Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

LVM4CSI: Enabling Direct Application of Pre-Trained Large Vision Models for Wireless Channel Tasks

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Frozen pre-trained vision models, applied to CSI rendered as images, can match or outperform task-specific neural networks for wireless channel estimation and sensing without any fine-tuning.

desk verdict Frozen CV models transferring to CSI tasks is a real new application, but the zero-shot DINO-X channel-estimation claim needs much better documentation and stronger baselines before the headline numbers should be taken at face value. read the letter →

arxiv 2507.05121 v1 pith:XOI7IIAE submitted 2025-07-07 cs.IT cs.AIcs.CVcs.LGmath.IT

classification cs.ITcs.AIcs.CVcs.LGmath.IT
keywords largevisionmodelschannelstateinformationestimationhumanactivityrecognitionuserlocalizationzero-shottransferobjectdetectionCSI-to-CVmapping
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

LVM4CSI claims that large vision models trained on natural images can be dropped, frozen, into wireless channel problems and perform as well as or better than networks designed and trained specifically for those problems. The paper demonstrates this on channel estimation, human activity recognition, and user localization by turning CSI into images and translating each task into a vision analogue. If the claim holds, it would remove much of the expert design and large CSI-specific training data currently required for AI-based channel acquisition and sensing.

What carries the argument

The machinery is a three-step workflow: translate the CSI task into a CV task, transform complex CSI into a three-channel image, and attach a minimal trainable layer to the frozen LVM's output. The equality carrying channel estimation is Eq. (13), which maps the center $(w_l, h_l)$ of a detected bounding box to normalized angle $\hat{\Theta}_l = 1 - w_l/(\beta M)$ and delay $\hat{T}_l = h_l/(\gamma N)$; gains are then fit by least squares and the channel is rebuilt from the estimated triplets. For sensing tasks, the load-bearing object is the frozen ConvNeXt feature vector, which replaces a custom CNN feature extractor; the whole trainable part is one dense layer for activity recognition or four dense layers for localization.

What would settle it

Run the channel-estimation pipeline on synthetic angular-delay CSI with known ground-truth paths at low SNR and high path count, and compare DINO-X's detected bounding-box centers against the true angle-delay pairs; if missed paths or a systematic bias appears as path number or noise grows, the zero-shot estimation claim fails in the regime the paper highlights.

Watch

Extended reading notes

Core claim

The discovery is that cross-modal transfer from natural images to wireless channels does not need fine-tuning: a prompt-driven object detector (DINO-X) finds multipath components in angular-delay CSI images, and a frozen classifier (ConvNeXt) produces feature vectors that a tiny trainable head turns into activity labels or positions. In channel estimation, box centers give path angles and delays through Eq. (13), and least-squares fitting of gains yields the full channel, improving NMSE by more than 9.61 dB over LMMSE at low SNR. In sensing, frozen ConvNeXt features plus a dense layer reach 95.8 percent activity accuracy with roughly one thirtieth of the trainable parameters of a task-specific recurrent model, and reduce localization error by about 40 percent relative to a trained CNN feature extractor at 2.5 dB SNR. The paper argues these results follow from the structural similarity between sparse CSI images and natural images, amplified by the scale of LVM pretraining.

Load-bearing premise

The load-bearing premise is that a detector trained on natural images will find signal paths in noisy angular-delay CSI images from a prompt like 'bright spots', and that the centers of its boxes line up with the true angles and delays without a systematic bias that Eq. (13) cannot absorb.

Editorial extensions

If this is right

  • Channel estimation can be performed with no training data for the path extractor: the paper needs only 50 CSI samples for evaluation, because DINO-X is used as-is.
  • A single frozen feature extractor serves both classification and regression: only the appended head changes between activity recognition and localization.
  • Performance scales with LVM capacity: among ConvNeXt variants, the largest model gives the best CSI task accuracy, pointing to model scale as a direct performance lever.
  • The same recipe works on both simulated channel data and real WiFi CSI measurements, suggesting the transfer is not tied to one simulator or band.
  • For sensing tasks, the trainable parameter count drops by more than an order of magnitude relative to train-from-scratch networks, easing deployment on resource-limited devices.

