Dynamical Multimodal Fusion with Mixture-of-Experts for Localizations

Bohao Wang; Chongwen Huang; Fenghao Zhu; Merouane Debbah; Qianqian Yang; Sami Muhaidat; Yongliang Shen; Zhaoyang Zhang; Zitao Shuai

REVIEW 5 major objections 6 minor 1 cited by

Dynamical Multimodal Fusion with Mixture-of-Experts for Localizations

T0 review · 5 major / 6 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 A soft mixture-of-experts fusion network keeps wireless localization sub-meter across three carrier bands.

desk verdict A solid MoE-based fusion architecture for ISAC localization with a promising OOD frequency test, but the trajectory split rule allows memory leakage that could inflate the headline gains. read the letter →

arxiv 2507.01337 v1 pith:6Q7CUUN6 submitted 2025-07-02 cs.IT eess.SPmath.IT

Bohao Wang , Zitao Shuai , Fenghao Zhu , Chongwen Huang , Yongliang Shen , Zhaoyang Zhang , Qianqian Yang , Sami Muhaidat

show 1 more author

Merouane Debbah

This is my paper

classification cs.ITeess.SPmath.IT

keywords 6Glocalizationintegratedsensingandcommunicationsfingerprintingmixture-of-expertsmultimodalfusionNLOSmitigationfrequencyrobustnessray-tracingsimulation

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

SCADF-MoE is proposed to solve two blockers for sub-meter 6G localization: the usefulness of each signal modality shifts with carrier frequency, and similar-looking fingerprints, especially in non-line-of-sight (NLOS) conditions, produce ambiguous position estimates. The paper's claim is that converting isolated fingerprint samples into short spatial trajectories and fusing channel-state information, angle-of-arrival, distance, and gain through learnable soft mixture-of-experts routers resolves both problems. On ray-tracing data from three urban layouts at 2.6, 6, and 28 GHz, the model reports average mean-squared error 63% lower and unseen-NLOS error 55% lower than the strongest prior fusion strategy, staying consistently below the $1\,\mathrm{m}^2$ sub-meter target. The work matters because it points to a fusion design that keeps localization accurate when the carrier band changes, without per-band retraining.

What carries the argument

The load-bearing mechanism is the soft mixture-of-experts fusion layer, in which a learnable routing matrix produces normalized input-to-expert dispatch weights and expert-to-output combination weights for every sample, so each token is a weighted mixture of all expert outputs rather than a hard assignment. Around that core, radius-based clustering turns isolated fingerprints into length-s trajectories; a modality-task MoE with s+1 task-specific routers shares the same experts but biases them with task embeddings; and a maximum-mean-discrepancy regularizer (a kernel-based distance between routing distributions) enforces diversity between dispatch and combine matrices. Together these components let the network dynamically favor the most informative modality per band and exploit inter-point correlations to resolve fingerprint twins.

What would settle it

Collect real over-the-air CSI, angle-of-arrival, distance, and gain measurements in the same three urban layouts at 2.6, 6, and 28 GHz, then train and test SCADF-MoE on those measurements; if average MSE rises above $1\,\mathrm{m}^2$ or a model trained on two bands loses sub-meter accuracy on the held-out band, the ray-tracing fidelity assumption would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that fusion for wireless fingerprint localization should be both frequency-adaptive and spatially structured, and that a soft mixture-of-experts backbone delivers both. The model's fusion layer learns sample-dependent dispatch and combination weights from a routing matrix, so the relative weight of CFR image tokens versus geometric scalars changes with the carrier band and propagation environment. A second mixture-of-experts layer with task-specific routers jointly regresses every point in a short neighborhood trajectory plus its centroid, so a single ambiguous fingerprint becomes a multi-task spatial constraint. An auxiliary maximum-mean-discrepancy loss keeps the task routers from collapsing onto identical experts, which the paper argues prevents gradient interference and stabilizes multi-task training. The reported outcome is consistent sub-meter MSE across three scenes and three bands, with a 63% average error reduction and a 55% reduction on unseen-NLOS error relative to the best baseline, and the authors position it as the first large-scale multimodal MoE for frequency-robust ISAC localization.

