Pith. sign in

REVIEW 5 major objections 8 minor 2 cited by

LLM4SG: Adapting Large Language Model for Scatterer Generation via Synesthesia of Machines

T0 review · 5 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a fine-tuned large language model can generate electromagnetic scatterer maps from LiDAR point clouds in V2V settings, with over 92% position accuracy and a few-shot performance that matches a full-sample ResNet.

desk verdict A plausible LLM-for-scatterer-generation paper whose headline accuracies rest on an under-specified ground-truth extraction and metric definitions; worth refereeing after the authors pin those down. read the letter →

arxiv 2505.17879 v3 pith:IUDKP7AM submitted 2025-05-23 eess.SP

Zengrui Han , Lu Bai , Ziwei Huang , Xiang Cheng This is my paper · ORCID
classification eess.SP
keywords largelanguagemodelscatterergenerationLiDARpointcloudSynesthesiaofMachinesvehicle-to-vehiclecommunicationsfew-shottransferraytracingLoRAfine-tuning
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

LLM4SG claims to replace expensive ray tracing with a pre-trained language model that reads LiDAR point clouds and outputs maps of electromagnetic scatterers, the points where radio paths bounce off objects. The paper builds a synthetic V2V dataset, SynthSoM-V2V, with paired LiDAR and ray-tracing-derived scatterer grids across two scenarios, two frequency bands, and two traffic densities. On that dataset it reports over 92% accuracy on scatterer position and over 85% on scatterer count, and shows that a few hundred transfer samples, under 2% of the full set, let the model match a full-sample-trained ResNet baseline. If these numbers hold, scatterer-level channel data for new environment conditions could be synthesized from sensing data alone, at a scale that AI-native 6G design is said to need.

What carries the argument

The load-bearing machinery is the conversion of LiDAR returns into a three-channel grid feature map with point density, maximum height, and averaged transceiver distance, which is patch-embedded into tokens together with a learnable frequency token; those tokens are processed by the first six transformer blocks of GPT-2, where only selected LoRA adapters are trainable. The output head applies a convolution and LeakyReLU to the LLM's hidden states to produce a single-layer scatterer grid map. This treats scatterer generation as a sequence-to-grid translation in token space: the pretrained transformer is the feature extractor, and the embedding module aligns the sensing-communication domain with the natural-language token domain.

What would settle it

Run the same LLM4SG training and evaluation protocol on SynthSoM-V2V with a fully documented, reproducible scatterer-extraction rule, for instance a fixed maximum number of bounces and a fixed path-clustering threshold, and check whether the reported 92%/85% accuracies and the sub-2% few-shot transfer results reproduce; if performance collapses under a changed but still reasonable labeling rule, the paper's quantitative claims depend on the unspecified label extraction.

Watch

Extended reading notes

Core claim

The paper's central claim is that the first six layers of GPT-2, adapted with LoRA in a task-guided selection of attention and feed-forward modules, learn the cross-modal mapping from LiDAR point-cloud grid features to scatterer grid maps. The paper reports that the resulting LLM4SG reaches over 92% position accuracy and over 85% quantity accuracy across the four V2V conditions constructed in SynthSoM-V2V, and that in few-shot frequency, scenario, and traffic-density transfer it matches or beats a full-sample-trained ResNet with fewer than 2% of the training samples. It also reports an average improvement of over 7.5% over the ResNet baseline across tested conditions, and a case study where a time autocorrelation function derived from LLM4SG-generated scatterers closely tracks the ray-tracing reference. The paper positions this as a pathway to massive, high-quality small-scale fading data for 6G systems.

Load-bearing premise

The labels assume that ray-tracing output can be converted into a well-defined set of scatterer points, but the paper never states how ray paths are grouped into scatterers, how many bounces are retained, or how scatterer coordinates are assigned to grid cells; if that extraction is ambiguous or scenario-dependent, the supervised target is not well defined and all reported accuracies are relative to an arbitrary labeling.