Reading between the lines

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

  • Beyond the paper, the same prompt-based detection recipe could be applied to other sparse RF images, such as radar range-angle maps or mmWave point-cloud projections, as a zero-shot path or scatterer detector.
  • The paper leaves implicit that the LVM advantage is largest where conventional estimators struggle: low SNR and many paths; at high SNR the gap narrows, suggesting the transfer is most valuable in precisely the regimes where task-specific models need the most data.
  • Because the LVM discards absolute power information, the localization case shows a likely general principle: LVM features will often need to be supplemented by a few physically meaningful scalars (e.g., channel power) rather than by more learned layers.
  • A testable extension: fine-tuning only the prompt or the head on a small CSI-labeled set, while still freezing the backbone, would show whether the zero-shot success is a ceiling or a starting point.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes LVM4CSI, a framework that applies frozen, pre-trained large vision models (LVMs) to wireless channel tasks without fine-tuning the vision backbone. CSI tasks are translated into analogous computer-vision tasks, complex-valued CSI is converted into RGB images, and lightweight trainable heads are added where needed. Three case studies are presented: zero-shot channel estimation using DINO-X as an instruction-driven object detector on angular-delay CSI images, human activity recognition using frozen ConvNeXt features with a dense classifier, and user localization using frozen ConvNeXt features plus dense regression layers. The authors report a channel-estimation NMSE improvement exceeding 9.61 dB over LMMSE, HAR accuracy up to 95.8%, and a roughly 40% localization error reduction relative to a custom CNN baseline, with far fewer trainable parameters than task-specific networks.

Significance. If the zero-shot channel-estimation result is robust, the paper would make a notable contribution by showing that features learned from natural images transfer to CSI acquisition without any training of the vision backbone. The framework is clearly presented, the workflow is sensible, and the use of frozen LVM parameters with only tiny task heads is a practical strength. The reported low trainable-parameter counts are also attractive. However, the central quantitative claims rest on thin statistical evidence and on an underspecified DINO-X detection pipeline, so the significance is currently conditional on a more complete validation.

