PaperlayerView exact revision on arXiv

Source-linked AI summary

Wireless Foundation Models: State-of-the-Art and Open Challenges

Alonso M. Pacheco Huachaca, Juan J. Rodriguez Rodriguez, Ahmed Aboulfotouh, Nelson L. S. da Fonseca, Carlos A. Astudillo, Hatem Abou-Zeid

arXiv:2609.04707v1eess.SPcs.AI

TL;DR

Wireless foundation-model research is fragmented across data, objectives, architectures, adaptation, and evaluation, complicating assessment of broad transfer. This survey systematically compares PHY WFMs across task families and design dimensions, finding growing but uneven evidence of reusable representations and identifying controlled evaluation and broader data as priorities.

  • Problem

    WFM literature is fragmented across modalities, objectives, architectures, adaptation strategies, and evaluation protocols, making broadly transferable models difficult to assess.

  • Method

    The survey analyzes WFM pretraining, architectures, adaptation, datasets, tasks, and distribution shifts across five PHY task families and multitask models.

  • Results

    33 of 66 dataset entries are synthetic, 31 are real, 30 are reused for pretraining and fine-tuning or evaluation, and 22 are explicitly public.

  • Takeaways & Limitations

    Reusable wireless representations are increasingly reported, but evidence varies across task families and evaluation settings and cannot yet be compared consistently.

Abstract

from arXiv · show

Wireless foundation models (WFMs) have emerged as a promising approach for learning reusable representations from large-scale wireless data and adapting them to downstream tasks. However, the rapidly growing literature remains fragmented across modalities, pretraining objectives, architectures, adaptation strategies, and evaluation protocols, making it difficult to assess progress toward broadly transferable models. This survey provides a systematic analysis of WFMs for physical-layer applications. We first introduce the main WFM design components, including pretraining, backbone architectures, and downstream adaptation. We then organize the literature into five physical-layer task families: signal recognition and demodulation, channel representation learning, RF sensing and localization, beam management, and spectrum sensing and monitoring, while separately examining multi-task PHY models. Across these categories, we analyze how existing models are pretrained, adapted, and evaluated, with particular attention to downstream task diversity and the distinction between in-distribution, partial-shift, and out-of-distribution transfer. Our analysis shows that current WFMs provide increasing evidence of reusable wireless representations, but this evidence varies considerably across task families and evaluation settings. Differences in datasets, modalities, architectures, pretraining objectives, adaptation protocols, and distribution shifts make it difficult to determine which design choices drive transfer and generalization. We conclude by identifying open directions for improving data availability, evaluation rigor, generalization, efficient adaptation, and real-world deployment, providing a unified framework for understanding the current WFM landscape and the requirements for developing more reusable foundation models for future physical-layer wireless systems.

I. INTRODUCTION

The survey addresses fragmented, task-specific WFM research by organizing physical-layer models around their design choices, downstream tasks, and transfer evaluation. It uses a structured literature-selection process to compare models and identify limitations and open directions.

  • Existing wireless deep-learning models often target fixed tasks, datasets, channel conditions, devices, or environments, limiting robustness across diverse 6G conditions.
  • WFMs aim to learn reusable wireless representations that adapt across tasks and deployment conditions while reducing dependence on task-specific labeled data.
  • The survey organizes PHY WFMs into signal recognition and demodulation, channel representation learning, RF sensing and localization, beam management, and spectrum sensing and monitoring.
  • Studies were selected through identification, screening, eligibility assessment, and inclusion, using criteria requiring wireless-domain pretraining and evaluation on at least one concrete PHY task.
  • Its comparison examines pretraining, modalities, architectures, adaptation strategies, datasets, downstream tasks, and distribution-shift settings.
  • The paper identifies limitations and open directions for scalable, generalizable, and deployable WFMs for next-generation wireless systems.

II. RELATED WORK

The survey positions itself as a specialized, task-centered analysis of wireless-native physical-layer foundation models. It distinguishes broad communications-AI surveys from model-level comparisons of reusable representations, adaptation, and transfer.

  • Prior surveys commonly emphasize large AI models, AI-native wireless systems, or broad 6G visions rather than wireless-native pretrained models.
  • Existing specialized surveys provide depth in areas such as channel modeling or prediction and control but do not consolidate PHY WFMs across multiple task families.
  • This survey uses PHY task families as the unit of analysis for systematic comparison across model design and evaluation.
  • What is a Wireless Foundation Model?: It applies a stricter foundation-model interpretation requiring general structure, reusable representations across downstream tasks, and meaningfully different task capabilities.
  • Formal Definition: The formal framework maps wireless observations through a shared backbone into latent representations, then uses lightweight task-specific heads for downstream predictions.
  • Evaluation: The survey distinguishes in-distribution, partial-shift, and out-of-distribution evaluation, with OOD settings providing stronger evidence of foundation-model behavior.