Editorial extensions

If this is right

  • Scatterer maps for a new frequency, scenario, or traffic density can be generated from LiDAR alone with a few hundred labeled samples, avoiding a full ray-tracing run for each new condition.
  • The predicted scatterer grid can be fed into a geometry-based channel model to obtain small-scale fading statistics; the paper's TACF comparison shows a close match to ray tracing.
  • Because the input is a tokenized grid rather than raw text, the GPT-2 backbone can be swapped for another LLM without redesigning the preprocessor or output module.
  • The paper reports an average improvement of over 7.5% over the ResNet baseline across the tested conditions, in addition to the few-shot transfer advantage.

Reading between the lines

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

  • A decisive test of the mechanism would replace the pretrained GPT-2 weights with randomly initialized transformer blocks of the same depth; if the performance survives, the contribution is the tokenization and embedding rather than the pretrained language knowledge.
  • The learnable frequency token suggests a general recipe for conditioning any sensing-to-EM mapping on scalar channel parameters, such as carrier frequency, antenna height, or bandwidth; extending it beyond the two tested bands would test that generality.
  • Because the dataset is entirely synthetic, the transfer claims are about simulated-to-simulated generalization; validating on real LiDAR data with measured or ray-traced scatterers in one scenario would test whether the mapping survives sensor noise and imperfect ground truth.
  • If the label-extraction ambiguity is resolved with a stable rule, the method could replace stochastic scatterer distributions in geometry-based stochastic channel models, generating full channel impulse responses for link-level simulation without ray tracing.
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 / 8 minor

Summary. The paper proposes LLM4SG, a GPT-2-based architecture that maps LiDAR point-cloud features to scatterer grid maps for V2V communications. It introduces SynthSoM-V2V, a synthetic dataset built from Wireless InSite ray tracing and AirSim across urban/suburban scenarios, 5.9 GHz and 28 GHz bands, and low/high traffic densities. A four-module design (preprocessor, embedding, backbone with LoRA, output) is evaluated against ResNet and SegNet baselines in full-sample and few-shot transfer tests. The central claims are that LLM4SG achieves over 92% scatterer position accuracy and over 85% quantity accuracy, and that with fewer than 2% of training samples it matches or exceeds a full-sample-trained ResNet. An ablation study and a TACF application case are also presented.

Significance. If the ground-truth extraction and metric issues are resolved, this would be a useful contribution to AI-native 6G channel data generation: it is among the first to apply an LLM to LiDAR-to-scatterer mapping, the dataset spans multiple scenarios, frequency bands, and traffic densities, and the few-shot transfer evaluation is a genuine out-of-sample test. The ablation supports the role of each architectural module, and the TACF application indicates practical value. However, the current manuscript leaves the supervised target under-specified and the headline metrics not fully reproducible, so the significance claims cannot yet be relied upon.