major comments (5)
  1. [Section IV-A, Eq. (13), and Section V-A1] The box-to-path-parameter mapping is underspecified, and this is load-bearing for the zero-shot channel-estimation claim. The paper does not state whether DINO-X outputs pixel coordinates or normalized [0,1] coordinates, does not give the prompt template, confidence threshold, non-maximum suppression, or any post-processing used to decide which detection boxes correspond to paths. Eq. (13) is only valid under a specific coordinate convention that is never stated. Without these details, the reader cannot reproduce the reported 9.61 dB gain or verify that the mapping is unbiased.
  2. [Section V-C1, Figs. 8 and 9] The channel-estimation evaluation uses only 50 CSI samples per setting and reports no error bars or statistical variability. Fig. 9 itself shows a missed path at SNR=0 dB, so the average results may be sensitive to a small number of favorable or unfavorable realizations. Moreover, the comparison is only against LS and LMMSE, not against task-specific NN estimators such as the YOLO-based path extractors cited in [29], [30] or learned denoisers such as [25]. The claim that LVM4CSI is 'comparable or superior to task-specific NNs' is not directly supported for channel estimation.
  3. [Section V-C2, Table II] The HAR result does not support the 'comparable or superior' wording in the abstract and conclusion: LVM4CSI's best accuracy is 95.8%, while the cited Recurrent ConFormer reaches 96.2%, and the only train-from-scratch comparisons are VGG19 and ResNet50. The statement that LVM4CSI achieves 'comparable or even superior accuracy' is therefore an overstatement for this task. Please temper the claim or compare against the broader set of task-specific HAR models used in the literature.
  4. [Section V-C3, Fig. 10] The localization evaluation compares LVM4CSI against two custom baselines (NoFeatExt-CSI and ConvFeatExt-CSI), not against existing CSI-based localization methods, and again no error bars or multiple runs are provided. The approximately 40% improvement is measured at one SNR setting against a simple hand-designed CNN, so the practical significance of the gain is unclear. Stronger baselines and confidence intervals are needed before claiming a general 40% localization-error reduction.
  5. [Section V-C2, paragraph after Table II, and Section VI] The sentence 'if LVMs like ConvNeXt are pre-trained on datasets that include some CSI, they can deliver competitive or superior performance with much less effort' is purely speculative and unsupported by any experiment in the paper. It also weakens the paper's core claim that no CSI-specific pre-training or fine-tuning is needed. Please remove this speculation or clearly separate it as a future direction rather than presenting it as a conclusion of the current results.
minor comments (5)
  1. [Abstract and Section IV] The phrase 'without any fine-tuning' is stronger than what the sensing case studies actually demonstrate: the LVM parameters are frozen, but the dense heads for HAR and localization are trained. Please rephrase the abstract and contributions to say the LVM backbone is frozen, while lightweight task heads are trained when needed.
  2. [Section V-A1 and Fig. 9] The number of test samples (50) is very small; a brief justification or a sensitivity analysis with different random seeds would substantially strengthen the channel-estimation results.
  3. [Section IV-A and Eq. (13)] The exact orientation of the angular-delay image should be stated explicitly; the factor 1 - w_l/(beta M) suggests a horizontal flip that is never explained in the text.
  4. [References] There are several typographical errors in the reference list, including 'HiSilion' in [4], '3th' in [35], and 'n Proc' in [45]; these should be corrected.
  5. [Fig. 2 caption] The caption notes that the detection boxes are unclear due to platform limitations; this weakens the qualitative evidence for DINO-X's detection ability and should be addressed with a clearer figure or a reproducible detection example.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central zero-shot channel estimation and frozen-feature sensing results are validated against external benchmarks and do not reduce to fitted parameters or self-citations.

full rationale

The paper's central claims are empirical transfer results. For channel estimation (Sections IV-A and V-C1), DINO-X is a frozen, externally pre-trained object detector; the paper reports no CSI training or parameter fitting, and the 9.61 dB margin is against the fixed LS/LMMSE benchmarks. Equation (13) is a coordinate conversion from detector box centers to angle/delay, not a fit of the detector to the target parameters, and Equation (14) reconstructs the channel from the detected parameters and LS gains, so the NMSE evaluation is independent of the method's training data. For human activity recognition and localization (Sections IV-B, IV-C, V-C2, and V-C3), ConvNeXt is frozen and only lightweight output layers are trained on the task labels; this is standard transfer learning, and the comparisons to train-from-scratch VGG19/ResNet50 and to the NoFeatExt-CSI/ConvFeatExt-CSI baselines are external. The paper cites prior work by the same authors (e.g., [15], [18], [29], [30]), but none of those citations is used to justify the central transfer premise; DINO-X [40], ConvNeXt [42], and the datasets [39], [44] are external sources. The speculative passage in Section V-C2 that LVMs pre-trained on CSI could deliver competitive performance is a future-direction remark, not a load-bearing derivation, and the conclusion's caveat about incorporating CSI into LVM pre-training is a stated limitation rather than a circular step. The main unresolved issues are implementation details and statistical support for the DINO-X prompt and box-coordinate convention; those are correctness risks, not circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The framework rests mainly on domain assumptions about CSI-to-image transfer and channel sparsity; it introduces no new physical entities. The few free parameters are design choices (oversampling, image size, head dimensions) rather than fitted physics constants.