Load-bearing premise

The load-bearing premise is that the ray-tracing simulations of the three urban scenes at 2.6, 6, and 28 GHz faithfully represent real urban propagation, including the NLOS multipath that creates ambiguous fingerprints; if the simulated data are more separable or less noisy than physical measurements, the reported sub-meter accuracy and frequency robustness may not carry over to deployment.

Editorial extensions

If this is right

With soft-MoE routing, the model stays below $1\,\mathrm{m}^2$ average MSE across dense urban, suburban, and urban-canyon scenes, whereas static concat and fully connected fusion baselines exceed it.
On the out-of-distribution split (train on 2.6 and 28 GHz, test on 6 GHz), the error remains below $1\,\mathrm{m}^2$ for SCADF-MoE while every compared baseline degrades sharply, indicating frequency-robust fusion.
Removing spatial-context sampling raises MSE by roughly 3 to 5 times, so converting isolated fingerprints into short trajectories is a necessary part of the method's accuracy.
Ablating the MMD diversity loss or collapsing the router to a single expert more than doubles error, especially in NLOS, so expert diversity is needed for stable multi-task training.
The 63% average MSE reduction and 55% unseen-NLOS reduction are reported relative to the strongest baseline SoftMoE, so the contribution is the combination of spatial context, task-specific routing, and diversity regularization, not any single component.

Reading between the lines

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

Beyond the paper: if real measurements reproduce the ray-tracing results, the same trajectory-plus-task-router formulation should transfer to other fingerprint modalities such as time-of-arrival or received signal strength without architectural changes.
Beyond the paper: the out-of-distribution result implies the router learns a frequency-continuous weighting; testing on an intermediate band absent from training, such as 10 GHz, would show whether that generalization is continuous.
Beyond the paper: inspecting the learned dispatch weights per band could reveal which modality the network actually trusts (for example, distance and AoA in open areas versus CFR in dense multipath), giving an interpretability check on the fusion.

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.

Desk Editor's Note

read the letter

The genuinely new thing here is the combination: soft MoE with task-specific routers, applied to multimodal wireless fingerprint fusion, plus trajectory-level spatial context via short point clusters and an MMD regularizer to keep experts diverse. That is a legitimate extension of the SoftMoE line into a practical ISAC localization problem, and the evaluation design is mostly right: three urban scenes, three carrier bands, and a held-out 6 GHz band for OOD testing. If the numbers hold, this is a useful engineering contribution for 6G positioning, not a revolution.

The main soft spot is the trajectory split rule. Section II-B says a test trajectory is accepted if it 'contains at least one point that never appears in training.' With s=5 and cluster radius 2 m, that means up to four of the five test points can be training points. Since the model jointly regresses all vertices and the centroid, the seen points provide strong spatial context for inferring the one unseen vertex—exactly the correlation the method is designed to exploit. The 'unseen-NLOS' gains and the 63%/55% headline numbers could be inflated by memorization rather than true generalization. This is a real internal-validity threat, and the text as written does not rule it out. If the split were actually disjoint (no shared physical points between train and test trajectories), the concern disappears, but the authors need to say that explicitly.

The other issues are less severe but worth fixing. The numbers in Table II are in units of 10^-2 m^2, which conflicts with the text citing 0.56 m^2 for the Mix setting. The paper promises a hyperparameter sensitivity analysis in Section IV but never delivers it. And despite claiming 95% confidence intervals are within marker size, no error bars or values appear anywhere. The reliance on synthetic ray-tracing data is a limitation—the authors acknowledge it in the future-work sentence—but it is not fatal for a methods paper.

On the positive side, the architecture is clearly described, the baselines are reasonable (Concat, FullCon, TransFusion, SoftMoE, all with the same spatial-context preprocessing), and the OOD frequency experiment is a good idea. The citation pattern looks fine; the self-citations are relevant prior work on multi-point NLOS localization.

This paper deserves a serious referee. The core idea is sound and the evaluation is thoughtful, but the trajectory split rule needs to be examined and likely fixed before the headline claims can be trusted. If that leakage is closed and the unit inconsistency and missing sensitivity analysis are addressed, it would be a solid conference or journal contribution.

