Wireless Foundation Models

• Wireless foundation models are large-scale, self-supervised neural architectures that learn universal representations from heterogeneous wireless data such as CSI and I/Q signals.
• They leverage transformer backbones and masked modeling techniques with physics-aware modules to capture electromagnetic constraints and enhance real-world robustness.
• These models enable scalable, multi-task support for tasks like channel estimation, beamforming, and localization, reducing retraining needs and optimizing resource usage in 6G networks.

A wireless foundation model (WFM) is a large-scale, self-supervised neural architecture that learns universal, task-agnostic representations from massive, heterogeneous wireless data, such as channel state information (CSI), raw I/Q signals, spectrograms, or network telemetry. Once pre-trained, a WFM supports a wide range of downstream wireless communication and sensing tasks—including channel estimation, prediction, beamforming, resource allocation, positioning, and wireless sensing—via zero-shot, few-shot, or lightweight fine-tuning, abstracting and unifying the fragmented module-specific solutions that have historically dominated the wireless domain (Liang et al., 4 Jun 2026, Alikhani et al., 2024). Unlike language or vision foundation models, the development of wireless FMs must contend with unique electromagnetic (EM) constraints, physical-layer latency/energy requirements, and propagation-intrinsic statistical structure.

1. Foundations and Motivations

Wireless foundation models have emerged to overcome the narrow, data-hungry, and non-generalizable nature of prior AI solutions in wireless communications (Liu et al., 27 Nov 2025, Liang et al., 4 Jun 2026). Traditional deep learning methods are typically task- or scenario-specific, necessitating retraining for every change in channel configuration, antenna geometry, pilot pattern, or deployment scenario. This leads to unmanageable storage/computation overheads and poor adaptability in heterogeneous 6G networks. WFMs, by contrast, offer:

• Universal representations: Task-agnostic embeddings that capture multipath structure, spatial/frequency dependencies, and device/environment variations (Alikhani et al., 2024, Liu et al., 27 Nov 2025).
• Scalability and adaptation: Zero- or few-shot transfer to new environments, devices, or tasks, with minimal or no retraining (Liu et al., 2024, Sheng et al., 8 Jul 2025).
• Unified multi-tasking: Single-model support for communication, sensing, and localization functionalities (Aboulfotouh et al., 18 Apr 2025, Liu et al., 27 Nov 2025, Aboulfotouh et al., 19 Nov 2025).
• Resource efficiencies: Parameter sharing, reduced training/serving costs, and quantization/distillation for real-time, on-device deployment (Zhang et al., 6 Nov 2025, Alikhani et al., 2024).

2. Model Architectures, Pretraining, and Modalities

The prevailing design of WFMs follows the transformer or masked autoencoder (MAE) paradigm, with several notable variants:

• Transformer backbones: Apply self-attention to sequences of CSI patches, I/Q tokens, or device features, capturing long-range dependencies and task-agnostic context (Alikhani et al., 2024, Liu et al., 27 Nov 2025, Sheng et al., 8 Jul 2025).
• Masked channel/signal modeling: Employ random or structured masking in space/time/frequency, reconstruct masked elements via MSE-based objectives, enforcing universal priors across wireless dimensions (Liu et al., 2024, Liu et al., 27 Nov 2025, Alikhani et al., 2024).
• Mixture of Experts (MoE): Use sparse MoE layers (e.g., CSI-SMoE in WiFo-2, S-R MoE in WiFo-CF) for scalable capacity and specialization to data heterogeneity (Liu et al., 27 Nov 2025, Xuanyu et al., 6 Aug 2025).
• Linear-time backbones: Integrate hybrid state-space models (e.g., WiMamba, ComHymba) for efficient sequence modeling at large channel dimensions, drastically reducing computation vs. standard transformers (Raviv et al., 27 Mar 2026, Yang et al., 22 May 2026).
• Physics-aware modules: Explicitly encode physics constraints, such as wave equivariance, as inductive biases to align learned representations with underlying EM propagation invariances, enhancing generalization, especially across sim-to-real gaps (Wang et al., 27 Jun 2026).
• Multimodal fusion: Jointly process I/Q, spectrogram, and CSI image-like modalities using shared ViT-based architectures with lightweight modality-specific embeddings (Aboulfotouh et al., 19 Nov 2025, Galeati, 15 Apr 2026).