free parameters (4)
  • DFT oversampling factors beta and gamma = 4
    Chosen by hand in Section V-A1; they set the 256x256 CSI image resolution in Eq. (12) and affect whether paths appear as separable bright spots.
  • CSI image input size for LVMs = 224x224
    Chosen in Sections V-A2 and V-A3 to match ConvNeXt input; requires resizing or upsampling the CSI matrices.
  • Localization head architecture = channel power expanded to 8, then dense layers 32, 16, 2
    Architecture chosen by hand in Section IV-C; the 8-dimensional channel power expansion and hidden sizes are not derived from theory.
  • HAR dense layer = 1 dense layer with softmax
    Chosen by hand in Section IV-B as a minimal decision layer.
assumptions (5)
  • domain assumption Multipath channel model H = sum_l alpha_l a(Theta_l) b(T_l)^T with L << M is valid for the simulated scenarios.
    Used to generate channel estimation data in Sections II-A and V-A1.
  • domain assumption In the oversampled angular-delay domain, signal paths appear as localized bright spots that can be detected by an object detector trained on natural images.
    This is the core transfer premise behind DINO-X path detection in Section III-A and Fig. 2.
  • domain assumption The center of a detected bounding box maps linearly to angle and delay via Eq. (13).
    Assumes no systematic offset between box center and path location; errors here directly enter the reconstructed channel.
  • domain assumption For human activity recognition, phase information can be discarded and the modulus-only CSI image preserves discriminative activity features.
    Stated in Section IV-B; used to build the 250x90 grayscale image.
  • domain assumption QuaDRiGa with the 3GPP 38.901 UMi NLOS model generates CSI-position pairs realistic enough to validate localization.
    Simulation setup in Section V-A3; no real localization CSI data is used.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LVM4CSI: Enabling Direct Application of Pre-Trained Large Vision Models for Wireless Channel Tasks." pith.science (2026). https://pith.science/paper/XOI7IIAE

@misc{pith2026250705121,
  author       = {Pith},
  title        = {Pith review of: LVM4CSI: Enabling Direct Application of Pre-Trained Large Vision Models for Wireless Channel Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XOI7IIAE}},
  note         = {Machine review of arXiv:2507.05121}
}
read the original abstract

Accurate channel state information (CSI) is critical to the performance of wireless communication systems, especially with the increasing scale and complexity introduced by 5G and future 6G technologies. While artificial intelligence (AI) offers a promising approach to CSI acquisition and utilization, existing methods largely depend on task-specific neural networks (NNs) that require expert-driven design and large training datasets, limiting their generalizability and practicality. To address these challenges, we propose LVM4CSI, a general and efficient framework that leverages the structural similarity between CSI and computer vision (CV) data to directly apply large vision models (LVMs) pre-trained on extensive CV datasets to wireless tasks without any fine-tuning, in contrast to large language model-based methods that generally necessitate fine-tuning. LVM4CSI maps CSI tasks to analogous CV tasks, transforms complex-valued CSI into visual formats compatible with LVMs, and integrates lightweight trainable layers to adapt extracted features to specific communication objectives. We validate LVM4CSI through three representative case studies, including channel estimation, human activity recognition, and user localization. Results demonstrate that LVM4CSI achieves comparable or superior performance to task-specific NNs, including an improvement exceeding 9.61 dB in channel estimation and approximately 40% reduction in localization error. Furthermore, it significantly reduces the number of trainable parameters and eliminates the need for task-specific NN design.

Figures

Figures reproduced from arXiv: 2507.05121 by the authors.