B. Pretraining Strategies

WFM pretraining spans supervised and self-supervised strategies designed to learn reusable wireless representations from data whose labels are costly. The survey contrasts objective families by their learning signals, uses, and limitations.

  • Wireless labels are often difficult to obtain, motivating self-supervised pretraining from raw I/Q samples, CSI, CIRs, spectrograms, radio maps, and sensing data.
  • Contrastive Learning: Contrastive learning aligns views of the same wireless instance and separates views from different instances, but depends strongly on augmentation and pair definitions.
  • Masked Reconstruction: Masked reconstruction recovers hidden input regions from visible context, exploiting dependencies across time, frequency, antennas, or space.
  • Latent Prediction: Latent prediction forecasts missing, future, or related representations without necessarily using negative samples or reconstructing raw waveforms.
  • Generative Pretraining: Generative pretraining produces complete or refined wireless-domain outputs through autoregressive or denoising processes, but physical consistency requires additional validation.
  • Temporal Prediction: Temporal prediction uses past observations to forecast future measurements, while its generalization across time scales must be explicitly evaluated.
  • Distillation: Distillation transfers teacher predictions or representations to a student model, supporting stronger representations or model compression under resource constraints.

7) Hybrid:

Hybrid and task-driven objectives combine complementary supervision signals or downstream task information, while backbone architectures determine how wireless inputs are represented and processed. The survey covers architectures ranging from CNNs and Transformers to multimodal, geometry-aware, diffusion, and lightweight MLP designs.

  • Hybrid: Hybrid objectives jointly optimize multiple losses to capture complementary wireless structure, but their representations are harder to interpret and tune.
  • Task-Driven Supervision: Task-driven supervision aligns a shared backbone with related wireless functions using labels, task descriptions, or task-specific losses, but may remain tied to included tasks and scenarios.
  • Backbone Architectures: CNNs capture local correlations in grid-like representations such as I/Q windows, spectrograms, and CSI matrices.
  • Backbone Architectures: Transformers dominate WFM backbones because attention can model long-range dependencies across structured wireless inputs.
  • Backbone Architectures: Mixture-of-Experts Transformers increase capacity by activating only selected experts for each input, enabling specialization across wireless conditions.
  • Backbone Architectures: CNN–Transformer, dual-branch, multimodal, and geometry-aware designs combine local extraction, broad context, modality-specific processing, or structural relationships.
  • Backbone Architectures: Diffusion Transformers support iterative wireless-signal refinement, whereas MLP encoders prioritize latency and parameter efficiency for edge-oriented WFMs.

4) RNN-based:

This section situates recurrent and state-space backbones within WFM design, then organizes physical-layer tasks and adaptation strategies for comparing transfer across settings.

  • Backbone architectures: RNNs model temporal dependencies through hidden-state updates and can combine with convolutions to capture local structure and temporal evolution.They are useful for time-evolving inputs such as spectrum traces and temporal CSI.
  • Backbone architectures: State-space models provide structured sequence processing, while bidirectional Mamba is adapted to capture relationships across multidimensional CSI.This avoids imposing a one-way temporal order on antenna and subcarrier dimensions.
  • Adaptation strategies: Adaptation strategies range from zero-shot transfer through frozen-head methods and linear probing to parameter-efficient updates with adapters, prompts, or low-rank methods.The strategies differ in update cost and adaptation flexibility.
  • Task organization: The survey organizes physical-layer WFMs using a hierarchical taxonomy spanning signal recognition, channel representation, sensing and localization, beam management, spectrum monitoring, and multitask PHY models.Task-level categories support consistent comparison of pretraining, adaptation, and evaluation.
  • Signal recognition and demodulation: The taxonomy distinguishes signal identification from demodulation: the former recognizes signal type, whereas the latter recovers transmitted symbols or bits.The distinction separates recognition-oriented transfer from information-recovery objectives.
  • Channel representation learning: Channel representation tasks include estimating or recovering incomplete channel information, forecasting future channel states, diagnosing channel conditions, and compressing CSI for feedback.These tasks cover present-state inference, temporal prediction, descriptive diagnosis, and bandwidth-efficient reporting.