Referee Report

5 major / 6 minor

Summary. The paper proposes SCADF-MoE, a soft mixture-of-experts architecture for multimodal wireless fingerprint localization. The method converts isolated ray-traced channel measurements at nearby points into short trajectories, adaptively fuses CFR and scalar geometric features through soft-MoE routers, and jointly regresses all trajectory vertices plus the trajectory centroid using task-specific routers, with an MMD regularizer for expert diversity. Experiments on the W AIR-D corpus extended to 2.6, 6, and 28 GHz across three urban scenes report sub-meter MSE, large gains over Concat, FullCon, TransFusion, and SoftMoE baselines, and improved performance on an unseen 6 GHz band when training on 2.6 and 28 GHz.

Significance. If the reported results hold, the paper makes a useful contribution to frequency-adaptive multimodal fusion for RF localization: the trajectory-level multi-task formulation is a sensible way to inject spatial context, the soft-MoE router provides a principled mechanism for per-sample modality weighting, and the MMD-diversity loss is a reasonable stabilizer. The paper also builds a fairly comprehensive simulation pipeline over multiple scenes and carrier bands, and it compares against several representative fusion baselines under matched encoders and heads. The main limitations are that all evidence is synthetic, the train/test trajectory split as described is not clearly spatially disjoint, the OOD-frequency experiment does not state whether test locations overlap training locations, and the reported units and hyperparameter-sensitivity claims are inconsistent with the text. These issues need to be resolved before the central generalization claim can be accepted.

major comments (5)

[II-B] The train/test trajectory construction is not spatially disjoint as written. Section II-B says a test trajectory is accepted only if it contains at least one point that never appears in training. With trajectory length s=5, a test trajectory can therefore contain up to four points that did appear in training. Because the model jointly regresses all five vertices and the centroid via the loss Lcoord in Section III-C, a memorized training point can supply strong spatial context for predicting the remaining unseen vertex. This directly threatens the internal validity of the headline gains in Fig. 3 and the 63%/55% reductions claimed in the conclusion. Please clarify whether every vertex of every test trajectory is excluded from the training set; if not, rerun the evaluation with a fully disjoint spatial split and report whether the conclusions change.
[IV-C] The OOD frequency experiment is confounded by possible location overlap. The paper trains on the 2.6 and 28 GHz bands and tests on 6 GHz, but it does not state whether the same physical locations appear in the training bands and the test band. If the same grid points are used, the model can memorize location-specific geometry from the 2.6/28 GHz training data and then rely on that memorized geometry even while receiving only 6 GHz fingerprints at test time, making the "unseen band" claim much weaker. Please specify whether the OOD test locations are disjoint from all training locations across bands, and if they are not, rerun the OOD experiment with a location-disjoint split.
[Table II and IV-C] There is a unit inconsistency in the main quantitative results. Table II is captioned "errors for antenna-array mix (×10^-2 m^2)", but the text in Section IV-C reports SCADF-MoE overall MSE = 0.56 m^2 for scene 00743, whereas Table II lists 0.34 in that cell. If the caption is literal, the table value would be 0.0034 m^2, which contradicts the text and Fig. 3. Please correct the units or the values and recompute the percentage reductions, since these numbers are load-bearing for the paper's central claim.
[IV] The paper promises a sensitivity analysis but does not deliver one. Section IV lists "(iii) sensitivity to key hyper-parameters" as one of the four evaluation angles, yet no sensitivity results appear in Section IV-D; the ablation study removes losses and routers but does not vary alpha, cluster radius r, trajectory length s, or minimum cluster size n_min. These parameters are free choices in Algorithm 1 and Table I, so the claimed consistency of the method is not yet supported. Please add the missing sensitivity experiments or explicitly remove the claim from the evaluation plan.
[II-A and IV-A] All results are based on ray-traced synthetic data, and the paper describes the scenes as "real urban layouts" without acknowledging the distinction between real geometry and simulated propagation. The central claim of sub-meter accuracy and NLOS robustness depends on the fidelity of the ray tracer, particularly for NLOS multipath. If the simulated fingerprints are more separable or less noisy than physical measurements, the reported gains may not transfer. Please add an explicit limitations paragraph, and where feasible validate on measured data or on ray traces with added noise and model perturbation; at minimum, the title and text should not imply experimental measurements.

minor comments (6)