Figure 1
Figure 1. A common framework of the existing LLM-based communication [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of a CSI image [29], [30] and a starry sky image [39]. Despite representing entirely distinct modalities, they appear similar at first glance. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Key framework of the proposed LVM4CSI approach, where the NN [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Key workflow of the proposed LVM4CSI approach, including three key steps, namely CSI-to-CV task translation, CSI-to-CV data transformation, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Key framework of the LVM4CSI-enabled channel estimation, where [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Key framework of the LVM4CSI-enabled CSI-based human activity [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Key framework of the LVM4CSI-enabled CSI-based user localization, [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: NMSE performance comparison of LVM-enabled channel estimation with LS and LMMSE methods across varying SNRs and path numbers. (a) The [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: An example of low-quality channel estimation using the LVM4CSI [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Mean localization error comparison between the LVM4CSI frame [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Cross-System Neural Precoder: Exploiting Structural Consistency for Fast Adaptation

    eess.SP 2026-07 conditional novelty 6.0 of 10

    AO subproblems for digital, analog, and RIS variables share reusable update structure, so a lightweight unfolded network adapts across precoding systems with far less data than a GNN baseline.

  2. Disruption of a Giant: Spectroscopic Identification of Members in the Periphery and Tidal Tails of $\omega$ Centauri

    astro-ph.GA 2026-05 unverdicted novelty 5.0 of 10

    VLT/FLAMES spectroscopy confirms 157 ω Centauri members at 93% success and traces tidal-tail debris to 3.2 deg with kinematics and metallicities continuous into the Fimbulthul stream.

Reference graph

Works this paper leans on

49 extracted references · 38 canonical work pages · cited by 2 Pith papers

  1. [29]

    Deep learning based fast downlink channel reconstruction for FDD massive MIMO systems,

    M. Li, Y . Han, X. Li, C.-K. Wen, and S. Jin, “Deep learning based fast downlink channel reconstruction for FDD massive MIMO systems,” in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC) , 2020, pp. 1–6

  2. [30]

    A novel dual-driven channel estimation scheme for spatially non-stationary fading environments,

    J. Jiang, L. Lian, T. Yu, Q. Shi, S. Zhang, X. Chen, and V . K. N. Lau, “A novel dual-driven channel estimation scheme for spatially non-stationary fading environments,” IEEE Trans. Wireless Commun. , vol. 23, no. 7, pp. 7027–7042, Jul. 2024

  3. [25]

    Deep denoising neural network assisted compressive channel estimation for mmwave intelligent reflecting surfaces,

    S. Liu, Z. Gao, J. Zhang, M. D. Renzo, and M.-S. Alouini, “Deep denoising neural network assisted compressive channel estimation for mmwave intelligent reflecting surfaces,” IEEE Trans. Veh. Technol. , vol. 69, no. 8, pp. 9223–9228, Aug. 2020

  4. [1]

    6G takes shape,

    J. G. Andrews, T. E. Humphreys, and T. Ji, “6G takes shape,” IEEE BITS Inform. Theory Mag., , vol. 4, no. 1, pp. 2–24, Mar. 2024

  5. [2]

    Future technology trends of terrestrial international mobile telecommunications systems towards 2030 and beyond,

    ITU-R M.2516-0, “Future technology trends of terrestrial international mobile telecommunications systems towards 2030 and beyond,” Tech. Rep., Nov. 2022, Accessed on Apr. 20, 2025. [Online]. Available: https: //www.itu.int/dms pub/itu-r/opb/rep/R-REP-M.2516-2022-PDF-E.pdf

  6. [3]

    The roadmap to 6G: AI empowered wireless networks,

    K. B. Letaief, W. Chen, Y . Shi, J. Zhang, and Y .-J. A. Zhang, “The roadmap to 6G: AI empowered wireless networks,” IEEE Commun. Mag., vol. 57, no. 8, pp. 84–90, Aug. 2019

  7. [4]

    Views on 6G Radio,

    3GPP 6GWS-250159, “Views on 6G Radio,” Huawei, HiSilion, Tech. Rep., Mar. 2025, Accessed on Apr. 20, 2025. [Online]. Available: https://www.3gpp.org/ftp/workshop/2025-03-10 3GPP 6G WS/Docs/6GWS-250159.zip

  8. [5]

    Advanced deep learning models for 6G: Overview, opportunities, and challenges,

    L. Jiao, Y . Shao, L. Sun et al. , “Advanced deep learning models for 6G: Overview, opportunities, and challenges,” IEEE Access, vol. 12, pp. 133 245–133 314, 2024

