construction,”IEEE Trans. Cogn. Commun. Netw., vol. 11, no. 2, pp. 738–750, Apr. 2025

    2025

  10. [10]

    CKMDiff: A generative diffusion model for CKM construction via inverse problems with learned priors,

    S. Fu, Y . Zeng, Z. Wu, D. Wu, J. Xu, S. Jin, C.-X. Wang, and X. Gao, “CKMDiff: A generative diffusion model for CKM construction via inverse problems with learned priors,”npj Wireless Technology, vol. 2, no. 1, p. 35, 2026

    2026

  11. [11]

    RadioFlow: Efficient radio map construction framework with flow matching,

    H. Jiaet al., “RadioFlow: Efficient radio map construction framework with flow matching,”arXiv preprint arXiv:2510.09314, 2025

    arXiv 2025

  12. [12]

    UniRM: A universal large model for multiband 3D radio map construction,

    X. Jianget al., “UniRM: A universal large model for multiband 3D radio map construction,”IEEE Journal on Selected Areas in Communications, vol. 44, pp. 2244–2258, 2026

    2026

  13. [13]

    DF-3DRME: A data- friendly learning framework for 3d radio map estimation based on super- resolution technique,

    L. Zhu, W. Zhu, S. Zhang, G. Caire, and L. Liu, “DF-3DRME: A data- friendly learning framework for 3d radio map estimation based on super- resolution technique,”arXiv preprint arXiv:2604.00676, 2026

    arXiv 2026

  14. [14]

    3-D radio map estimation based on active measurement trajectory selection,

    Z. Chen, H. Wang, and D. Guo, “3-D radio map estimation based on active measurement trajectory selection,”IEEE Wireless Commun. Lett., vol. 14, no. 7, pp. 1884–1888, July 2025

    2025

  15. [15]

    3D radio map reconstruction based on generative adver- sarial networks under constrained aircraft trajectories,

    T. Huet al., “3D radio map reconstruction based on generative adver- sarial networks under constrained aircraft trajectories,”IEEE Trans. Veh. Technol., vol. 72, no. 6, pp. 8250–8255, June 2023

    2023

  16. [16]

    RadioGen3D: 3D radio map generation via adversarial learning on large-scale synthetic data,

    J. Chenet al., “RadioGen3D: 3D radio map generation via adversarial learning on large-scale synthetic data,”arXiv preprint arXiv:2602.18744, 2026

    arXiv 2026

  17. [17]

    3DCKMDiff: 3D CKM construction based on gener- ative diffusion for environment-aware low-altitude communication and sensing,

    H. Ronget al., “3DCKMDiff: 3D CKM construction based on gener- ative diffusion for environment-aware low-altitude communication and sensing,” inProc. IEEE/CIC ICCC Workshops, 2025

    2025

  18. [18]

    3D-RadioDiff: An altitude-conditioned diffusion model for 3D radio map construction,

    L. Zhao, Z. Fei, X. Wang, J. Luo, and Z. Zheng, “3D-RadioDiff: An altitude-conditioned diffusion model for 3D radio map construction,” IEEE Wireless Commun. Lett., vol. 14, no. 7, pp. 1969–1973, July 2025

    1969

  19. [19]

    RadioDiff-3D: A 3D×3D radio map dataset and generative diffusion based benchmark for 6G environment-aware com- munication,

    X. Wanget al., “RadioDiff-3D: A 3D×3D radio map dataset and generative diffusion based benchmark for 6G environment-aware com- munication,”IEEE Trans. Netw. Sci. Eng., vol. 13, pp. 3773–3789, 2026

    2026

  20. [20]

    A fine-grained 3D radio map construction paradigm with ultra-low sampling rates by large generative models,

    Z. Liuet al., “A fine-grained 3D radio map construction paradigm with ultra-low sampling rates by large generative models,”IEEE Journal on Selected Areas in Communications, vol. 44, pp. 4397–4413, 2026

    2026

  21. [21]

    PILOT: One physics-integrated generation framework to unify 2d and 3d radio map construction,

    W. Huang, H. Sun, and J. Chen, “PILOT: One physics-integrated generation framework to unify 2d and 3d radio map construction,”arXiv preprint arXiv:2604.23533, 2026

    Pith/arXiv arXiv 2026

  22. [22]

    FARM: Foundational aerial radio map for intelligent low-altitude networking,

    S. Gao, J. Liang, Y . Yuan, W. Lu, G. Shen, and L. Yang, “FARM: Foundational aerial radio map for intelligent low-altitude networking,” arXiv preprint arXiv:2604.17362, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    6D channel knowledge map construction via bidirectional wireless Gaussian splatting,

    J. Zhou, C. Hu, G. Wu, Z. Ren, H. Hu, J. Zhang, R. Zhang, and J. Xu, “6D channel knowledge map construction via bidirectional wireless Gaussian splatting,”arXiv preprint arXiv:2510.26166, 2025

    arXiv 2025

  24. [24]

    LAMBDA: A low-altitude multimodal base dataset for uav sensing and communication,

    L. Zhou, P. Rao, C. Zhang, J. Mo, S. Sun, Z. Chen, and M. Tao, “LAMBDA: A low-altitude multimodal base dataset for uav sensing and communication,”arXiv preprint arXiv:2607.03826, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inProc. Adv. Neural Inf. Process. Syst., vol. 33, 2020, pp. 6840–6851

    2020

  26. [26]

    Flow matching for generative modeling,

    Y . Lipmanet al., “Flow matching for generative modeling,” inProc. 11th Int. Conf. Learn. Representations, 2023

    2023

  27. [27]

    Latent diffusion model without variational autoencoder,

    M. Shiet al., “Latent diffusion model without variational autoencoder,” arXiv preprint arXiv:2510.15301, 2026

    arXiv 2026

  28. [28]

    Sim ´eoniet al., “Dinov3,”arXiv preprint arXiv:2508.10104, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Dynamical diffusion: Learning temporal dynamics with diffusion models,

    X. Guoet al., “Dynamical diffusion: Learning temporal dynamics with diffusion models,”arXiv preprint arXiv:2503.00951, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Dominant path prediction model for urban scenarios,

    R. Wahl, G. W ¨olfle, P. Wertz, P. Wildbolz, and F. Landstorfer, “Dominant path prediction model for urban scenarios,” in14th IST Mobile and Wireless Communications Summit, 2005, pp. 1–5

    2005

  31. [31]

    Sionna RT: Differentiable ray tracing for radio prop- agation modeling,

    J. Hoydiset al., “Sionna RT: Differentiable ray tracing for radio prop- agation modeling,” inProc. IEEE Globecom Workshops (GC Wkshps), Dec. 2023, pp. 317–321

    2023