major comments (5)
  1. [II.B, Eq. (17)] The supervised target is under-specified. The paper states that Wireless InSite generates 'electromagnetic propagation scatterers and CIR data' (Section II.B), but never specifies how RT output is converted into the scatterer set S_g(t) used in Eq. (17): whether S_g(t) contains all ray interaction points, only first-order or last-bounce interactions, whether nearby points are merged or clustered, or how coordinates are assigned to the 10x10 grid cells. Since both Ppos (Eq. (20)) and Pnum (Eq. (21)) are computed against this ground truth, an ambiguous extraction makes the headline accuracy figures (over 92% / 85%) relative to an arbitrary labeling. Please provide the exact extraction rule, including any clustering or thresholding steps, or release the extraction code.
  2. [IV.A.5, Eq. (21)] Pnum as defined is incomparable and has undefined behavior. The threshold Sth is never given a value, so the numeric Pnum values in Tables III-V cannot be reproduced or compared across settings. In addition, when phi_g(t) = 0, the fraction |phi_g - phi_hat_g| / phi_g is undefined (division by zero); the paper does not state how empty ground-truth cells are treated in the numerator or whether they are excluded from the denominator. Please specify Sth, define the handling of zero-density cells, and report results for a range of Sth to show that the metric is not threshold-sensitive.
  3. [IV.B, Tables III-IV] The performance claims lack statistical support. Section IV.B states that 'We performed experiments using various hyper-parameter values and selected the optimal results for comparison', but no error bars, number of random seeds, or significance tests are reported. The reported gains of 2 to 6 percentage points in Ppos and 5 to 9 percentage points in Pnum could be within run-to-run variation or a product of validation-set overfitting. Please report mean and standard deviation over at least three to five seeds, and state the hyperparameter selection procedure (e.g., selection by validation-set NMSE) so that the 'optimal results' are well-defined.
  4. [III.A, Table II] Several load-bearing hyperparameters are missing, preventing reproduction. Section III.A defines a ground-removal threshold H_G but never gives its value or how it is chosen; LoRA rank, alpha, and the exact target-layer subset used for the 'task guided' fine-tuning are not reported in Table II or the text; and the parameter counts in Table VI (5.46/86.19 M) are unexplained. Without these, the architecture's claimed efficiency and the reproducibility of the results are in question. Please include all missing hyperparameters and clarify what the two numbers in Table VI denote, e.g., trainable versus total parameters.
  5. [Abstract and V, with Section III] The paper frames LLM4SG as learning a 'general mapping relationship from LiDAR point clouds to electromagnetic scatterers' and concludes that it can supply 'massive and high-quality channel small-scale fading data', but both training and evaluation use Wireless InSite RT labels, so the model is a learned surrogate of RT rather than an independent physical derivation. The few-shot transfers to unseen frequency, scenario, and VTD combinations are genuine out-of-sample tests and are a strength, but the claims of 'high-quality' and of replacing ray tracing should be explicitly scoped to 'as judged by agreement with RT labels', and ideally supplemented by a real-measurement validation or a clear statement that such validation is left to future work.
minor comments (8)
  1. [IV.A.4, Eq. (19)] The summation index N in the NMSE formula is not defined; it should be the number of grid cells n_x * n_y, and the subscript g is missing in the phi terms.
  2. [Table I] The text refers to the 'ScaR network', but the abbreviation ScaR is not defined anywhere; this appears to be a typo for LLM4SG.
  3. [III.B, Eq. (12)] The symbol F is used for both the feature grid map and the fully connected network in Eq. (12), which is confusing; please use a different symbol for the network.
  4. [Figs. 6-10] The 'About XXX samples' annotations are not accompanied by a clear marker on the curves; please state how these sample counts are derived and add error bars or confidence bands to the accuracy curves.
  5. [Table VI] The two parameter counts for LLM4SG are unexplained; the caption should state whether the first number is the trainable parameter count and the second is the total parameter count.
  6. [Fig. 3 and other figures] Several figures contain non-English text (e.g., Chinese text in Fig. 3); please provide English translations for all labels and annotations.
  7. [II, after Section II.B] SynthSoM-V2V is claimed as a contribution, but no download link or access mechanism is provided; please indicate how readers can obtain the dataset or state clearly that it will be released upon publication.
  8. [IV.C.2, Figs. 8-9] The text says the scenario-transfer study covers 'three distinct environments', but the figures only show transfers involving crossroad, forking road, and wide lane in pairs; please clarify whether the pair forking road to wide lane is omitted or presented elsewhere.

Circularity Check

1 steps flagged · score 6.0 of 10

Target leakage in preprocessing: the LiDAR input is filtered by a VR ellipsoid defined from the RT scatterers that the network is trained to predict.

  1. self definitional [Section III.A (Preprocessor Module), before Eq. (1)]
    "Then, LiDAR point clouds undergo visibility region (VR) [38] filtering, which is widely used in channel modeling. To be specific, scatterers are visible and contribute to channel characterization only if they are within the VR. In [38], the VR is defined as an ellipsoid with the Tx and Rx as its two focuses. The major axis 2a(t) of the ellipsoid is calculated as the sum of the distances from the scatterers to transceivers."