Show all 49 references
  1. [6]

    Towards explainable AI for channel estimation in wireless communications,

    A. K. Gizzini, Y . Medjahdi, A. J. Ghandour, and L. Clavier, “Towards explainable AI for channel estimation in wireless communications,” IEEE Trans. Veh. Technol., vol. 73, no. 5, pp. 7389–7394, May 2024

  2. [7]

    Massive MIMO channel prediction: Kalman filtering vs. machine learning,

    H. Kim, S. Kim, H. Lee, C. Jang, Y . Choi, and J. Choi, “Massive MIMO channel prediction: Kalman filtering vs. machine learning,” IEEE Trans. Commun., vol. 69, no. 1, pp. 518–528, Jan. 2021

  3. [8]

    Overview of deep learning- based CSI feedback in massive MIMO systems,

    J. Guo, C.-K. Wen, S. Jin, and G. Y . Li, “Overview of deep learning- based CSI feedback in massive MIMO systems,” IEEE Trans. Commun., vol. 70, no. 12, pp. 8017–8045, Dec. 2022

  4. [9]

    AI empowered channel semantic acquisition for 6G integrated sensing and communication networks,

    Y . Zhang, Z. Gao, J. Zhao et al. , “AI empowered channel semantic acquisition for 6G integrated sensing and communication networks,” IEEE Network, vol. 38, no. 2, pp. 45–53, Mar. 2024

  5. [10]

    Twenty- five years of advances in beamforming: From convex and nonconvex op- timization to learning techniques,

    A. M. Elbir, K. V . Mishra, S. A. V orobyov, and R. W. Heath, “Twenty- five years of advances in beamforming: From convex and nonconvex op- timization to learning techniques,” IEEE Signal Process. Mag. , vol. 40, no. 4, pp. 118–131, Jun. 2023

  6. [11]

    AI/ML for beam management in 5G-Advanced: A standardization perspective,

    Q. Xue, J. Guo, B. Zhou, Y . Xu, Z. Li, and S. Ma, “AI/ML for beam management in 5G-Advanced: A standardization perspective,”IEEE Veh. Technol. Mag., vol. 19, no. 4, pp. 64–72, Dec. 2024

  7. [12]

    5G NR positioning enhancements in 3GPP Release-18,

    H.-S. Cha, G. Lee, A. Ghosh, M. Baker, S. Kelley, and J. Hofmann, “5G NR positioning enhancements in 3GPP Release-18,” IEEE Commun. Standards Mag., vol. 9, no. 1, pp. 22–27, Mar. 2025

  8. [13]

    Rethinking bias- variance trade-off for generalization of neural networks,

    Z. Yang, Y . Yu, C. You, J. Steinhardt, and Y . Ma, “Rethinking bias- variance trade-off for generalization of neural networks,” in Proc. 37th Int. Conf. Mach. Learn. (ICML) , 2020, pp. 10 767–10 777

  9. [14]

    A survey of large language models,

    W. X. Zhao, K. Zhou, J. Li et al., “A survey of large language models,” arXiv preprint arXiv:2303.18223 , 2023

  10. [15]

    Prompt-enabled large AI models for CSI feedback,

    J. Guo, Y . Cui, C.-K. Wen, and S. Jin, “Prompt-enabled large AI models for CSI feedback,” arXiv preprint arXiv:2501.10629 , 2025

  11. [16]

    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

  12. [17]

    LLM4CP: Adapting large language models for channel prediction,

    B. Liu, X. Liu, S. Gao, X. Cheng, and L. Yang, “LLM4CP: Adapting large language models for channel prediction,” J. Commun. Inf. Netw. , vol. 9, no. 2, pp. 113–125, Jun. 2024

  13. [18]