[Fig. 2] The caption and axis label contain the typo "Spacial"; it should read "Spatial".
[III-A and III-D] In Section III-A, Step 3, there is a stray word "exactly." at the end of the sentence defining the dispatch and combination weights; it should be removed. In Section III-D, the kernel k(x,y) and K(x,y) are defined twice with the same expression; please consolidate the two definitions.
[IV-A] The text says "All results average five independent runs; 95% confidence intervals are within the marker size," but no confidence-interval values or visible error bars are reported in any figure or table. Please provide the CI values or state explicitly that they are omitted.
[IV-A] The baseline description is internally inconsistent: it says all baselines are augmented with the same spatial-context sampling, but the SoftMoE bullet says "no spatial-context awareness is built in." Please clarify whether the baselines receive trajectory-length inputs or only single-point fingerprints, since this affects the interpretation of the comparison.
[Conclusion] The conclusion reports "cuts average MSE by 63%," while Section IV-C reports 61.1% for the Mix setting on scene 00743 and Table II implies a different average across scenes. Please define the averaging convention precisely and make the numbers consistent across the abstract, figures, table, and conclusion.
[II-B] The sentence "From the training points subset, we draw length-s trajectories" appears to contradict the next sentence about test trajectories containing unseen points. Likely the first phrase should read "testing points subset" or the sampling procedure should be described more carefully; please fix this ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported improvements are empirical comparisons, not derivations from the fitted inputs or from a self-citation chain.

full rationale

This is an empirical machine-learning paper rather than a formal derivation, so circularity can only arise if a reported 'prediction' is, by construction, identical to the data or parameters used to make it. That is not the case here. The training objective (Eq. 5: L = L_coord + alpha L_MMD), the soft-MoE fusion blocks, the modality-task MoE heads, and the trajectory-sampling procedure are all specified before evaluation and do not incorporate the test-set MSE. In the OOD experiment the model is trained on 2.6 and 28 GHz and only then tested on 6 GHz, so the frequency-robustness claim is not fitted to the tested band. Baselines are given identical encoders, heads, and spatial-context sampling, so the reported 63% MSE reduction and 55% unseen-NLOS reduction are comparative measurements against fixed alternatives, not identities. The rule that a test trajectory is accepted only if it 'contains at least one point that never appears in training' could weaken the strength of the word 'unseen' as a generalization claim, because up to s-1 points in a length-s test trajectory may have been seen in training; however, that is a data-split and leakage concern, not a circular reduction of a prediction to a fitted input. The self-citations ([4], [8], [10]) and the W AIR-D dataset citation ([16]) supply components, baselines, or data, but none of the central conclusions is justified solely by those citations, and no uniqueness theorem or ansatz is imported from the authors' prior work in a way that forces the result. No specific circular step can be quoted from the paper, so the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical entity. Its central results rest on several hand-set hyperparameters (alpha, r, s, nmin, expert count) and on the fidelity of synthetic ray-tracing, none of which is independently validated.

free parameters (5)

MMD regularization weight alpha = 0.1
Set empirically in Section III-D and Table I to balance diversity and accuracy; no sensitivity analysis is shown.
Cluster radius r = 2 m
Chosen as the smallest radius with median cluster size above a threshold; Table I fixes 2 m but the selection threshold is not reported.
Trajectory length s = 5
Fixed in Table I and central to the multi-task trajectory formulation.
Minimum cluster size nmin = not reported
Threshold used to select r* in Section II-B, not quantified in the paper.
Expert count and hidden dimension = not reported
Architecture hyperparameters for the Soft MoE stack and Modality-Task MoE are not listed, yet they determine capacity and routing behavior.

assumptions (3)

domain assumption Ray-tracing channel synthesis is a faithful proxy for real 2.6/6/28 GHz urban propagation, including NLOS multipath.
All fingerprints in Section II-A and IV-A are generated by a ray-tracer; if the simulation is cleaner than physical channels, the absolute errors and OOD conclusions are optimistic.
domain assumption Trajectory splitting with at least one unseen point per test trajectory prevents meaningful spatial leakage between training and test.
Section II-B; if neighboring points in the same 2 m cluster leak geometry into the test trajectory, the reported per-trajectory errors understate per-point localization error.
domain assumption The three urban scenes and the held-out 6 GHz band represent the deployment distribution.
Section IV-C uses only one unseen band; generalization to other bands and other physical environments is assumed rather than demonstrated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Dynamical Multimodal Fusion with Mixture-of-Experts for Localizations." pith.science (2026). https://pith.science/paper/6Q7CUUN6