    The preprocessing step that defines the model input removes all LiDAR points outside a VR ellipsoid whose major axis is computed from "the scatterers"—the very RT scatterer locations that later form the target scatterer set S_g(t) in Eq. (17). Thus the input feature map F is not raw LiDAR; it is LiDAR masked by a geometric region derived from the ground-truth output. At deployment, reproducing this filter would require knowing the scatterers being predicted, so the reported Ppos/Pnum and transfer results are evaluated on inputs that already contain label-dependent information. The paper does not state a LiDAR-only method for estimating 2a(t).

full rationale

The central mapping claim is compromised by a self-definitional input step: Section III.A filters the LiDAR point clouds using a visibility-region ellipsoid, and the ellipsoid's major axis is defined from RT scatterer distances. Because those scatterers are the prediction target (Eq. (17)), the input is not independent of the output. This is the one concrete circular step found. The remaining concerns—the unspecified RT-to-scatterer extraction rule in Section II.B and the undefined Pnum denominator in Eq. (21) for zero-density cells—are reproducibility/correctness defects rather than circularity. The many self-citations ([20], [22], [34], [38]) support the dataset platform and prior exploratory work but are not themselves the derivation. Aside from the VR leakage, the few-shot transfer tests to unseen frequency, scenario, and VTD conditions are genuine out-of-sample evaluations, so the paper does not reduce entirely to a fit; the circularity is partial and localized to the input construction.

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

The method is supervised learning on synthetic ray-traced labels; there is no first-principles derivation. Key free parameters are metric thresholds, ground-filtering thresholds, and unreported LoRA settings. The main assumed premises are the fidelity of the AirSim/Wireless InSite alignment and the validity of the RT scatterer labels.

free parameters (4)
  • Sth (Pnum accuracy threshold) = not reported
    Defines Pnum in Eq. (21) but is never given; the headline 'over 85% quantity accuracy' is not computable without it.
  • HG (ground removal threshold) = not reported
    Used to filter ground points in LiDAR in Section III.A; it is a hand-chosen threshold that affects input features.
  • LoRA rank and alpha and target layer subset = not reported
    The paper states LoRA is applied to a carefully selected subset of layers but gives no rank, alpha, or exact modules, making the method irreproducible.
  • Architecture grid and patch sizes = 80x80 input, 10x10 output, 8x8 patch, 6 GPT-2 layers, D=768
    Chosen by hand in Table I; these resolutions determine the granularity of the scatterer map and the complexity of the model.
assumptions (4)
  • domain assumption AirSim LiDAR and Wireless InSite electromagnetic space are precisely aligned, so scatterer positions from RT correspond to physical objects in the LiDAR scene.
    Stated in Section II.A: '3D models in physical environments are properly imported from AirSim into Wireless InSite to achieve precise alignment.' This is assumed, not verified quantitatively.
  • domain assumption Ray-tracing scatterers extracted from Wireless InSite provide valid ground truth for electromagnetic scatterers.
    Section II.B relies on RT to generate scatterers but does not define the extraction; the whole supervised training depends on label validity.
  • domain assumption The LiDAR feature grid (density, max height, average distance) is a sufficient representation of the environment for predicting scatterers.
    Section III.A designs F1-F3 but no analysis shows sufficiency; material properties and occluded objects are not represented.
  • domain assumption The visibility region ellipsoid criterion from [38] correctly identifies scatterers relevant to channel characterization.
    Section II.A applies VR filtering from [38], assuming this geometry is appropriate for the V2V scenarios.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM4SG: Adapting Large Language Model for Scatterer Generation via Synesthesia of Machines." pith.science (2026). https://pith.science/paper/IUDKP7AM

@misc{pith2026250517879,
  author       = {Pith},
  title        = {Pith review of: LLM4SG: Adapting Large Language Model for Scatterer Generation via Synesthesia of Machines},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IUDKP7AM}},
  note         = {Machine review of arXiv:2505.17879}
}
read the original abstract