3) RF Sensing and Localization:

The survey defines RF sensing and localization as a major physical-layer task family and reports that the literature is growing, with especially frequent coverage of sensing, localization, and channel-oriented capabilities.

  • RF sensing and localization: RF sensing uses wireless signals to infer physical, spatial, or behavioral information about users and environments.The category treats wireless signals as sensing modalities rather than only communication carriers.
  • RF sensing and localization: Localization and spatial inference combines localization, angle-of-arrival estimation, and ranging because all extract spatial structure from physical-layer observations.The outputs differ, but the shared goal is spatial inference.
  • RF sensing and localization: The taxonomy separately covers activity recognition, RF fingerprinting, and environment reconstruction for behavior, device identity, and surrounding physical or radio structure.These distinctions separate human activity, transmitter-specific imperfections, and propagation-space reconstruction.
  • Beam management: Beam-management tasks select or adapt beams under changing channel or mobility conditions, while precoding focuses on transmit-side beamforming or spatial precoding vectors.The survey keeps precoding distinct because it is more directly tied to transmission design.
  • Spectrum sensing and monitoring: Spectrum sensing and monitoring cover activity detection, segmentation, forecasting, and anomaly detection across time and frequency.These tasks identify occupied regions, model spectrum dynamics, or detect abnormal RF patterns.
  • Literature coverage: RF Sensing and Localization appears in 30/49 (61.2%) models and Channel Representation Learning in 28/49 (57.1%), while spectrum tasks appear in 6/49 (12.2%).The distribution indicates stronger coverage of sensing and channel-oriented capabilities than spectrum-oriented WFMs.
  • Literature coverage: Channel Representation Learning is the dominant category for 18/49 (36.7%) models, followed by Multi-Task PHY at 14/49 (28.6%), while 35 of 49 models cover multiple fine-grained tasks.Cross-task evaluation is therefore common alongside dominant-category assignment.

C. Datasets

The survey compares WFM datasets by usage, modality, realism, accessibility, and labels, then relates modality and objective choices to transfer and adaptation across signal tasks.

  • Dataset overview: Among 66 dataset entries, 33 (50.0%) are synthetic, 31 (47.0%) real, and 2 (3.0%) combine both sources.Synthetic and real data are therefore nearly balanced in the reviewed dataset landscape.
  • Dataset overview: Dataset reuse is common: 30 entries (45.5%) serve both pretraining and fine-tuning or evaluation, while only 22 (33.3%) are explicitly public.Accessibility remains limited despite frequent reuse across the WFM pipeline.
  • Dataset overview: Channel-oriented representations account for 35 entries (53.0%), compared with 22 I/Q-related (33.3%) and 9 spectrogram-related (13.6%) entries.CSI, CIR, and channel-based data dominate the summarized dataset modalities.
  • Signal recognition and demodulation: Signal recognition and demodulation models use I/Q, spectrogram, and multimodal inputs, with mixed-modality inputs appearing in 7 of 12 models.Hybrid and masked-reconstruction objectives are most common, with 3 models each.
  • Data modalities: Raw I/Q preserves waveform information, whereas spectrograms organize time-frequency patterns but may hide phase information important for demodulation.Multimodal and channel-aware inputs add complementary waveform, time-frequency, or propagation context.
  • Dataset design: Controlled testbed data improve interpretability but limit diversity, while heterogeneous aggregation broadens variation and makes transfer gains harder to attribute.Variation can span technologies, devices, locations, sampling rates, frequencies, and channel conditions.
  • Pretraining objectives: SpectrumFM finds that a hybrid masked-reconstruction and next-slot-prediction objective performs best among the compared objectives.The comparison provides objective-level evidence rather than only evaluating a complete WFM pipeline.
  • Adaptation strategies: Lightweight adaptation is most convincing for recognition, whereas demodulation requires handling channel mixing, interference, noise, and overlapping user signals.The survey cautions against directly generalizing frozen-head or LoRA evidence from recognition to demodulation.

D. Evaluation Shift

Evaluation evidence varies by task and shift definition: recognition studies often broaden conditions without isolating the source of transfer, while channel evaluation must account for heterogeneous inputs and acquisition conditions. Stronger comparisons require separating dataset, modality, task, and adaptation effects.