    Exploring the potential of large language models for massive MIMO CSI feedback,

    Y . Cui, J. Guo, C.-K. Wen, S. Jin, and E. Tong, “Exploring the potential of large language models for massive MIMO CSI feedback,” arXiv preprint arXiv:2501.10630, 2025

  14. [19]

    BeamLLM: Vision- empowered mmwave beam prediction with large language models,

    C. Zheng, J. He, G. Cai, Z. Yu, and C. G. Kang, “BeamLLM: Vision- empowered mmwave beam prediction with large language models,” arXiv preprint arXiv:2503.10432 , 2025

  15. [20]

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

    Y . Zhang, H. Yin, W. Li, E. Bjornson, and M. Debbah, “Port-LLM: A port prediction method for fluid antenna based on large language models,” arXiv preprint arXiv:2502.09857 , 2025

  16. [21]

    Large language model enabled multi-task physical layer network,

    T. Zheng and L. Dai, “Large language model enabled multi-task physical layer network,” arXiv preprint arXiv:2412.20772 , 2024

  17. [22]

    Deep transfer learning for gesture recognition with WiFi signals,

    Q. Bu, G. Yang, X. Ming, T. Zhang, J. Feng, and J. Zhang, “Deep transfer learning for gesture recognition with WiFi signals,” Pers. Ubiquitous Comput., pp. 1–12, Jun. 2022

  18. [23]

    U-shape networks are unified backbones for human action understanding from Wi-Fi signals,

    F. Wang, Y . Gao, B. Lan, H. Ding, J. Shi, and J. Han, “U-shape networks are unified backbones for human action understanding from Wi-Fi signals,” IEEE Internet Things J. , vol. 11, no. 6, pp. 10 020– 10 030, Mar. 2024

  19. [24]

    Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising,

    K. Zhang, W. Zuo, Y . Chen, D. Meng, and L. Zhang, “Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising,” IEEE Trans. Image Process. , vol. 26, no. 7, pp. 3142–3155, Jul. 2017

  20. [26]

    Image super-resolution using deep convolutional networks,

    C. Dong, C. C. Loy, K. He, and X. Tang, “Image super-resolution using deep convolutional networks,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 38, no. 2, pp. 295–307, Feb. 2016

  21. [27]

    Deep learning for super- resolution channel estimation in reconfigurable intelligent surface aided systems,

    W. Shen, Z. Qin, and A. Nallanathan, “Deep learning for super- resolution channel estimation in reconfigurable intelligent surface aided systems,” IEEE Trans. Commun. , vol. 71, no. 3, pp. 1491–1503, Mar. 2023

  22. [28]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2016, pp. 779–788

  23. [31]

    Deep learning based user grouping for FD-MIMO systems exploiting statistical channel state information,

    S. Ji, Q. Wang, S. Wu, J. Tian, X. Li, and W. Wang, “Deep learning based user grouping for FD-MIMO systems exploiting statistical channel state information,” China Commun., vol. 18, no. 7, pp. 183–196, Jul. 2021

  24. [32]

    ImageNet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,”Commun. ACM, vol. 60, no. 6, p. 84–90, Jun. 2017

  25. [33]

    Convolutional neural networks based indoor Wi-Fi localization with a novel kind of CSI images,

    H. Li, X. Zeng, Y . Li, S. Zhou, and J. Wang, “Convolutional neural networks based indoor Wi-Fi localization with a novel kind of CSI images,” China Commun., vol. 16, no. 9, pp. 250–260, Sep. 2019

  26. [34]

    Deep convolutional neural networks for indoor localization with CSI images,

    X. Wang, X. Wang, and S. Mao, “Deep convolutional neural networks for indoor localization with CSI images,” IEEE Trans. Netw. Sci. Eng. , vol. 7, no. 1, pp. 316–327, Jan.-Mar. 2020

  27. [35]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” in Proc. 3th Int. Conf. Learn. Representations (ICLR), 2015, pp. 10 767–10 777

  28. [36]