In this paper, a novel large language model (LLM)-based method for scatterer generation (LLM4SG) is proposed for sixth-generation (6G) artificial intelligence (AI)-native communications. To provide a solid data foundation, we construct a new synthetic intelligent sensing-communication dataset for Synesthesia of Machines (SoM) in vehicle-to-vehicle (V2V) communications, named SynthSoM-V2V, covering multiple V2V scenarios with multiple frequency bands and multiple vehicular traffic densities (VTDs). Leveraging the powerful cross-modal representation capabilities of LLMs, LLM4SG is designed to capture the general mapping relationship from light detection and ranging (LiDAR) point clouds to electromagnetic scatterers via SoM. To address the inherent and significant differences across multi-modal data, synergistically optimized four-module architecture, i.e., preprocessor, embedding, backbone, and output modules, are designed by considering sensing characteristics and electromagnetic propagation. The embedding module achieves effective cross-domain alignment of the sensing-communication domain and the natural language domain.The backbone network is adapted in a task-guided manner with low rank adaptation (LoRA), where a carefully selected subset of layers is fine tuned to preserve general knowledge and reduce training cost. The proposed LLM4SG is evaluated for scatterer generation by benchmarking against ray-tracing (RT) and conventional deep learning models. Simulation results demonstrate that the proposed LLM4SG achieves superior performance in both full-sample and cross-condition generalization testing. It significantly outperforms conventional deep learning models across different frequency bands, scenarios, and VTDs, and demonstrates the capability to provide the massive and high-quality channel small-scale fading data required by AI-native 6G systems.

Figures

Figures reproduced from arXiv: 2505.17879 by the authors.

Figure 1
Figure 1. Crossroads in AirSim and Wireless InSite and detailed parameter setting of vehicular trajectories under high and low VTDs. (a) Crossroad in AirSim [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Forking roads in AirSim and Wireless InSite and detailed parameter setting of vehicular trajectories under high and low VTDs. (a) Forking road in ➢ 输入：多场景（分岔路口、十字路口）、多车流量密度（高、低）下的LiDAR点云；频点（28 GHz/5.9 GHz） [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. An illustration of the network architecture of LLM4SG. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: An illustration of the patching operation of LiDAR point cloud features. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: An illustration of the position encoding. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Generalization performance in frequency band transfer in urban [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Generalization performance in frequency band transfer in suburban [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Generalization performance in scenario transfer. (a) Location eval [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 11
Figure 11. Figure 11: Comparison of simulated TACFs and RT-based TACFs. [PITH_FULL_IMAGE:figures/full_fig_p011_11.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. WiFo-CF: Wireless Foundation Model for CSI Feedback

    eess.SP 2025-08 unverdicted novelty 6.0 of 10

    WiFo-CF is a pretrained wireless foundation model that handles heterogeneous CSI feedback configurations and transfers to localization tasks.

  2. Foundation Model Empowered Synesthesia of Machines (SoM): AI-native Intelligent Multi-Modal Sensing-Communication Integration

    eess.SP 2025-06 conditional novelty 4.0 of 10

    The paper proposes a systematic classification and two roadmaps for using foundation models (LLMs and wireless foundation models) to design Synesthesia of Machines systems for 6G, with preliminary case-study evidence ...

Reference graph

Works this paper leans on

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

  1. [1]

    Channel nonstationarity and consis- tency for beyond 5G and 6G: A survey,

    X. Cheng, Z. Huang, and L. Bai, “Channel nonstationarity and consis- tency for beyond 5G and 6G: A survey,” IEEE Commun. Surveys Tuts. , vol. 24, no. 3, pp. 1634–1669, Thirdquarter 2022

    work page 2022

  2. [2]

    Hybrid far- and near-field modeling for reconfigurable intelligent surface assisted V2V channels: A sub-array partition based approach,

    H. Jiang, B. Xiong, H. Zhang, and E. Basar, “Hybrid far- and near-field modeling for reconfigurable intelligent surface assisted V2V channels: A sub-array partition based approach,” IEEE Trans. Wireless Commun., vol. 15, no. 1, pp. 16–26, Mar. 2020

    work page 2020

  3. [3]

    A tone-based AoA estimation and multiuser precoding for millimeter wave massive MIMO,

    L. Zhao et al., “A tone-based AoA estimation and multiuser precoding for millimeter wave massive MIMO,” IEEE Trans. Commun. , vol. 65, no. 12, pp. 5209–5225, Dec. 2017

    work page 2017

  4. [4]

    Multi-user millimeter wave MIMO with full-dimensional lens antenna array,

    Y . Zeng, L. Yang, and R. Zhang, “Multi-user millimeter wave MIMO with full-dimensional lens antenna array,” IEEE Trans. Wireless Com- mun., vol. 17, no. 4, pp. 2800–2814, Apr. 2018

    work page 2018

  5. [5]

