---
title: Channel Foundation Models (CFMs)
url: https://www.emergentmind.com/topics/channel-foundation-models-cfms
type: topic
---

# Channel Foundation Models (CFMs)

Searching arXiv for recent and foundational papers on Channel Foundation Models to ground the article.
Channel Foundation Models (CFMs) are domain-specific foundation models that learn universal, reusable representations of channel state information from large-scale, heterogeneous, and often unlabeled channel data, then transfer those representations across downstream tasks such as channel estimation, prediction, beamforming, positioning, CSI compression and feedback, and sensing. In the wireless literature, CFMs are described as pretrained, universal channel feature extractors and as large, transferable models trained on diverse CSI to learn a “universal channel representation,” with adaptation performed by full fine-tuning, lightweight fine-tuning, or zero-/few-shot transfer [2507.13637][2512.15562].

## 1. Conceptual foundations

The CFM literature frames wireless channels as a pretraining domain in their own right rather than as a sequence of isolated supervised tasks. A survey introducing the concept “for the first time” presents CFMs as a unified framework built around self-supervised learning on large unlabeled channel datasets, motivated by heavy dependence on labeled data, poor generalization under domain shift, and task silos across channel estimation, feedback, beam selection, localization, and sensing [2507.13637]. Related work sharpens that definition in several directions: predictive CFMs are trained on cross-domain CSI to produce transferable priors for current-slot estimation [2512.15562]; representation CFMs learn compact, task-agnostic embeddings from realistic multi-antenna channels [2505.09160]; generalized wireless foundation models aim to work without scenario-specific finetuning across reconstruction and prediction tasks [2601.18200].

A recurrent theme is that the wireless channel exhibits structured variability across space, time, frequency, propagation environment, numerology, antenna configuration, mobility, and SNR. The rationale for large-scale pretraining is therefore not merely data volume, but exposure to heterogeneous propagation statistics. In the predictive CFM setting, training across many TDL profiles, speeds, subcarrier spacings, pilot patterns, and SNRs is used to induce “environment-agnostic features” and strong cross-scenario transferability [2512.15562]. HeterCSI formulates the same issue as “dual heterogeneity”: scale heterogeneity in $T$, $K$, and $A$, and scenario diversity across indoor, urban, rural, LoS/NLoS, carrier frequency, and hardware [2601.18200].

The contrast with classical pipelines is explicit. Traditional supervised estimators learn direct mappings for a fixed task and often degrade under distribution shifts, whereas CFMs separate representation learning from downstream adaptation. In one formulation, they bring a “pretrain once, adapt everywhere” paradigm to the physical layer by learning channel-intrinsic features from large unlabeled corpora [2502.11965].

## 2. Self-supervised learning paradigms

The survey literature organizes CFM training into generative, discriminative, and hybrid paradigms, with predictive modeling now functioning as a closely related fourth pattern in task-oriented channel modeling [2507.13637][2512.15562].

| Paradigm | Core mechanism | Representative instances |
|---|---|---|
| Generative SSL | Masked reconstruction of CSI or related channel tensors | WiFo [2412.08908], WiMAE [2505.09160], CSI-MAE [2601.03789] |
| Discriminative SSL | Contrastive alignment of positive channel views or modalities | CSI-CLIP [2502.11965], CSI-CLIP++ [2606.25714] |
| Hybrid SSL | Joint reconstruction and contrastive learning | ContraWiMAE [2505.09160] |
| Predictive priors | Forecast future CSI from channel history, then refine with current evidence | Predictive CFM for channel estimation [2512.15562] |

Generative SSL is most commonly instantiated as masked channel modeling. The survey writes the masked objective as
$$
\mathcal{L}_{\text{mask}} = \lVert \mathbf{M}\odot \mathbf{x} - \mathbf{M}\odot f_\theta(\mathbf{x}) \rVert_2^2,
$$
and treats masked reconstruction as the dominant pretraining template for CFM-like systems [2507.13637]. WiFo unifies time-domain and frequency-domain channel prediction as reconstruction from partial observations,
$$
H = \Phi_{\mathrm{rec}}(H[\Omega]),
$$
with pretraining driven by random, time-masked, and frequency-masked reconstruction tasks over a space-time-frequency tensor [2412.08908]. CSI-MAE uses masked MSE over masked patches only, with a high mask ratio of $75\%$, while WiMAE reports that $mr = 0.60$ is optimal for forcing the encoder to learn global structure rather than local memorization [2601.03789][2505.09160].