  • Signal recognition and demodulation: OOD recognition comparisons broaden evaluation, but simultaneous changes in dataset, modality, signal regime, or task setup obscure the source of transfer.This limitation affects comparisons involving SpectrumFM, Multimodal WFM, and LatentWave.
  • Signal recognition and demodulation: WirelessJEPA improves over the contrastive IQFM baseline in most OOD cases, but the gain is not uniform across recognition targets.The result supports reporting where shifted transfer occurs, not only whether it occurs.
  • Signal recognition and demodulation: SID and Dem provide different transfer evidence: recognition tests class separability, whereas demodulation tests symbol recovery under changed communication conditions.Demodulation shifts include channel, interference, user, SNR, modulation, and antenna changes.
  • Channel representation learning: Channel WFM evaluation spans CSI or CIR and may incorporate maps, sensing data, spectrograms, RSRP, or scene-aware signals, creating heterogeneous transfer conditions.CSI is dominant, but its structure and reliability depend on acquisition, system configuration, and propagation conditions.
  • Channel representation learning: Scaling channel pretraining helps only when it covers the propagation and system variations relevant to deployment.WiFo reduces dependence on a single configuration, but transfer remains constrained by uncovered heterogeneity dimensions.
  • Channel representation learning: Masked reconstruction dominates channel pretraining, while prediction, generative, contrastive, and hybrid objectives encode different assumptions about useful channel structure.Controlled comparisons are needed because objectives preserve different properties and may not benefit every downstream task equally.

C. Adaptation Under Channel Shift

Channel adaptation depends on the downstream output space and on whether pretraining covers relevant variation. RF sensing and localization add modality and physical-context constraints, while masking and latent prediction emphasize different structures rather than guaranteeing task-agnostic transfer.

  • Channel adaptation: WiFo supports zero-shot evaluation across 12 datasets, while WiFo-CF handles multiple configurations through heterogeneous pretraining and an MoE architecture.WiFo-CF is evaluated under ID and OOD settings on simulated and real-world measurements.
  • Channel adaptation: Non-ideal CSI can confound transfer because clean, pilot-estimated, degraded, compressed, and feedback-constrained inputs differ before propagation structure is evaluated.Channel evaluations should vary CSI quality and acquisition conditions rather than rely only on ideal homogeneous inputs.
  • Channel adaptation: Channel prediction can require less retraining when pretraining captures diverse space–time–frequency variation, whereas estimation and CSI feedback require task-specific output mappings.Feedback models must satisfy channel dimensions, feedback rates, and reconstruction requirements.
  • RF sensing and localization: RF sensing WFMs use diverse inputs including CSI, CIR, I/Q, spectrograms, radar, and environmental or radio maps, making modality choice central to adaptation.A ViT-style model supports several RF sensing inputs, while radar introduces distinctive chirp, range-Doppler, and temporal-spatial structure.
  • RF sensing and localization: Masked reconstruction targets different physical structures depending on whether the model masks propagation delays, time–frequency patterns, or spatial RF fields.RadioFM, 6G-RadioFM, and FM-RME apply the objective to different sensing representations.
  • RF sensing and localization: Latent prediction is not task-agnostic by default because masking geometry selects different wireless dependencies and favors different downstream tasks.WirelessJEPA and CSI-JEPA report broader task evaluations, while LatentWave shows distinct effects from frequency versus region masking.

C. Model Adaptation and Efficiency

Adaptation efficiency in wireless foundation models includes label, measurement, parameter, computation, and inference costs. Existing evidence shows gains under some shifted settings, but physical-environment changes and multimodal contributions remain difficult to isolate.

  • Data and adaptation efficiency: RadioFM reaches comparable localization accuracy with ten times less reference data, reducing the cost of fingerprinting campaigns.Localization efficiency depends on target-domain reference measurements as well as trainable parameters.
  • Data and adaptation efficiency: FARM consistently outperforms RadioUNet across OOD shifts in altitude or coverage, carrier frequency, and antenna pattern.The evaluation concerns aerial radio-map reconstruction from sparse RSS observations.
  • Data and adaptation efficiency: 6G WavesFM shares one backbone across sensing, communication, and localization, with a reported convergence speedup of up to 5×.Task-specific heads and LoRA reduce the need for separate models, while CSI-MAE uses lightweight decoder fine-tuning and zero-shot transfer.
  • Evaluation boundaries: Most RF sensing evaluations provide limited evidence for physical-environment shifts such as changed layout, device placement, user behavior, or sensing geometry.New datasets alone do not establish transfer across changed sensing situations.
  • Evaluation boundaries: JEPA-based models show that downstream behavior depends strongly on the selected prediction target and masking geometry.This indicates that latent prediction alone does not guarantee a general-purpose wireless representation.
  • Evaluation boundaries: Multimodal gains remain difficult to attribute because additional I/Q, CSI, spectrogram, radar, or environmental inputs may themselves provide informative signals.Modality ablations and missing-modality tests are needed to separate representation gains from input-information gains.