    Machine learning for reliable mmWave systems: Blockage prediction and proactive handoff,

    A. Alkhateeb, I. Beltagy, and S. Alex, “Machine learning for reliable mmWave systems: Blockage prediction and proactive handoff,” in Proc. IEEE GlobalSIP’18, Anaheim, CA, USA, Nov. 2018, pp. 1055–1059

    work page 2018

  6. [6]

    Deep learning for mmWave beam and blockage prediction using sub-6 GHz channels,

    M. Alrabeiah and A. Alkhateeb, “Deep learning for mmWave beam and blockage prediction using sub-6 GHz channels,” IEEE Trans. Commun., vol. 68, no. 9, pp. 5504–5518, Sept. 2020

    work page 2020

  7. [7]

    A novel SAGE algorithm for estimating parameters of wideband spatial nonstationary wireless channels with antenna polariza- tion,

    Z. Zhou et al., “A novel SAGE algorithm for estimating parameters of wideband spatial nonstationary wireless channels with antenna polariza- tion,” IEEE Trans. Antennas Propagat. , vol. 71, no. 9, pp. 7457–7472, Sept. 2023

    work page 2023

  8. [8]

    A cluster-based three-dimensional channel model for vehicle-to-vehicle communications,

    M. Yang et al. , “A cluster-based three-dimensional channel model for vehicle-to-vehicle communications,” IEEE Trans. Veh. Technol., vol. 68, no. 6, pp. 5208–5220, Jun. 2019

    work page 2019

Show all 45 references
  1. [9]

    A general 3D space-time-frequency non- stationary model for 6G channels,

    Z. Huang and X. Cheng, “A general 3D space-time-frequency non- stationary model for 6G channels,” IEEE Trans. Wireless Commun. , vol. 20, no. 1, pp. 535–548, Jan. 2021

    2021

  2. [10]

    Research on MIMO channel capacity in complex indoor environment based on deterministic channel model,

    Y .-X. Wang, Z.-Y . Liu, and L.-X. Guo, “Research on MIMO channel capacity in complex indoor environment based on deterministic channel model,” in Proc. IEEE ICCC’21 , Xiamen, China, 2021, pp. 405–409

    2021

  3. [11]

    An investigation into the spatial heterogeneity of MIMO device with 3D deterministic channel modelling using single-probe anechoic chamber method,

    Q. Zhang, T. H. Loh, Z. Huang, and F. Qin, “An investigation into the spatial heterogeneity of MIMO device with 3D deterministic channel modelling using single-probe anechoic chamber method,” in Proc. EuCAP’23, Florence, Italy, Mar. 2023, pp. 1–5

    2023

  4. [12]

    Foundation model empowered Synesthesia of Ma- chines (SoM): AI-native intelligent multi-modal sensing-communication integration,

    X. Cheng et al. , “Foundation model empowered Synesthesia of Ma- chines (SoM): AI-native intelligent multi-modal sensing-communication integration,” IEEE Trans. Netw. Sci. Eng. , early access 2025

    2025

  5. [13]

    Large multi-modal models (LMMs) as universal foundation models for AI-native wireless systems,

    S. Xu et al., “Large multi-modal models (LMMs) as universal foundation models for AI-native wireless systems,” IEEE Netw. , vol. 38, no. 5, pp. 10–20, Sept. 2024

    2024

  6. [14]

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

    L. Jiao et al. , “Advanced deep learning models for 6G: Overview, opportunities, and challenges,” IEEE Access , vol. 12, pp. 133245– 133314, Jun. 2024

    2024

  7. [15]

    Artificial intelligence enabled radio propagation for communications—part I: Channel characterization and antenna-channel optimization,

    C. Huang et al. , “Artificial intelligence enabled radio propagation for communications—part I: Channel characterization and antenna-channel optimization,” IEEE Trans. Antennas Propagat., vol. 70, no. 6, pp. 3939– 3954, Jun. 2022

    2022

  8. [16]

    A power-angle-spectrum based clustering and tracking algorithm for time-varying radio channels,