Pretraining relies on large-scale, diverse, heterogeneous datasets: simulated (3GPP/QuaDriGa/DeepMIMO), ray-traced, or over-the-air measured CSI and I/Q streams, spanning frequencies from sub-6 GHz to THz, array sizes from SISO to massive MIMO, and environments from rural to dense urban (Liu et al., 27 Nov 2025, Xuanyu et al., 6 Aug 2025, Alikhani et al., 2024).

3. Fundamental Limits and Scaling Laws

WFMs are fundamentally constrained by the intrinsic nonlinear (manifold) dimension $d_{\mathrm{NL}}$ of the wireless propagation environment, a physical property derived from Maxwell's equations, antenna aperture, bandwidth, and the number of resolvable scatterers (Cheng et al., 8 May 2026). Key findings:

• Intrinsic dimension $d_{\mathrm{NL}}$: Wireless channels occupy a manifold of dimension $d_{\mathrm{NL}}\approx 5$–$35$ in real-world environments, much lower than typical semantic model spaces in language (e.g., $d\sim1000$).
• Model scaling law: Performance improves rapidly with model size $M$ as $\mathrm{Err}(M) \propto M^{-\alpha}$, $\alpha\approx 0.41$ at small $M$, but saturates when $M\gtrsim O(10^{d_{\mathrm{NL}}})$ (typically $d_{\mathrm{NL}}$0–$d_{\mathrm{NL}}$1M parameters) (Cheng et al., 8 May 2026).
• Pilot-aided test-time adaptation: Beyond scaling ceilings, test-time training (TTT)—adaptation of decoder parameters using pilot symbols—affords greater accuracy gains per FLOP than further increasing model size (Cheng et al., 8 May 2026).
• Physics constraints dominate: Channel geometry, not compute budget, sets the upper limit for beneficial WFM scaling.

4. Algorithmic Frameworks and System Integration

WFMs are integrated into wireless systems following both centralized and federated learning paradigms:

• Cloud–Edge–Device orchestration: Three-tier architectures with cloud FMaaS providers, edge servers hosting quantized FM weights, and device-level lightweight adaptation (e.g., LoRA, adapters, prompt vectors) (Chen et al., 2023).
• Federated learning with FM personalization: Hierarchical FL (federated adapter tuning, split-model training, and personalization via PEFT/LoRA) allows model adaptation without large-scale weight syncing, reducing communication overhead in resource-constrained wireless networks (Chen et al., 2023).
• Split-model/edge inference: Partition model computation between device front-end and edge back-end for privacy/computation trade-offs (Chen et al., 2023).

Fundamental resource-accuracy trade-offs are formalized as constrained optimizations over per-round communication $d_{\mathrm{NL}}$2, computation $d_{\mathrm{NL}}$3, and on-device storage $d_{\mathrm{NL}}$4, subject to network bandwidth $d_{\mathrm{NL}}$5 and device power $d_{\mathrm{NL}}$6 (Chen et al., 2023).

5. Downstream Applications and Performance

WFMs have demonstrated state-of-the-art accuracy and versatility across numerous downstream tasks without retraining or with minimal adaptation:

Task	Typical Metrics	Representative Models	Zero-/Few-shot Results
Channel pred./est.	NMSE [dB], SE [bps/Hz]	WiFo, WiFo-2, ComHymba	WiFo-2 outperforms full-shot baselines by 2.5–3 dB NMSE (Liu et al., 27 Nov 2025)
CSI feedback	NMSE, latency, ESE	WiFo-CF	WiFo-CF (zero-shot) surpasses fully-trained baselines on OOD configs (Xuanyu et al., 6 Aug 2025)
Beam prediction	Top-1 accuracy, SE	LWM, WiFo-2	LWM/SWiFo-2 reduce required training data by up to 60% (Alikhani et al., 2024, Liu et al., 27 Nov 2025)
Wireless localization	MAE [m]	LWLM	25–87% reduction in error vs. non-pretrained baselines (Pan et al., 15 May 2025)
Ambient intelligence	AUROC, F1	AM-FM	Outperforms scratch models on 9 tasks, robust to unseen devices/env (Zhu et al., 4 Feb 2026)
Multi-modal tasks	Acc./MAE across modals	Multimodal WFM, WavesFM	Matches/bests single-modality models in sensing, classification (Aboulfotouh et al., 19 Nov 2025, Aboulfotouh et al., 18 Apr 2025)