B. Pretraining and Model Architectures

Beam-management pretraining increasingly shapes representations around beam-relevant channel differences, combining masked modeling, contrastive learning, and multimodal alignment. Results show that representation design and pretraining scale both affect transfer to beam prediction.

  • Beam-oriented representation learning: CSI–CIR alignment improves positioning error by 22% but beam-management accuracy by only 1% over supervised methods.The contrast highlights that useful channel alignment does not automatically preserve beam-discriminative information.
  • Beam-oriented representation learning: Scaling CSI–CIR alignment across broader DeepMIMO settings yields beam-prediction Top-1 gains of up to 19.31 percentage points.The result suggests broader exposure to beam-relevant channel variation can make alignment more useful for beam prediction.
  • Beam-oriented representation learning: WiMAE reaches 39.9% top-1 accuracy with only 1% of training data for codebook size 32, outperforming LWM and raw-channel baselines.ContraWiMAE further improves linear separability as beam-class complexity increases by adding a contrastive component to masked channel modeling.
  • System-level control: MMIMO-Prc-FM supports downstream precoding through a frozen backbone using masked CSI pretraining, but evaluation remains tied to the original data and system setting.This illustrates both the potential and the setting dependence of masked channel pretraining for system-level control.

C. Evaluation Shift

Beam-management evaluation ranges from controlled shifts that preserve label meaning to stronger cross-context transfer using environmental sensing. These settings test whether learned representations retain beam-relevant information beyond the original CSI conditions.

  • Controlled distribution shifts: Controlled shifts vary scenarios, bands, SNR levels, or simulated channel conditions while keeping beam-label construction compatible with channel and codebook assumptions.This design moves beyond exact pretraining conditions without making the downstream label incomparable.
  • Cross-context transfer: WiFo-M2 aligns multimodal sensing with CSI and evaluates beam prediction in unseen SynthSoM-Twin, DeepSense-6G, and ViWi scenarios.The model tests whether environmental sensing can serve as a proxy for beam-relevant channel structure when direct CSI is unavailable or incomplete.
  • Cross-context transfer: 91.0%, 100%, and 100% Top-1/3/5 beam accuracy are reported for the image-based WiFo-M2 variant with a frozen backbone in measured DeepSense-6G.The evaluation represents a substantially different sensing and dataset context from direct CSI-based transfer.

D. Adaptation Strategies

Adaptation results show that beam transfer often requires task-specific alignment, while spectrum models vary in how readily their representations support detection, classification, forecasting, segmentation, and zero-shot estimation. Across tasks, transfer depends on representation-task compatibility and system conditions.

  • Beam adaptation: With 50 labeled samples per class, LoRA reaches 52.6% beam-prediction accuracy versus 42.5% for supervised training and 35.5% for linear probing.With 500 samples per class, LoRA reaches 94.1%, while the supervised baseline reaches 89.5% and linear probing reaches 42.7%.
  • Beam adaptation: ContraWiMAE exceeds LWM by 42.3 percentage points under linear probing, but the gap narrows to 6.4 percentage points with a stronger ResNet-Wide head.A more expressive downstream model can recover part of the beam/codebook mapping from a weaker representation.
  • Beam adaptation: Beam transfer requires compatible antenna, user, channel-dimension, codebook, or deployment settings because beam labels are tied to the system configuration.Environment-aware inputs can help when CSI is missing, delayed, or unreliable.
  • Beam adaptation: Channel similarity alone does not guarantee beam discrimination because beam selection depends on mapping channel states to a specific codebook or precoding rule.Beam-oriented pretraining, broader alignment, or downstream adaptation is therefore needed to expose channel differences that change the PHY action.
  • Spectrum adaptation: Spectrum WFMs use spectrograms, I/Q, PSD, or spectrum maps, producing representations focused respectively on occupancy, signal dynamics, or spatial radio-environment structure.The survey table distinguishes these modalities and task-oriented formulations across spectrum sensing and monitoring models.
  • Spectrum adaptation: SSRadio converges faster after freezing its masked-spectrogram backbone, but remains slightly below the from-scratch forecasting baseline.Segmentation is further affected by dataset distribution differences and the mismatch between regression pretraining and classification.
  • Spectrum adaptation: Multimodal WFM reaches 96.40% for interference detection with linear probing and 99.60% with LoRA, while interference classification rises from 58.25% to 66.93%.The larger adaptation benefit for classification indicates that finer decision boundaries require more task-specific adjustment.
  • Spectrum adaptation: FM-RME transfers without fine-tuning to an unseen radio-map dataset with different temporal and spectral parameters and outperforms the compared baselines.Its geometry-aware pretraining and masked recovery target reusable spatial–temporal–spectral structure.