@misc{pith2026250701337,
  author       = {Pith},
  title        = {Pith review of: Dynamical Multimodal Fusion with Mixture-of-Experts for Localizations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6Q7CUUN6}},
  note         = {Machine review of arXiv:2507.01337}
}

read the original abstract

Multimodal fingerprinting is a crucial technique to sub-meter 6G integrated sensing and communications (ISAC) localization, but two hurdles block deployment: (i) the contribution each modality makes to the target position varies with the operating conditions such as carrier frequency, and (ii) spatial and fingerprint ambiguities markedly undermine localization accuracy, especially in non-line-of-sight (NLOS) scenarios. To solve these problems, we introduce SCADF-MoE, a spatial-context aware dynamic fusion network built on a soft mixture-of-experts backbone. SCADF-MoE first clusters neighboring points into short trajectories to inject explicit spatial context. Then, it adaptively fuses channel state information, angle of arrival profile, distance, and gain through its learnable MoE router, so that the most reliable cues dominate at each carrier band. The fused representation is fed to a modality-task MoE that simultaneously regresses the coordinates of every vertex in the trajectory and its centroid, thereby exploiting inter-point correlations. Finally, an auxiliary maximum-mean-discrepancy loss enforces expert diversity and mitigates gradient interference, stabilizing multi-task training. On three real urban layouts and three carrier bands (2.6, 6, 28 GHz), the model delivers consistent sub-meter MSE and halves unseen-NLOS error versus the best prior work. To our knowledge, this is the first work that leverages large-scale multimodal MoE for frequency-robust ISAC localization.

Figures

Figures reproduced from arXiv: 2507.01337 by the authors.

**Figure 1.** Overview of our SCADF-MoE method. channel signatures while still being distinguishable at a global scale [11], suggesting that localization should be formulated as a multi-task problem that jointly regresses the coordinates of all points in a short trajectory. SCADF-MoE (i) uses soft MoE routers to re-weight CSI, AoA, and related cues on a per-sample basis, automatically adapting to carrier-frequency changes, and (i… view at source ↗

**Figure 2.** Effectiveness of spatial context pre-processing. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** MSE across carrier bands: (top) Mix, (bottom) OOD. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Ablation experiment of SCADF-MoE in antenna mix. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

Variational Mixture of Graph Neural Experts for Alzheimer's Disease Recognition across Frequency Bands in EEG Brain Networks
cs.LG 2025-10 conditional novelty 4.0 of 10

VMoGE, a variational mixture of per-frequency-band graph experts, reports AUC up to 0.89 for Alzheimer's vs. healthy EEG and links learned band weights to known dementia markers.

Reference graph

Works this paper leans on

19 extracted references · 13 canonical work pages · cited by 1 Pith paper

[1]

Wireless large AI model: Shaping the AI-native future of 6G and beyond,

F. Zhu, X. Wang, X. Li, M. Zhang, Y . Chen, and etc., “Wireless large AI model: Shaping the AI-native future of 6G and beyond,” [Online] available: https://arXiv.org/abs/2504.14653 , Apr. 2025

arXiv 2025
[2]

Large- scale AI in telecom: Charting the roadmap for innovation, scalability, and enhanced digital experiences,

A. Shahid, A. Kliks, A. Al-Tahmeesschi, A. Elbakary, and etc., “Large- scale AI in telecom: Charting the roadmap for innovation, scalability, and enhanced digital experiences,” [Online] available: https://arXiv.org/ abs/2503.04184, Mar. 2025

arXiv 2025
[3]

Robust deep learning-based physical layer communications: Strategies and approaches,

F. Zhu, X. Wang, C. Zhu, T. Gong, Z. Yang, and etc., “Robust deep learning-based physical layer communications: Strategies and approaches,” IEEE Netw., pp. 1–1, May. 2025

work page 2025
[4]

Multi-sources information fusion learning for multi- points NLOS localization,

B. Wang, F. Zhu, M. Liu, C. Huang, Q. Yang, A. Alhammadi, Z. Zhang, and M. Debba, “Multi-sources information fusion learning for multi- points NLOS localization,” in Proc. Veh. Tech. Conf. (VTC), Jun. 2024

work page 2024
[5]

