Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

A single pre-trained model converts camera and LiDAR views into wireless-channel features that improve beam prediction, channel estimation, channel interpolation, and channel prediction without retraining its backbone.

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

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

T0 review · deepseek-v4-flash

2026-08-03 10:39 UTC pith:P23XI5XM

load-bearing objection A competent, broad proof-of-concept for a sensing-to-CSI foundation model; the universal-gains claim overreaches the evidence, but the framework and evaluation are worth referee time. the 4 major comments →

arxiv 2601.09179 v3 pith:P23XI5XM submitted 2026-01-14 eess.SP

WiFo-M²: Empower Wireless Communications With Plug-and-Play Environment Sensing via Foundation Model

Haotian Zhang , Shijian Gao , Xiang Cheng This is my paper
classification eess.SP
keywords foundation modelmulti-modal sensingcontrastive learningout-of-band channel-aware featuresbeam predictionchannel estimationchannel interpolationchannel prediction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that the usual way of using sensing to help wireless links—hand-designing a small model for each physical-layer task—can be replaced by one foundation model. The model, WiFo-M2, learns to turn camera images and LiDAR point clouds into out-of-band channel-aware features: representations aligned, in a shared latent space, with the actual channel state information at matching moments. Once pre-trained, its backbone stays frozen; downstream PHY algorithms only add a small trainable adapter. Across beam prediction, channel estimation, channel interpolation, and channel prediction, the frozen features improve most task-specific baselines, and they transfer to unseen scenarios and system configurations. A sympathetic reader would take the core claim as: sensory observations carry reusable information about the radio channel, and contrastive pre-training can extract it once, for many tasks.

Core claim

WiFo-M2 is a sequence-to-sequence feature extractor with separate image and LiDAR encoders followed by a GRU that rolls sensory features forward to future timestamps. Its pre-training objective, ContraSoM, aligns each predicted sensory feature with the CSI feature at the same future timestamp via contrastive learning, so the model learns to anticipate channel-relevant scene information. Once pre-trained, the backbone is frozen and a small adapter is trained alongside each downstream PHY algorithm. The paper reports that WiFo-M2 mostly improves beam prediction, channel estimation, channel interpolation, and channel prediction across seven seen and three unseen datasets, with the most reliable

What carries the argument

The load-bearing object is WiFo-M2's temporal feature-extrapolation mechanism: a gated recurrent unit (GRU) that consumes a fixed historical window of frame-level visual and point-cloud features and emits channel-aware features for multiple future timestamps, which are then contrastively aligned with CSI features sampled at those same timestamps. This alignment makes the extrapolated sensory features channel-aware without needing paired RF data at deployment time. Supporting machinery includes a unified sensing pipeline that labels the receiver's bounding box with an angular color code for images and labels LiDAR clusters as receiver, building, or other, creating deterministic sensory-to-CSI

Load-bearing premise

The load-bearing premise is that the GRU extrapolator can predict future out-of-band channel-aware features from a short history of sensory frames across the 25–50 ms inter-frame gap; if the scene changes faster than the sensor frame rate, the predicted features are stale and the downstream gains vanish.

What would settle it

On a scene where a vehicle or blocker moves fast enough to change the dominant propagation path between two sensor frames, compare channel-prediction NMSE using WiFo-M2 features generated by extrapolation versus features computed from actually captured future frames; if extrapolated features yield no improvement over the baseline without sensing, the central claim about bridging the perception gap is falsified.