C. Evaluation Shift

Evaluation-shift evidence varies across spectrum and cross-category WFMs. Most studies remain close to the pretraining representation view, while stronger tests change data sources, interference conditions, spectrum views, or task settings.

  • Spectrum WFMs: Spectrum WFMs provide limited OOD evidence because most evaluations remain within a single spectrum view.SSRadio and 6G-RadioFM use spectrogram-based tasks, while SpectrumFM uses I/Q-based sensing and anomaly detection.
  • Spectrum WFMs: FM-RME tests transfer across radio-environment settings with different temporal and spectral parameters, but its spatial–temporal–spectral output limits direct comparison with other spectrum tasks.The evaluation concerns radio-map estimation rather than interference recognition or spectrogram segmentation.
  • Spectrum WFMs: Multimodal WFM evaluates heterogeneous pretraining against over-the-air LTE I/Q recordings with DSSS interference, providing a clearer shifted-setting test.This setting changes both the data source and interference condition relative to pretraining.
  • Spectrum WFMs: Geometry-aware feature extraction supports zero-shot radio-map estimation across changed radio-environment settings by retaining useful spatial propagation relationships.The evidence concerns transfer beyond the original map configuration.
  • Spectrum WFMs: Task difficulty differs across spectrum evaluations: activity or interference detection can be easier than interference-type classification, while segmentation adds time–frequency localization demands.Therefore, strong transfer on one spectrum task does not establish general spectrum reuse.
  • Cross-category PHY WFMs: Cross-category models test representation breadth, shared architectures, task conditioning, or physical context, but task count alone is a weak signal of generality.The surveyed models include broad downstream testing, shared backbones, task-specific interfaces, and context-aware designs.
  • Cross-category PHY WFMs: No surveyed cross-category work isolates transfer from shared representations versus task-specific conditioning or context injection.Controlled ablations separating these sources of gains remain absent.

A. OOD Evaluation for WFM Robustness

The survey identifies OOD evaluation, predictive pretraining, specialized architectures, efficient deployment, multimodal context, decision support, and wireless world models as directions for more robust WFMs. It concludes that current transfer evidence is promising but insufficiently comparable or broad to establish general reuse.

  • OOD evaluation: OOD evaluation should become central because many WFM studies test familiar sources or partial shifts rather than clearly different downstream distributions.Wireless-specific evaluation should distinguish meaningful distribution changes from shifts that remain close to pretraining conditions.
  • Pretraining: Predictive pretraining can encode wireless-specific structure through adaptive targets and masks emphasizing antenna, temporal, frequency, or spatial dependencies.The survey proposes moving beyond manually designed masking rules toward learned target selection.
  • Pretraining: Future predictive WFMs should model related dynamics jointly, such as how user movement changes channels and beam-selection requirements.Current approaches often predict channel conditions, angles, or traffic separately despite their physical relationships.
  • Architectures: Specialized backbones, including state-space and mixture-of-experts designs, should be evaluated for long-sequence efficiency and heterogeneous wireless configurations.Most reviewed WFMs still rely on Transformer or ViT backbones.
  • Deployment: Deployment evaluations should report latency, memory, trainable parameters, and adaptation time alongside downstream accuracy.These costs determine feasibility under edge and real-time constraints.
  • Multimodal and physical context: Multimodal and environment-aware inputs may improve context-aware transfer, but they complicate attribution because gains can arise from representations, added context, or adaptation protocols.The survey highlights maps, scenes, sensing data, and physical context as both useful and confounding factors.
  • Wireless World Models: Wireless world models would learn how channels, interference, spectrum activity, or beam quality evolve under changing deployment conditions.This extends WFM evaluation beyond controlled changes toward temporal and physical modeling for prediction, adaptation, monitoring, and resource management.
  • Conclusion: The survey finds growing but uneven evidence that wireless representations transfer across tasks, datasets, and operating conditions, while differing designs prevent clear attribution of transfer gains.Broader downstream evaluation alone is insufficient without controlled distribution shifts and adaptation comparisons.
Loading 2609.04707v1…