Fingerprint-based localization for massive MIMO-OFDM system with deep convolutional neural networks,

X. Sun, C. Wu, X. Gao, and G. Y . Li, “Fingerprint-based localization for massive MIMO-OFDM system with deep convolutional neural networks,” IEEE Trans. Veh. Technol., vol. 68, no. 11, pp. 10 846–10 857, Sept. 2019

work page 2019
[6]

Sub-THz Channel Measurements at 158 GHz and 300 GHz in a Street Canyon Environment

W. Keusgen, A. Schultze, M. Peter, and T. Eichler, “Sub-THz channel measurements at 158 GHz and 300 GHz in a street canyon environment,” [Online] available: https://arxiv.org/abs/2203.04404 , Mar. 2022

work page Pith review arXiv 2022
[7]

Achieving centimeter- accuracy indoor localization on WiFi platforms: A multi-antenna ap- proach,

C. Chen, Y . Chen, Y . Han, H.-Q. Lai, and etc., “Achieving centimeter- accuracy indoor localization on WiFi platforms: A multi-antenna ap- proach,” IEEE Internet Things J., vol. 4, no. 1, pp. 122–134, Nov. 2016

work page 2016
[8]

Dynamic modeling of patients, modalities and tasks via multi-modal multi-task mixture of experts,

C. Wu, Z. Shuai, Z. Tang, L. Wang, and L. Shen, “Dynamic modeling of patients, modalities and tasks via multi-modal multi-task mixture of experts,” in Proc. Int. Conf. Learn. Represent. (ICLR) , Jan. 2025

work page 2025

Show all 19 references

[9]

A belief propagation algorithm for multipath-based SLAM,

E. Leitinger, F. Meyer, F. Hlawatsch, K. Witrisal, F. Tufvesson, and M. Z. Win, “A belief propagation algorithm for multipath-based SLAM,” IEEE Trans. Wireless Commun., vol. 18, no. 12, pp. 5613–5629, Sept. 2019

2019
[10]

Multi-sources fusion learning for multi-points NLOS localization in OFDM system,

B. Wang, Z. Shuai, C. Huang, Q. Yang, Z. Yang, R. Jin, A. Al Hammadi, Z. Zhang, C. Yuen, and M. Debbah, “Multi-sources fusion learning for multi-points NLOS localization in OFDM system,” IEEE J. Sel. Topics Signal Process., vol. 18, no. 7, pp. 1339 – 1350, Sept. 2024

2024
[11]

MoLoc: On distinguishing fingerprint twins,

W. Sun, J. Liu, C. Wu, Z. Yang, X. Zhang, and Y . Liu, “MoLoc: On distinguishing fingerprint twins,” in Proc. IEEE Inter. Conf. Distrib. Compu. Sys. (ICDCS) , Dec. 2013

2013
[12]

Localization approach based on ray-tracing simulations and fingerprinting techniques for indoor–outdoor scenarios,

A. Del Corte-Valiente, J. M. G ´omez-Pulido, O. Guti ´errez-Blanco, and J. L. Castillo-Sequera, “Localization approach based on ray-tracing simulations and fingerprinting techniques for indoor–outdoor scenarios,” Energies, vol. 12, no. 15, p. 2943, Jul. 2019

2019
[13]

Wireless localization with spatial-temporal robust fingerprints,

D. Li, J. Xu, Z. Yang, C. Wu, J. Li, and N. D. Lane, “Wireless localization with spatial-temporal robust fingerprints,” ACM Trans. Sens. Netw., vol. 18, no. 1, pp. 1–23, Oct. 2021

2021
[14]

A wireless fingerprint positioning method based on wavelet transform and deep learning,

D. Li and Z. Niu, “A wireless fingerprint positioning method based on wavelet transform and deep learning,” ISPRS Int. J. Geo-inf , vol. 10, no. 7, p. 442, Mar. 2021

2021
[15]

Multiple-camera people localization in an indoor environment,

V . A. Petrushin, G. Wei, and A. V . Gershman, “Multiple-camera people localization in an indoor environment,” Knowl. Inf. Syst., vol. 10, no. 2, pp. 229–241, May. 2006

2006
[16]