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

If this is right

  • One frozen sensory backbone can serve beam prediction, channel estimation, channel interpolation, and channel prediction at once, eliminating per-task sensing pipeline design.
  • Because only a lightweight adapter is retrained, adopting WiFo-M2 for a new PHY action adds at most 0.26M trainable parameters in the tested configurations.
  • The temporal extrapolation means a base station can keep using sensing during gaps between sensor frames, provided the scene evolves slowly relative to the 25–50 ms lookahead.
  • The reported transfer to unseen scenarios and measured data indicates the learned OOB features are not tied to the pre-training scenes or carrier frequencies.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • As an extension beyond the paper, the same CSI-aligned feature space could plausibly support untested actions such as handover, rate adaptation, localization, or near-field beam management, since these tasks rely on the same spatial scene information.
  • A high-speed scene where the dominant propagation path changes between sensor frames should be a stress test: gains from extrapolated features should shrink, suggesting an adaptive lookahead or higher sensor frame rate as a natural extension.
  • The weak or negative results on very large antenna arrays imply the sensory prior is spatially coarse; adding fine-grained angular labels or fusing the sensory prior with recent RF measurements during pre-training could sharpen it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes WiFo-M2, a 'foundation model' that converts multi-modal environment sensing (RGB images and LiDAR point clouds) into out-of-band channel-aware features via a contrastive pre-training strategy called ContraSoM. The model ingests a historical window of sensory data and uses a GRU to extrapolate future features, aligning them with CSI features via symmetric InfoNCE and diffusion-based LiDAR augmentation. The frozen backbone plus small adapters is evaluated on beam prediction, channel estimation, channel interpolation, and channel prediction, across 7 seen and 3 unseen datasets (one with measured data). The paper claims universal, plug-and-play performance gains and strong cross-scenario generalization.

Significance. If the claims were fully substantiated, WiFo-M2 would be a meaningful step: a single frozen multi-modal encoder that improves several PHY actions across system configurations and unseen scenarios, with low adaptation overhead (≤0.26M extra parameters) and inference latency within real-time budgets (9.35 ms at 40 fps). The experimental scope is genuinely broad: four PHY actions, multiple baselines, ablations isolating temporal extrapolation, diffusion augmentation, and contrastive vs ImageNet weights, plus a measured-data generalization test. The main weaknesses are that the central extrapolation mechanism is not directly validated, the 'universal gains' framing is contradicted by several degradation results, and statistical variability is not quantified.