    Medical image segmentation review: The success of U-Net,

    R. Azad, E. K. Aghdam, A. Rauland et al., “Medical image segmentation review: The success of U-Net,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 12, pp. 10 076–10 095, Dec. 2024

  29. [37]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 770–778

  30. [38]

    SrcSense: Robust WiFi-based motion source recognition via signal-informed deep learning,

    G. Zhu, B. Wang, W. Gao, Y . Hu, C. Wu, and K. J. Ray Liu, “SrcSense: Robust WiFi-based motion source recognition via signal-informed deep learning,” IEEE J. Sel. Areas Sens. , vol. 2, pp. 40–53, Jan. 2025

  31. [39]

    DeepSpaceYoloDataset: Annotated astronomical images captured with smart telescopes,

    O. Parisot, “DeepSpaceYoloDataset: Annotated astronomical images captured with smart telescopes,” Data, vol. 9, no. 1, p. 12, Jan. 2024

  32. [40]

    DINO-X: A unified vision model for open-world object detection and understanding,

    T. Ren, Y . Chen, Q. Jiang et al. , “DINO-X: A unified vision model for open-world object detection and understanding,” arXiv preprint arXiv:2411.14347, 2024

  33. [41]

    Weight distillation: Transferring the knowledge in neural network parameters,

    Y . Lin, Y . Li, Z. Wang, B. Li, Q. Du, T. Xiao, and J. Zhu, “Weight distillation: Transferring the knowledge in neural network parameters,” in Proc. 59th Annu. Meet. Assoc. Comput. Linguist. (ACL) , 2021, pp. 2076–2088

  34. [42]

    A ConvNet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A ConvNet for the 2020s,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 11 976–11 986

  35. [43]

    Newtonized orthog- onal matching pursuit: Frequency estimation over the continuum,

    B. Mamandipoor, D. Ramasamy, and U. Madhow, “Newtonized orthog- onal matching pursuit: Frequency estimation over the continuum,” IEEE Trans. Signal Process., vol. 64, no. 19, pp. 5066–5081, Oct. 2016

  36. [44]

    A survey on behavior recognition using WiFi channel state information,

    S. Yousefi, H. Narui, S. Dayal, S. Ermon, and S. Valaee, “A survey on behavior recognition using WiFi channel state information,” IEEE Commun. Mag., vol. 55, no. 10, pp. 98–104, Oct. 2017

  37. [45]

    CSI- StripeFormer: Exploiting stripe features for CSI compression in massive MIMO system,

    Q. Hu, H. Kang, H. Chen, Q. Huang, Q. Zhang, and M. Cheng, “CSI- StripeFormer: Exploiting stripe features for CSI compression in massive MIMO system,” in n Proc. IEEE Conf. Comput. Commun. (INFOCOM), 2023, pp. 1–10

  38. [46]

    CSWin transformer: A general vision transformer backbone with cross-shaped windows,

    X. Dong, J. Bao, D. Chen et al., “CSWin transformer: A general vision transformer backbone with cross-shaped windows,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2022, pp. 12 124–12 134

  39. [47]

    CompFi: Partially con- nected neural network using complex CSI data for indoor localization,

    M. Kim, C. Kim, D. Han, and J.-K. K. Rhee, “CompFi: Partially con- nected neural network using complex CSI data for indoor localization,” in Proc. IEEE 91st Veh. Technol. Conf. (VTC-Spring) , 2020, p. 1–5

  40. [48]

    Recurrent Conformer for WiFi activity recognition,

    M. Shang and X. Hong, “Recurrent Conformer for WiFi activity recognition,” IEEE/CAA J. Automatica Sinica , vol. 10, no. 6, pp. 1491– 1493, Jun. 2023

  41. [49]

    Searching for MobileNetV3,

    A. Howard, M. Sandler, G. Chu et al., “Searching for MobileNetV3,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , 2019, p. 1314–1324

Pith tools

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