• Hardware realization: Quantized (e.g., INT8) WFMs deployed on embedded hardware (Jetson AGX Orin) achieve inference in $d_{\mathrm{NL}}$71–3ms, supporting real-time PHY-layer operation (Zhang et al., 6 Nov 2025, Liu et al., 27 Nov 2025).
• Transfer and reliability: WFMs yield reliabity-aware predictions (e.g., NMSE confidence tokens), enabling cross-task adaptation for handover, scheduling, and resource allocation (Liu et al., 27 Nov 2025).
• Multi-tasking: One WFM backbone can support >20 tasks with <10% of the parameters/training cost of dedicated models per task (Liu et al., 27 Nov 2025, Sheng et al., 8 Jul 2025).

6. Advanced Topics: Physical Inductive Biases and Explainability

Learning wireless physics, not just statistical correlations, is critical for robust, generalizable WFMs:

• Wave equivariance: Enforcing equivariance to phase, time, frequency, or spatial ramps (wave propagation symmetries) ensures that networks generalize across simulated and real-world channel distributions, closing the sim-to-real gap (Wang et al., 27 Jun 2026).
• Plug-and-play equivariant modules: Lightweight modules before/after the backbone align input CSI to remove arbitrary phase trends and restore physical invariance (Wang et al., 27 Jun 2026).
• Explainable generalization: Empirical results show that physics-informed WFMs (phys-WFM) can reduce normalized MSE by 5–15dB on unseen environments compared to vanilla transformer-based models (Wang et al., 27 Jun 2026).

7. Research Challenges and Future Directions

Developing robust, scalable WFMs opens several technical frontiers (Chen et al., 2023, Liang et al., 4 Jun 2026, Cheng et al., 8 May 2026):

1. Heterogeneity: Devices and data sources vary in compute, memory, and sample distribution. Adaptive client selection, resource-aware FL, and robust optimizers are needed.
2. Scalability: FM parameter sizes may be 10^8–10^12; parameter-efficient tuning, expert routing, and knowledge distillation (Tiny-WiFo) mitigate synchronization/computation overhead (Zhang et al., 6 Nov 2025).
3. Privacy/Security: Model inversion, prompt leakage, and adaptation in federated settings require secure aggregation and differential privacy.
4. Physics-grounded inductive bias: Systematic construction of frameworks respecting physical EM laws (beyond wave equivariance), including causality, sparsity, and hardware impairments.
5. Resource management: Joint optimization across bandwidth, power, and FM service-level agreements via cross-layer models.
6. Reliability and trustworthiness: Preventing model hallucinations, quantifying uncertainty, and hybrid symbolic-neural evaluation strategies (Chen et al., 2023, Liang et al., 4 Jun 2026).
7. Benchmarking and datasets: Lifting reliance on simulation while enabling open, standardized, high-fidelity measurement repositories for pretraining at true “foundation” scale (Liu et al., 27 Nov 2025, Liang et al., 4 Jun 2026).
8. Edge-device collaboration and orchestration: Split computing and real-time adaptation for seamless cloud–edge–device intelligence (Chen et al., 2023).
9. Agentic foundation models: Integration of reasoning, multi-agent coordination, and autonomous orchestration, moving beyond prediction to full-scale network management (Liang et al., 4 Jun 2026).

In summary, wireless foundation models provide a unifying abstraction for AI-native wireless networks: scalable, physics-informed, and adaptable engines for channel acquisition, prediction, sensing, and control. Ongoing research into model scaling, domain-aligned inductive biases, and resource-efficient orchestration will be foundational to the evolutionary path of 6G and beyond (Chen et al., 2023, Liu et al., 27 Nov 2025, Cheng et al., 8 May 2026, Wang et al., 27 Jun 2026, Raviv et al., 27 Mar 2026, Alikhani et al., 2024, Liang et al., 4 Jun 2026).