    C. Huang et al., “A power-angle-spectrum based clustering and tracking algorithm for time-varying radio channels,” IEEE Trans. Veh. Technol., vol. 68, no. 1, pp. 291–305, Jan. 2019

    2019

  9. [17]

    Trajectory-joint clustering algorithm for time-varying channel modeling,

    C. Huang et al., “Trajectory-joint clustering algorithm for time-varying channel modeling,” IEEE Trans. Veh. Technol., vol. 69, no. 1, pp. 1041– 1045, Jan. 2020

    2020

  10. [18]

    A Novel Deep Learning Based Time-Varying Multipath Component Tracking Algorithm,

    H. Wang and Z. Sun, “A Novel Deep Learning Based Time-Varying Multipath Component Tracking Algorithm,” in Proc. IEEE VTC’24 , Singapore, Singapore, 2024, pp. 1–5

    2024

  11. [19]

    A scatterer localization method using large-scale antenna array systems,

    G. Zhang et al. , “A scatterer localization method using large-scale antenna array systems,” in Proc. IEEE CAMA’22 , Guangzhou, China, 2022, pp. 1–4

    2022

  12. [20]

    Intelligent multi-modal sensing-communication inte- gration: Synesthesia of Machines,

    X. Cheng et al. , “Intelligent multi-modal sensing-communication inte- gration: Synesthesia of Machines,” IEEE Commun. Surveys Tuts. , vol. 26, no. 1, pp. 258–301, Firstquarter 2024

    2024

  13. [21]

    An intelligent path loss prediction approach based on integrated sensing and communications for future vehicular networks,

    Z. Wei et al. , “An intelligent path loss prediction approach based on integrated sensing and communications for future vehicular networks,” IEEE Open J. Comput. Soc. , vol. 5, pp. 170–180, Apr. 2024

    2024

  14. [22]

    Scatterer recognition for multi-modal intelligent vehicular channel modeling via Synesthesia of Machines,

    Z. Huang, L. Bai, Z. Han, and X. Cheng, “Scatterer recognition for multi-modal intelligent vehicular channel modeling via Synesthesia of Machines,” IEEE Wirel. Commun. Lett. , vol. 14, no. 7, pp. 1899–1903, Jul. 2025

    1903

  15. [23]

    Pretrained transformers as universal computation engines,

    K. Lu, A. Grover, P. Abbeel, and I. Mordatch, “Pretrained transformers as universal computation engines,” arXiv preprint arXiv:2103.05247 , vol. 1, 2021

    2021 arXiv

  16. [24]

    Language models are few-shot learners,

    T. Brown et al., “Language models are few-shot learners,” Adv. Neural Inf. Process. Syst. , vol. 33, pp. 1877–1901, 2020

    1901

  17. [25]

    DiffusionGPT: LLM-driven text-to-image generation system,

    J. Qin et al. , “DiffusionGPT: LLM-driven text-to-image generation system,” arXiv preprint arXiv:2401.10061 , 2024

    2024

  18. [26]

    Medical video generation for disease progression simulation,

    X. Cao et al. , “Medical video generation for disease progression simulation,” arXiv preprint arXiv:2411.11943 , 2024

    2024 arXiv

  19. [27]

    Enhancing table representations with LLM-powered synthetic data generation,

    D. Yang et al. , “Enhancing table representations with LLM-powered synthetic data generation,” arXiv preprint arXiv:2411.03356 , 2024

    2024 arXiv

  20. [28]

    GenG: An LLM-based generic time series data gener- ation approach for edge intelligence via cross-domain collaboration,

    X. Zhou et al., “GenG: An LLM-based generic time series data gener- ation approach for edge intelligence via cross-domain collaboration,” in Proc. IEEE INFOCOM’24 , Vancouver, Canada, May 2024, pp. 1–6

    2024

  21. [29]

    LLM4CP: Adapting large language models for channel prediction,

    B. Liu et al. , “LLM4CP: Adapting large language models for channel prediction,” J. Commun. Inf. Netw., vol. 9, no. 2, pp. 113–125, Jun. 2024

    2024

  22. [30]

    Csi-LLM: A novel downlink channel prediction method aligned with LLM pre-training,