major comments (4)
  1. [Sec. IV-C, Eq. (8); Sec. IV-D2; Table IV] The central mechanism that lets WiFo-M2 provide OOB features when fresh sensory data is unavailable—the GRU extrapolator producing v^x_tau for tau in [T, T+Delta_tau]—is never directly validated. No metric (cosine similarity or L2 distance between predicted and true z^tau_C) is reported as a function of horizon or scenario. The ablation in Table IV only gives an average NMSE drop when TFE is removed, which cannot distinguish extrapolation error from representation quality. Because the contrastive loss in Eqs. (16)–(17) trains against future CSI features, extrapolation fidelity is load-bearing for the plug-and-play claim. The paper's own results (BP L1/L2/L3/L7 below Vision-BP in Fig. 7; CI degradation on L3 in Fig. 9; CP no gain on L5 in Table III) are consistent with stale predicted features. Please add direct extrapolation-quality evaluation and report horizon-dependent performance, in
  2. [Abstract; Sec. V-C1–V-C4] The claim of 'universal performance gains' is contradicted by multiple results. On BP, WiFo-M2-Img is below Vision-BP on L1, L2, L3, L7 (Fig. 7); on CI, the paper acknowledges 'marginal performance degradation' on L3 (Sec. V-C3, Fig. 9); on CP, Table III shows no gain on L5 and WiFo-M2-LiDAR degrades the WiFo baseline on L1; on unseen C1, WiFo-M2 multi-modal BP (83.9% Top-1) is below MM-BP (88.7%). The conclusion and abstract should be rephrased to 'substantial but not universal gains', and the paper should analyze the conditions (array size, pilot sparsity, scene dynamics, baseline convergence) under which the OOB features help or hurt. Currently the qualitative summary overstates the evidence.
  3. [Sec. V-C and Figs. 7–9; Tables III–VI] None of the experimental curves or tables report error bars, confidence intervals, or multiple-seed variation. Many comparisons are within 1–2 dB or 1–2 percentage points (e.g., BP Top-1 gaps in Fig. 7; CE curves in Fig. 8; CP gains in Table III). The ablation means in Table IV are averages over L1–L7 with no variance. For a paper whose central claim is consistent gains across datasets and actions, statistical variability must be quantified. Please run at least 3–5 seeds for the adapter training and baselines, or provide significance tests, and report mean ± std.
  4. [Sec. V-A, Sec. V-E] Cross-scenario generalization is evidenced by only one unseen dataset (C1) that exercises the full set of PHY actions; S1 has no CSI so only BP is evaluated, and V1 provides only images. The claim of 'strong generalization to unseen scenarios' is therefore based on one synthetic campus dataset for CE/CI/CP, plus BP on measured data. Given that the pre-training covers only two scenarios (M3SC and SynthSoM), the evidence is too thin to support 'universal' or 'strong' generalization. Please either add more unseen datasets (at least one additional measured dataset with CSI) or substantially soften the generalization claim in the abstract and conclusion.
minor comments (5)
  1. [Sec. IV-C, Eq. (8)] The output dimensions of the GRU, projection layer, and CSI feature vector are all denoted d; please clarify whether the GRU hidden state dimension equals d or whether a projection is needed. Also, the notation for the prediction window in Eq. (8) (T to T+Delta_tx) is clear, but the sentence 'the resulting projected feature sequence corresponds to the prediction window...' should explicitly state which timestamps are used as positive pairs in pre-training.
  2. [Sec. V-D, Table IV] The comparison 'WiFo-M2-Img w/ INet Weight' replaces the ResNet-34 backbone with ImageNet weights, but the GRU and projection layers are presumably still randomly initialized or trained from scratch. Please specify how these non-backbone components are initialized and trained in this ablation, so the comparison isolates the backbone pre-training rather than confounding it with untrained temporal layers.
  3. [Fig. 8] The figure caption says 'CE Performance of CENN and FCDAMP', but the two rows are not labeled in the caption. Please indicate which row corresponds to which baseline, and consider adding a legend for the curves (WiFo-M2, WiFo-M2-Img, WiFo-M2-LiDAR, baseline) directly in the figure.
  4. [Sec. V-C2, Eq. (20)] In Eq. (20), the received signal is written as Y = F_B^H F_R^H h x + F_B^H F_R^H N. Since F_B is set to an identity matrix during CE, the notation is fine, but the dimensions of x and N should be made explicit (Pilot sequence length T, noise matrix N∈C^{N×T}) to avoid confusion.
  5. [General] The phrase 'WiFo-M$^2$' uses math mode inconsistently with the running text (e.g., 'WiFo-M 2' appears in many places). Please standardize the notation. Also, Tables I and II would benefit from a short statement explaining how 'L5' and 'L6' merge multiple links with different antenna configurations, since the table entries list two configurations separated by a slash.

Circularity Check

0 steps flagged

No significant circularity: the downstream evaluation is held out and the pre-training objective is a standard supervised alignment, not a re-statement of the test metrics.

full rationale

WiFo-M2's derivation chain is: (i) process LiDAR/RGB into receiver-labeled representations (Sec. IV-B); (ii) extract frame-level features with ResNet/PointNet and model temporal dynamics with a GRU (Eqs. (6)-(8)); (iii) pre-train with ContraSoM, an InfoNCE contrastive objective aligning sensory features with WiFo-extracted CSI features (Eqs. (14)-(19)); (iv) freeze the backbone and train only small adapters/downstream schemes on T-T splits, evaluating on T-V and on unseen scenarios C1/S1/V1 (Secs. V-C and V-E). The reported test quantities (beam Top-k accuracy, CE/CI/CP NMSE) are not the same as the contrastive loss, so improvements are not guaranteed by construction. The temporal extrapolator (Eq. (8)) is supervised by future CSI features during pre-training, but its utility is measured through downstream tasks on held-out data, not by checking that the pre-training objective itself decreases. Self-citations (EIW [3], SoM [4], M3SC [26], SynthSoM [27], WiFo [23]) provide motivation, data, and a feature extractor/baseline; these are published resources and the paper benchmarks against WiFo rather than assuming its validity. No equation reduces to its own input, and no fitted test value is renamed as a prediction. The extrapolation error across the inter-frame gap is a real robustness concern, but it is a correctness/validation gap, not circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The paper introduces no new physical entities, forces, or conserved quantities. Its load-bearing assumptions are the correctness of the sensory-to-CSI positive-pair construction, the reliability of GRU-based temporal extrapolation, and the transferability of contrastively aligned features to frozen-backbone downstream tasks. The main hand-set knobs are contrastive/diffusion loss weights, diffusion sampler parameters, and unstated pipeline thresholds such as y_build and sequence lengths.