Discriminative SSL enters the CFM literature mainly through physically paired views. CSI-CLIP treats CIR and CSI as naturally aligned multimodal data and uses CLIP-style contrastive learning to align them in a shared embedding space [2502.11965]. CSI-CLIP++ keeps the same physical idea but emphasizes scalability and stronger transfer across PHY, RAN, and ISAC tasks by aligning frequency-domain CSI and delay-domain CIR [2606.25714]. In both cases, the physical equivalence of the two domains provides positive pairs without requiring synthetic augmentations.

Hybrid SSL combines structural reconstruction and discriminative separation. ContraWiMAE augments masked reconstruction with InfoNCE on AWGN-generated positive pairs, using
$$
L(\theta,\phi,\psi)=\alpha L_{\mathrm{rec}}(\theta,\phi)+(1-\alpha)L_{\mathrm{con}}(\theta,\psi),
$$
with $\alpha=0.9$ in the reported experiments [2505.09160]. The intended effect is to improve linear separability and data efficiency beyond what reconstruction alone can achieve.

Predictive CFMs depart from static masked recovery by learning priors over channel evolution. In the predictive channel-estimation framework, a decoder-only time-series transformer forecasts the current slot from the previous slot,
$$
f_\theta(H^{(i-1)}) \to \hat H_{\mathrm{PFM}}^{(i)},
$$
and sparse pilots then calibrate the prediction rather than define the estimate from scratch [2512.15562].

## 3. Architectural patterns and channel representations

CFM architectures closely follow the chosen pretraining objective, but several structural motifs recur: patchified CSI tokens, transformer backbones, modality-specific encoders for paired channel views, and explicit handling of complex-valued data.

Masked-autoencoder CFMs typically convert complex CSI into two real-valued channels and preserve antenna–subcarrier or space–time–frequency topology through positional encoding. WiFo represents MISO-OFDM CSI as a space–time–frequency tensor $H \in \mathbb{C}^{T\times K\times N}$, converts it to a real tensor with separate real and imaginary channels, and applies non-overlapping $3$D patching with patch size $(4,4,4)$ before transformer encoding [2412.08908]. CSI-MAE instead aggregates the MIMO dimension into an antenna-domain axis, builds a two-channel real tensor
$$
X=[\mathrm{Re}(\tilde H),\mathrm{Im}(\tilde H)] \in \mathbb{R}^{N_a\times N_f\times 2},
$$
then uses 2D sine–cosine positional embeddings over the antenna–subcarrier grid [2601.03789]. WiMAE similarly splits each complex patch into real and imaginary parts and explores patch shapes $(1,16)$ and $(4,4)$, reporting better downstream generalization for $(1,16)$ and higher reconstruction SNR for $(4,4)$ [2505.09160].

Contrastive multimodal CFMs use dual encoders instead of encoder–decoder reconstruction. CSI-CLIP and CSI-CLIP++ build one branch for CSI and one for CIR, relying on the Fourier correspondence between the two domains rather than explicit analytical regularization. CSI-CLIP++ states the discrete OFDM relation as
$$
H[k]=\sum_{n=0}^{N-1} h[n] e^{-j2\pi kn/N}, \qquad
h[n]=\frac{1}{N}\sum_{k=0}^{N-1} H[k] e^{+j2\pi kn/N},
$$
and uses paired CSI–CIR realizations to learn a shared embedding space [2606.25714]. CSI-CLIP uses ResNet50 encoders adapted to two-channel real/imaginary inputs, while CSI-CLIP++ reports ViT-B/16 and ViT-L/16 backbones, with downstream adaptation through a lightweight two-layer MLP head on the pretrained CSI encoder [2502.11965][2606.25714].

Task-oriented CFMs incorporate stronger inductive bias into the downstream pipeline. The predictive estimator in “Reducing Pilots in Channel Estimation With Predictive Foundation Models” combines three modules: a decoder-only time-series transformer adapted from TimesFM, a pilot processing network based on a Vision Transformer, and a learned residual fusion block. The predictive backbone performs univariate decomposition over real–imaginary channel sequences, patching with $L_{\mathrm{pat}}=32$, causal attention, and decoder-only forecasting; the pilot encoder uses $L_{\mathrm{ViT}}=10$ layers, enhanced FFN blocks with $3\times3$ depth-wise convolutions, and Adaptive LayerNorm conditioned on $\sigma^2$ [2512.15562]. The final estimate is produced by hidden-state fusion,
$$
U''=[U',R], \qquad \hat H \equiv \hat Z=\mathrm{FusionResBlock}(U''),
$$
which empirically outperforms weighting and attention fusion in that study [2512.15562].