    S. Fan et al., “Csi-LLM: A novel downlink channel prediction method aligned with LLM pre-training,” arXiv preprint arXiv:2409.00005, 2024

    2024 arXiv

  23. [31]

    Beam prediction based on large language models,

    Y . Sheng et al. , “Beam prediction based on large language models,” IEEE Wireless Commun. Lett. , early access 2025

    2025

  24. [32]

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

    Y . Cui et al. , “Exploring the potential of large language models for massive MIMO CSI feedback,” arXiv preprint arXiv:2501.10630, 2025

    2025 arXiv

  25. [33]

    Large AI model for delay-doppler domain channel prediction in 6G OTFS-based vehicular networks,

    J. Xue et al. , “Large AI model for delay-doppler domain channel prediction in 6G OTFS-based vehicular networks,” arXiv preprint arXiv:2503.01116, 2025

    2025 arXiv

  26. [34]

    SynthSoM: A synthetic intelligent multi-modal sensing-communication dataset for Synesthesia of Machines (SoM),

    X. Cheng et al. , “SynthSoM: A synthetic intelligent multi-modal sensing-communication dataset for Synesthesia of Machines (SoM),” Sci. Data, vol. 12, pp. 819–833, May 2025

    2025

  27. [35]

    Wireless InSite

    Remcom. Wireless InSite. [Online]. Available: https://www.remcom.com/wireless-insite-em-propagation-software [Publication date: Jan. 2017, Accessed date: Mar. 2022]

    2017

  28. [36]

    AirSim: High-fidelity visual and physical simulation for autonomous vehicles,

    S. Shah, D. Dey, C. Lovett, and A. Kapoor, “AirSim: High-fidelity visual and physical simulation for autonomous vehicles,” in Field and Service Robotics, M. Hutter and R. Siegwart, Eds. Cham, Switzerland: Springer, 2018, pp. 621–635

    2018

  29. [37]

    M 3SC: A generic dataset for mixed multi-modal (MMM) sensing and communication integration,

    X. Cheng et al. , “M 3SC: A generic dataset for mixed multi-modal (MMM) sensing and communication integration,” China Commun., vol. 20, no. 11, pp. 13–29, Nov. 2023. 13

    2023

  30. [38]

    A multi-modal UA V-to-ground channel model for 6G intelligent sensing-communication integration,

    L. Bai, M. Lu, Z. Huang, and X. Cheng, “A multi-modal UA V-to-ground channel model for 6G intelligent sensing-communication integration,” IEEE Trans. Commun. , accepted, 2025

    2025

  31. [39]

    In defense of grid features for visual question answer- ing,

    H. Jiang et al., “In defense of grid features for visual question answer- ing,” in Proc. IEEE CVPR’20, Seattle, W A, USA, Jun. 2020, pp. 10267– 10276

    2020

  32. [40]

    Attention is all you need,

    A. Vaswani et al., “Attention is all you need,” Adv. Neural Inf. Process. Syst., vol. 30, pp. 5998–6008, 2017

    2017

  33. [41]

    Language models are unsupervised multitask learn- ers,

    A. Radford et al., “Language models are unsupervised multitask learn- ers,” OpenAI blog, vol. 1, no. 8, pp. 1–9, 2019

    2019

  34. [42]

    Reluplex made more practical: Leaky ReLU,

    J. Xu et al. , “Reluplex made more practical: Leaky ReLU,” in Proc. IEEE ISCC’20, Rennes, France, Jul. 2020, pp. 1–7

    2020

  35. [43]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in Proc. ICLR’15, May. 2015, pp. 1–12

    2015

  36. [44]

    Multi-modal intelligent channel modeling: A new mod- eling paradigm via Synesthesia of Machines,

    L. Bai et al. , “Multi-modal intelligent channel modeling: A new mod- eling paradigm via Synesthesia of Machines,” IEEE Commun. Surveys Tuts., accepted, 2025

    2025

  37. [45]

    Technical Specification Group Radio Access Network; Study on Channel Model for Frequencies From 0.5 to 100 GHz (Release 14) , document TR 38.901 Version 14.2.0, 3GPP, Sep. 2017. [Online]. Available: http://www.3gpp.org/DynaReport/38901.htm

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.