free parameters (7)
  • Contrastive temperature tau_e = 0.1
    Sec. V-B; controls the sharpness of the InfoNCE distribution and is chosen by hand.
  • LiDAR contrastive loss weights lambda_1, lambda_2, lambda_3 = 1, 0.25, 0.25
    Eq. (17), Sec. V-B; balances original vs diffusion-augmented views.
  • Diffusion loss weight lambda_diff = 0.3
    Eq. (19); balances contrastive and diffusion training objectives.
  • Diffusion reverse sampler stride u and number of steps U = u=8, U=24
    Eq. (12), Sec. IV-D1; chosen by hand for efficient deterministic reverse generation.
  • Diffusion noise initialization levels m1, m2 = uniform over {200,400}
    Eq. (13); chosen to generate two distinct augmented views.
  • Building-point threshold y_build = not specified
    Sec. IV-B2 uses a geometric rule B={p_i : y_i > y_build}; the threshold is scenario-specific and not quantified.
  • Sensory sequence lengths n_I and n_L = not specified explicitly
    Eqs. (6)-(7); the historical window lengths for image and LiDAR are load-bearing for temporal extrapolation but their values are not stated.
axioms (6)
  • domain assumption Angular-encoding of the matched bounding box creates valid positive pairs between processed sensory frames and CSI of the specific receiver
    Sec. IV-B, Steps 1-3; if YOLOv8 detection, tracking, or angular matching fails, the contrastive pairs are incorrect and alignment degrades.
  • domain assumption Future OOB channel-aware features can be extrapolated from historical sensory sequences across the inter-frame gap
    Sec. IV-C Stage 3 and Fig. 5; the GRU extrapolation is supervised during pre-training, but the paper assumes scene dynamics change slowly enough for reliable look-ahead.
  • domain assumption CSI features from the pre-trained WiFo model [23] are a sufficient representation of the channel for contrastive alignment
    Sec. IV-C1; the WiFo encoder is fixed and self-cited; if its STF features are biased or lossy, all aligned sensory features inherit that bias.
  • domain assumption Contrastive alignment in latent space transfers to diverse downstream PHY actions without fine-tuning the backbone
    Sec. V-C; the plug-and-play claim rests on this transfer, with only small adapters trained per task.
  • domain assumption Diffusion-augmented LiDAR features remain valid positive samples for contrastive learning
    Sec. IV-D1; if the feature-space augmentation destroys spatial/semantic structure, the contrastive objective is corrupted.
  • standard math Standard deep-learning components (ResNet, PointNet, GRU, InfoNCE, DDPM) are correct and applicable as used
    Used throughout Sec. IV; not re-derived, but standard in the literature.

pith-pipeline@v1.3.0-alltime-deepseek · 20140 in / 12396 out tokens · 120556 ms · 2026-08-03T10:39:12.413105+00:00 · methodology

0 comments
read the original abstract

The emerging convergence of next-generation wireless networks and agentic artificial intelligence (AI) is inspiring a new vision: embodied intelligent network entities utilize environmental sensing to refine their physical-layer (PHY) actions. Despite a growing body of preliminary work, prevailing small and task-specific AI models require extensive manual design of data pre-processing, network architecture, and fine-tuning, leaving them tightly coupled to particular PHY actions, system configurations, and deployment scenarios. To address this, we propose a paradigm shift with WiFo-M$^2$, a foundation model that enables environment sensing to be easily integrated into PHY actions, delivering universal performance gains. To extract generalizable out-of-band (OOB) channel-aware features from environment sensing, we introduce ContraSoM, a contrastive pre-training strategy. Once pre-trained, WiFo-M$^2$ infers future OOB channel-aware features from historical sensory data and strengthens feature robustness via modality-specific data augmentation. Experiments show that WiFo-M$^2$ improves the performance of a comprehensive suite of fundamental PHY actions, demonstrating strong generalization to unseen scenarios.