W AIR-D: Wireless AI research dataset,

Y . Huangfu, J. Wang, S. Dai, R. Li, J. Wang, C. Huang, and Z. Zhang, “W AIR-D: Wireless AI research dataset,” [Online] available: https:// arxiv.org/abs/2212.02159, Dec. 2022

2022 arXiv
[17]

On deep learning-based massive MIMO indoor user localization,

M. Arnold, S. Dorner, S. Cammerer, and etc., “On deep learning-based massive MIMO indoor user localization,” in Proc. Int. Workshop on Signal Process. Advances in Wireless Commu. (SPAWC) , Aug. 2018

2018
[18]

Swin-loc: Transformer-based CSI fingerprinting indoor localization with MIMO ISAC system,

X. Xu, F. Zhu, S. Han, Z. Yu, H. Zhao, B. Wang, and P. Zhang, “Swin-loc: Transformer-based CSI fingerprinting indoor localization with MIMO ISAC system,” IEEE Trans. Veh. Technol., Mar. 2024

2024
[19]

From sparse to soft mixtures of experts,

J. Puigcerver, C. Riquelme, B. Mustafa, and N. Houlsby, “From sparse to soft mixtures of experts,” [Online] available: https://arxiv.org/abs/ 2308.00951, Aug. 2023

2023 arXiv

Pith tools

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

[1] [1]

Wireless large AI model: Shaping the AI-native future of 6G and beyond,

F. Zhu, X. Wang, X. Li, M. Zhang, Y . Chen, and etc., “Wireless large AI model: Shaping the AI-native future of 6G and beyond,” [Online] available: https://arXiv.org/abs/2504.14653 , Apr. 2025

arXiv 2025

[2] [2]

Large- scale AI in telecom: Charting the roadmap for innovation, scalability, and enhanced digital experiences,

A. Shahid, A. Kliks, A. Al-Tahmeesschi, A. Elbakary, and etc., “Large- scale AI in telecom: Charting the roadmap for innovation, scalability, and enhanced digital experiences,” [Online] available: https://arXiv.org/ abs/2503.04184, Mar. 2025

arXiv 2025

[3] [3]

Robust deep learning-based physical layer communications: Strategies and approaches,

F. Zhu, X. Wang, C. Zhu, T. Gong, Z. Yang, and etc., “Robust deep learning-based physical layer communications: Strategies and approaches,” IEEE Netw., pp. 1–1, May. 2025

work page 2025

[4] [4]

Multi-sources information fusion learning for multi- points NLOS localization,

B. Wang, F. Zhu, M. Liu, C. Huang, Q. Yang, A. Alhammadi, Z. Zhang, and M. Debba, “Multi-sources information fusion learning for multi- points NLOS localization,” in Proc. Veh. Tech. Conf. (VTC), Jun. 2024

work page 2024

[5] [5]

Fingerprint-based localization for massive MIMO-OFDM system with deep convolutional neural networks,

X. Sun, C. Wu, X. Gao, and G. Y . Li, “Fingerprint-based localization for massive MIMO-OFDM system with deep convolutional neural networks,” IEEE Trans. Veh. Technol., vol. 68, no. 11, pp. 10 846–10 857, Sept. 2019

work page 2019

[6] [6]

Sub-THz Channel Measurements at 158 GHz and 300 GHz in a Street Canyon Environment

W. Keusgen, A. Schultze, M. Peter, and T. Eichler, “Sub-THz channel measurements at 158 GHz and 300 GHz in a street canyon environment,” [Online] available: https://arxiv.org/abs/2203.04404 , Mar. 2022

work page Pith review arXiv 2022

[7] [7]

Achieving centimeter- accuracy indoor localization on WiFi platforms: A multi-antenna ap- proach,

C. Chen, Y . Chen, Y . Han, H.-Q. Lai, and etc., “Achieving centimeter- accuracy indoor localization on WiFi platforms: A multi-antenna ap- proach,” IEEE Internet Things J., vol. 4, no. 1, pp. 122–134, Nov. 2016

work page 2016

[8] [8]

Dynamic modeling of patients, modalities and tasks via multi-modal multi-task mixture of experts,

C. Wu, Z. Shuai, Z. Tang, L. Wang, and L. Shen, “Dynamic modeling of patients, modalities and tasks via multi-modal multi-task mixture of experts,” in Proc. Int. Conf. Learn. Represent. (ICLR) , Jan. 2025

work page 2025