Generalized pretraining under variable input size requires architectural or systems mechanisms beyond standard ViTs. HeterCSI keeps a ViT-style encoder–decoder but adds scale-aware adaptive batching and double masking so that variable $T$, $K$, and $A$ do not collapse training through zero-padding and gradient conflict [2601.18200]. At a broader multimodal scale, WiCo—introduced in the MMICM framework—uses modality-specific encoders for RGB, depth, LiDAR, maps, and RF metadata, a unified transformer backbone with hierarchical tokenization and sparse attention, physics-informed decoders, and parameterized adapters conditioned on band, scenario, and scale [2603.10343].

## 4. Data regimes, pretraining corpora, and transfer

The CFM literature is unusually explicit about pretraining diversity because cross-scenario transfer is treated as a primary capability rather than a side effect. Wireless CFMs are trained on simulated or ray-traced corpora that vary carrier frequency, channel model, antenna layout, OFDM configuration, mobility, and environment.

Several large pretraining sets are repeatedly used. WiMAE and ContraWiMAE pretrain on 1.14M DeepMIMO samples at $3.5$ GHz across $15$ scenarios and evaluate on $14{,}840$ samples from $6$ unseen scenarios [2505.09160]. CSI-CLIP pretrains on more than $700{,}000$ CSI samples from $35$ heterogeneous DeepMIMO scenarios, then evaluates transfer to new tasks and to a Sionna RT urban cellular scenario at $3.5$ GHz [2502.11965]. CSI-CLIP++ also uses more than $0.7$M DeepMIMO samples across $35$ scenarios and evaluates on seven unseen DeepMIMO scenarios plus cross-simulator transfer to Sionna RT [2606.25714]. WiFo pretrains on a heterogeneous QuaDRiGa corpus of $160$K CSI samples from $16$ datasets spanning carrier frequencies from $1.5$ to $5.9$ GHz, variable $T$, $K$, and UPA sizes, and user speed ranges from $0$–$10$ to $120$–$300$ km/h [2412.08908]. CSI-MAE uses approximately $1.45$ million Sionna-generated CSI samples under 3GPP TR 38.901, spanning UMi, UMa, and RMa; carrier frequencies $0.7$, $2.4$, $3.5$, $4.9$, and $5$ GHz; subcarrier spacings $15$, $30$, and $60$ kHz; and UE velocities uniformly sampled in $[0,27.78]$ m/s [2601.03789]. The predictive channel-estimation CFM is trained on $768{,}000$ trajectories times $10$ slots, i.e., $7{,}680{,}000$ samples, covering TDL-A/B/C, speeds $30$–$300$ km/h, subcarrier spacings $\{15,30\}$ kHz, SNRs $5$–$25$ dB, and “2P”/“4P” pilot patterns, with TDL-D LoS reserved for zero-shot tests [2512.15562].

HeterCSI makes the data-handling problem itself part of the method. It pretrains on $40$ QuaDRiGa-generated MISO-OFDM datasets, each with $12{,}000$ samples split $9{,}000/1{,}000/2{,}000$, and evaluates on $12$ distinct zero-shot datasets spanning indoor, RMa, UMa, and UMi conditions across frequencies including $2.1$, $2.4$, $2.6$, $4.8$, $6.2$, $26.5$, $29.0$, $37.5$, $39.0$, and $42.0$ GHz [2601.18200]. Its central empirical insight is that mixed-scale training induces destructive gradient interference, while scenario diversity is constructive once scales are aligned. The reported gradient cosine statistics are $85.96\%$ negative pairs for mixed-scale batches, dropping to $0.01\%$ and $5.56\%$ for same-scale and similar-scale batches across diverse scenarios [2601.18200].

Transfer protocols vary. Some CFMs are explicitly zero-shot: WiFo is designed for instant use on new channel-prediction configurations without fine-tuning [2412.08908], and HeterCSI emphasizes operation without scenario-specific finetuning [2601.18200]. Others mix frozen-backbone and full-finetuning regimes. CSI-MAE reports a lightweight decoder finetuning strategy that freezes the pretrained encoder for extrapolation and feedback, and full-parameter finetuning for positioning [2601.03789]. WiMAE and ContraWiMAE emphasize linear probing and small downstream models, while CSI-CLIP and CSI-CLIP++ use lightweight two-layer MLP heads on a pretrained CSI encoder [2505.09160][2502.11965][2606.25714].