Figures

Figures reproduced from arXiv: 2601.09179 by Haotian Zhang, Shijian Gao, Xiang Cheng.

Figure 1
Figure 1. Figure 1: Comparisons between conventional RF-based/sensing-refined wireless communications paradigm, and the proposed [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An overview of the considered scenario. IV. WIFO-M2 FRAMEWORK: SENSING PROCESSING, FEATURE EXTRACTION, AND PRE-TRAINING In this section, we first analyze the fundamental challenges faced by developing WiFo-M2 to achieve scalable environ￾ment sensing-refined PHY actions for intelligent network entities. Then, we introduce the core designs: the unified environment sensing processing pipeline, the architectur… view at source ↗
Figure 3
Figure 3. Figure 3: Overall framework of WiFo-M2 , illustrating the two-stage pipeline: (left) pre-training stage with ContraSoM strategy realizing SoM processing, and (right) inference stage for optimizing various PHY actions in a plug-and-play manner. B. Unified Environment Sensing Processing Pipeline 1) Image Processing Pipeline: The image processing pipeline converts raw images containing many potential re￾ceivers into a … view at source ↗
Figure 5
Figure 5. Figure 5: Temporal feature extrapolation in WiFo-M [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overview of multi-modal sensory data from all datasets. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The Top-k accuracies (k ∈ {1, 3, 5}) of the WiFo-M2 -based BP scheme, Vision-BP scheme, and MM-BP scheme. subcarriers. We assume that the BS adopts the hybrid (analog/digital) structure and is equipped with NRF RF chains. The digital and analog precoding matrices are represented by FB ∈ C NRF×NRF and FR ∈ C N×NRF with ∥FRFB∥ 2 F = NRF. FB is set as an identity matrix INRF during CE. The quantization bit nu… view at source ↗
Figure 8
Figure 8. Figure 8: CE Performance of CENN and FCDAMP with WiFo-M [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Performance gain brought by WiFo-M2 , WiFo-M2 -Img, and WiFo-M2 -LiDAR on LPCCNet and SR-CI. polated result into a super-resolution network to en￾hance the accuracy of CI. LPCCNet [34] is a Lapla￾cian pyramid-based channel completion network that progressively reconstructs the high-resolution channel image from a low-resolution input through a stepwise up-sampling architecture. Experiment Settings: We assu… view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. WiFo-MiSAC: A Wireless Foundation Model for Multimodal Sensing and Communication Integration via Synesthesia of Machines (SoM)

    eess.SP 2026-04 unverdicted novelty 6.0

    WiFo-MiSAC is a task-agnostic foundation model that unifies multimodal wireless signals via tokenization and self-supervised learning with SS-DMoE to achieve strong few-shot performance on beam prediction and channel ...

  2. Paradigm Shift from Statistical Channel Modeling to Digital Twin Prediction: An Environment-Generalizable ChannelLM for 6G AI-enabled Air Interface

    eess.SP 2026-04 unverdicted novelty 6.0

    ChannelLM-driven digital twin architecture reduces channel prediction error by 4.23 dB in unseen environments versus small AI models while achieving 70 ms end-to-end latency.

  3. Foundation Models for Wireless Communications: From PHY Intelligence to Network Autonomy

    eess.SP 2026-06 unverdicted novelty 2.0

    Surveys adaptation of foundation models to wireless tasks across off-the-shelf, wireless-native, and agentic paradigms for 6G PHY intelligence and network autonomy.

Reference graph

Works this paper leans on

38 extracted references · 3 linked inside Pith · cited by 3 Pith papers

  1. [1]

    Toward Edge General Intelligence With Agentic AI and Agentification: Concepts, Technologies, and Future Directions,

    R. Zhanget al., “Toward Edge General Intelligence With Agentic AI and Agentification: Concepts, Technologies, and Future Directions,”IEEE Commun. Surveys Tuts., vol. 28, pp. 4285–4318, 1st Quart., 2026

    2026

  2. [2]

    Embodied AI-Enhanced Vehicular Networks: An Inte- grated Vision Language Models and Reinforcement Learning Method,

    R. Zhanget al., “Embodied AI-Enhanced Vehicular Networks: An Inte- grated Vision Language Models and Reinforcement Learning Method,” IEEE Trans. Mobile Comput., vol. 24, no. 11, pp. 11494–11510, Nov. 2025

    2025

  3. [3]

    Embodied Intelligent Wireless (EIW): Synesthesia of Machines Empowered Wireless Communications,

    X. Chenget al., “Embodied Intelligent Wireless (EIW): Synesthesia of Machines Empowered Wireless Communications,” arXiv:2511.22845, 2025

    arXiv 2025

  4. [4]

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

    X. Chenget al., “Intelligent multi-modal sensing-communication inte- gration: Synesthesia of machines,”IEEE Commun. Surveys Tuts., vol. 26, no. 1, pp. 258–301, 1st Quart., 2024

    2024

  5. [5]

    Multi-Modal Sensing-Aided Channel Prediction for 6G mmWave Massive Antenna Systems,

    J. Moon, K. A. Ngo, B. Lee and B. Shim, “Multi-Modal Sensing-Aided Channel Prediction for 6G mmWave Massive Antenna Systems,”IEEE Trans. Cogn. Commun. Netw., vol. 12, pp. 2798–2813, 2026

    2026

  6. [6]

    Vision-Assisted Near- Field Channel Estimation for XL-MIMO Systems,

    Q. Zhang, Z. Zhou, Q. Liu and L. Yang, “Vision-Assisted Near- Field Channel Estimation for XL-MIMO Systems,”IEEE Trans. Cogn. Commun. Netw., vol. 12, pp. 2979–2988, 2026

    2026

  7. [7]

    Synesthesia of machines (SoM)-enhanced wideband multi-user CSI learning with LiDAR sens- ing,

    H. Zhang, S. Gao, X. Cheng and L. Yang, “Synesthesia of machines (SoM)-enhanced wideband multi-user CSI learning with LiDAR sens- ing,”IEEE Trans. Veh. Technol., vol. 74, no. 8, pp. 13237–13242, Aug. 2025

    2025

  8. [8]

    Integrated sensing and communications toward proactive beamforming in mmWave V2I via multi-modal feature fusion (MMFF),

    H. Zhang, S. Gao, X. Cheng and L. Yang, “Integrated sensing and communications toward proactive beamforming in mmWave V2I via multi-modal feature fusion (MMFF),”IEEE Trans. Wireless Commun., vol. 23, no. 11, pp. 15721–15735, Nov. 2024

    2024

  9. [9]

    Multi-modality sensing in mmWave beamforming for connected vehicles using deep learning,

    M. B. Mollah, H. Wang, M. A. Karim, and H. Fang, “Multi-modality sensing in mmWave beamforming for connected vehicles using deep learning,”IEEE Trans. Cogn. Commun. Netw., vol. 12, pp. 327–341, 2026

    2026

  10. [10]

    Camera based mmWave beam prediction: Towards multi-candidate real-world scenarios,

    G. Charan, M. Alrabeiah, T. Osman, and A. Alkhateeb, “Camera based mmWave beam prediction: Towards multi-candidate real-world scenarios,”IEEE Trans. Veh. Technol., vol. 74, no. 4, pp. 5897–5913, Apr. 2025

    2025

  11. [11]

    Vision Image-Aided Near-Field Beam Training for Internet of Vehicles Communication Systems: From Daytime to Nighttime,

    W. Huanget al., “Vision Image-Aided Near-Field Beam Training for Internet of Vehicles Communication Systems: From Daytime to Nighttime,”IEEE Trans. Cogn. Commun. Netw., vol. 12, pp. 2937–2949, 2026

    2026

  12. [12]

    Synesthesia of Machines (SoM)-Aided Online FDD Precoding via Heterogeneous Multi-Modal Sensing: A Vertical Federated Learning Approach,

    H. Zhang, S. Gao, W. Wen, X. Cheng, and L. Yang, “Synesthesia of Machines (SoM)-Aided Online FDD Precoding via Heterogeneous Multi-Modal Sensing: A Vertical Federated Learning Approach,”IEEE Trans. Mobile Comput., doi: 10.1109/TMC.2025.3648908, early access

    arXiv 2025

  13. [13]

    Learning transferable visual models from natural language supervision,

    A. Radfordet al., “Learning transferable visual models from natural language supervision,” inProc. Int. Conf. Mach. Learn. (ICML), 2021, pp. 8748–8763

    2021

  14. [14]

    DiffCL: A Diffusion- Based Contrastive Learning Framework With Semantic Alignment for Multimodal Recommendations,

    Q. Song, J. Hu, L. Xiao, B. Sun, X. Gao and S. Li, “DiffCL: A Diffusion- Based Contrastive Learning Framework With Semantic Alignment for Multimodal Recommendations,”IEEE Trans. Neural Netw. Learn. Syst., vol. 36, no. 10, pp. 18587–18597, Oct. 2025

    2025

  15. [15]

    Contrastive Reg- istration for Unsupervised Medical Image Segmentation,

    L. Liu, A. I. Aviles-Rivero and C. -B. Sch ¨onlieb, “Contrastive Reg- istration for Unsupervised Medical Image Segmentation,”IEEE Trans. Neural Netw. Learn. Syst., vol. 36, no. 1, pp. 147–159, Jan. 2025

    2025

  16. [16]

    Multi-modal graph contrastive learning for micro-videorecommendation,

    Z. Yi, X. Wang, I. Ounis, and C. Macdonald, “Multi-modal graph contrastive learning for micro-videorecommendation,” inProc.45th Int. ACM SIGIR Conf. Res. Develop. Inf. Retrieval, 2022, pp. 1807–1811

    2022

  17. [17]

    A MIMO Wireless Channel Foundation Model via CIR-CSI Consistency

    J. Jiang, W. Yu, Y . Li, Y . Gao, and S. Xu, “A MIMO Wireless Channel Foundation Model via CIR-CSI Consistency”, arXiv:2502.11965, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Large Wire- less Localization Model (LWLM): A Foundation Model for Positioning in 6G Networks

    G. Pan, H. Chen, S. Zhang, C. Hager, and H. Wymeersch, “Large Wire- less Localization Model (LWLM): A Foundation Model for Positioning in 6G Networks”, arXiv:2505.10134, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    When Vision-Language Model (VLM) Meets Beam Prediction: A Multimodal Contrastive Learning Framework

    J. Wang,et al., “When Vision-Language Model (VLM) Meets Beam Prediction: A Multimodal Contrastive Learning Framework”, arXiv preprint arXiv:2508.00456, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    Wireless Multimodal Foundation Model (WMFM): Integrat- ing Vision and Communication Modalities for 6G ISAC Systems

    M. Farzanullah, H. Zhang, A. B. Sediq, A. Afana, M. Erol- Kantarci, “Wireless Multimodal Foundation Model (WMFM): Integrat- ing Vision and Communication Modalities for 6G ISAC Systems”, arXiv:2512.23897, 2025

    arXiv 2025

  21. [21]

    Deep residual learning for image recognition,

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

    2016

  22. [22]

    PointNet: Deep learning on point sets for 3D classification and segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “PointNet: Deep learning on point sets for 3D classification and segmentation,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2017, pp. 77–85

    2017

  23. [23]

    WiFo: Wireless Foundation Model for Channel Prediction,

    B. Liu, S. Gao, X. Liu, X. Cheng, and L. Yang, “WiFo: Wireless Foundation Model for Channel Prediction,”Sci. China Inf. Sci., vol. 68, no. 1, pp. 1–13, Jun. 2025

    2025

  24. [24]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inProc. Int. Conf. Neural Inf. Process. Syst. (NIPS), 2020, pp. 6840– 6851

    2020

  25. [25]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2020, pp. 9729–9738

    2020

  26. [26]

    M 3SC: A Generic Dataset for Mixed Multi-Modal (MMM) Sensing and Communication Integration,

    X. Chenget al., “M 3SC: A Generic Dataset for Mixed Multi-Modal (MMM) Sensing and Communication Integration,”China Communica- tions, vol. 20, no. 11, pp. 13–29, Nov. 2023

    2023

  27. [27]

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

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

    2025

  28. [28]

    Multimodal deep learning empowered millimeter-wave beam prediction,

    B. Shi, M. Li, M.-M. Zhao, M. Lei, and L. Li, “Multimodal deep learning empowered millimeter-wave beam prediction,” inProc. IEEE 99th Veh. Technol. Conf. (VTC-Spring), Jun. 2024, pp. 1–6

    2024

  29. [29]

    SynthSoM- Twin: A Multi-Modal Sensing-Communication Digital-Twin Dataset for Sim2Real Transfer via Synesthesia of Machines

    J. Chen, Z. Huang, X. Cai, X. Cheng, and L. Yang,“SynthSoM- Twin: A Multi-Modal Sensing-Communication Digital-Twin Dataset for Sim2Real Transfer via Synesthesia of Machines”, arXiv:2511.11503, 2025

    arXiv 2025

  30. [30]

    ViWi: A Deep Learning Dataset Framework for Vision-Aided Wireless Communica- tions,

    M. Alrabeiah, A. Hredzak, Z. Liu, and A. Alkhateeb, “ViWi: A Deep Learning Dataset Framework for Vision-Aided Wireless Communica- tions,” inProc. IEEE 91st Veh. Technol. Conf. (VTC-Spring), May. 2020, pp. 1–5

    2020

  31. [31]

    DeepSense 6G: A large-scale real-world multi- modal sensing and communication dataset,

    A. Alkhateebet al., “DeepSense 6G: A large-scale real-world multi- modal sensing and communication dataset,”IEEE Commun. Mag., vol. 61, no. 9, pp. 122–128, Sep. 2023

    2023

  32. [32]

    Sparse Channel Estimation and Hybrid Precoding Using Deep Learning for Millimeter Wave Massive MIMO,

    W. Ma, C. Qi, Z. Zhang, and J. Cheng, “Sparse Channel Estimation and Hybrid Precoding Using Deep Learning for Millimeter Wave Massive MIMO,”IEEE Trans. Commun., vol. 68, no. 5, pp. 2838–2849, May. 2020

    2020

  33. [33]

    Deep learning-based beamspace channel estimation in mmWave massive MIMO systems,

    Y . Zhang, Y . Mu, Y . Liu, T. Zhang, and Y . Qian, “Deep learning-based beamspace channel estimation in mmWave massive MIMO systems,” IEEE Wireless Commun. Lett., vol. 9, no. 12, pp. 2212–2215, Dec. 2020

    2020

  34. [34]

    Deep Learning Super- Resolution-Based Channel Completion for Massive MISO Systems,

    K. Zu, Y . He, H. Chen, Y . Zheng and M. Haardt, “Deep Learning Super- Resolution-Based Channel Completion for Massive MISO Systems,” IEEE Signal Process. Lett., vol. 32, pp. 2254–2258, 2025

    2025

  35. [35]

    Channel estimation in IRS-enhanced mmWave system with super-resolution network,

    Y . Wang, H. Lu, and H. Sun, “Channel estimation in IRS-enhanced mmWave system with super-resolution network,”IEEE Commun. Lett., vol. 25, no. 8, pp. 2599–2603, Aug. 2021

    2021

  36. [36]

    Nerf2: Neural radio-frequency radiance fields,

    X. Zhao, Z. An, Q. Pan, and L. Yang, “Nerf2: Neural radio-frequency radiance fields,” inProc. ACM Annu. Int. Conf. Mobile Comput. Netw. (MobiCom), 2023, pp. 1–15

    2023

  37. [37]

    Fire: enabling reciprocity for fdd mimo systems,

    Z. Liu, G. Singh, C. Xu, and D. Vasisht, “Fire: enabling reciprocity for fdd mimo systems,” inProc. ACM Annu. Int. Conf. Mobile Comput. Netw. (MobiCom), 2021, pp. 628–641

    2021

  38. [38]

    Accurate Channel Prediction Based on Transformer: Making Mobility Negligible,

    H. Jiang, M. Cui, D. W. K. Ng and L. Dai, “Accurate Channel Prediction Based on Transformer: Making Mobility Negligible,”IEEE J. Sel. Areas Commun., vol. 40, no. 9, pp. 2717–2732, Sept. 2022

    2022