## 5. Downstream tasks and reported empirical behavior

CFMs are evaluated across a broader task spectrum than most earlier wireless self-supervised models: CSI reconstruction, time-domain prediction, frequency-domain prediction, channel estimation, channel identification, beam prediction, feedback, extrapolation, positioning, and system-level BER.

| System | Main tasks | Representative reported outcome |
|---|---|---|
| Predictive CFM estimator [2512.15562] | Channel estimation and BER | NMSE improves by $1$–$5$ dB; “2P” matches LMMSE “4P,” implying up to $50\%$ pilot reduction |
| HeterCSI [2601.18200] | Reconstruction, time prediction, frequency prediction | Versus WiFo, NMSE improves by $7.19$ dB, $4.08$ dB, and $5.27$ dB; training latency reduced by about $53\%$ |
| CSI-CLIP [2502.11965] | Positioning, beam management, channel identification | Positioning mean error distance reduced by $22\%$ on average across $35$ scenarios |
| CSI-CLIP++ [2606.25714] | Channel identification, beam prediction, positioning | Beam Top-1 accuracy improves by up to $19.31$ percentage points |
| CSI-MAE [2601.03789] | Extrapolation, feedback, positioning | Finetuned positioning reaches RMSE $0.718$ m, with $90\%$ errors within $1.37$ m |
| WCFM with NPI suppression [2509.15993] | Channel prediction from degraded CSI | Best NMSE across SINR, predicted subcarrier count $N_p$, and downstream training ratio $R_t$ |

Predictive channel estimation is one of the clearest demonstrations of a task-oriented CFM. The predictive estimator reports that PFM-aided channel estimation improves NMSE by $1$–$5$ dB over LMMSE, CNN, and ViT baselines, that at $30$–$90$ km/h “2P” pilots achieve NMSE comparable to LMMSE with “4P,” and that end-to-end BER with LDPC and $64$-QAM is reduced by at least $3$ dB in the high-SNR regime at $90$ km/h [2512.15562]. Zero-shot tests are central to the argument: the same system reports more than $2$ dB gain for an unseen speed of $150$ km/h, more than $4$ dB gain for an unseen $N_t=16$ setting at SNR $\ge 20$ dB, and competitive performance on unseen TDL-D LoS despite the distribution shift [2512.15562].

Prediction-oriented CFMs show a similar pattern. WiFo reports average NMSE $0.210$ for time-domain prediction and $0.158$ for frequency-domain prediction on its multi-dataset evaluation, outperforming Transformer, LSTM, 3D ResNet, PAD, and LLM4CP variants while remaining zero-shot on unseen configurations [2412.08908]. HeterCSI then uses WiFo as the state-of-the-art zero-shot benchmark and reports average NMSE(dB) gains of $7.19$ dB for CSI reconstruction, $4.08$ dB for time-domain prediction, and $5.27$ dB for frequency-domain prediction, while also surpassing best full-shot competitors on average by $1.33$ dB, $1.58$ dB, and $3.14$ dB on the same three tasks [2601.18200].

Representation CFMs are strongest on discriminative downstream tasks. CSI-CLIP reports an average $22\%$ reduction in positioning error distance over supervised baselines across $35$ scenarios, about $1\%$ average improvement in beam-management accuracy, and gains or parity in channel identification [2502.11965]. CSI-CLIP++ strengthens the beam-prediction result substantially, with Top-1 gains up to $19.31$ percentage points in O1_60 and consistent improvements across seven unseen scenarios, while also improving cross-simulator positioning in Sionna RT by $27.45\%$, $39.27\%$, and $17.52\%$ in the three reported sectors [2606.25714]. In ablations, CSI-CLIP++ reports that contrastive CSI–CIR alignment is stronger than masked reconstruction for PHY and RAN discrimination, while masked reconstruction remains competitive for fine-grained regression in ISAC positioning [2606.25714].

MAE-style CFMs report strong reconstruction-oriented and transfer-oriented behavior. CSI-MAE states that its frozen decoder strategy meets or exceeds supervised baselines on channel extrapolation and channel feedback across all five evaluation scenarios, while full finetuning yields the best reported results, including feedback NMSE values such as $-48.25$ dB for RMa-$2.4$ and $-50.76$ dB for UMi-$5$ [2601.03789]. For positioning on its 3GPP dataset, finetuning achieves RMSE $0.718$ m and places $90\%$ of errors within $1.37$ m, compared with RMSE $1.151$ m and $90\%$ within $2.17$ m for the supervised baseline [2601.03789]. WiMAE and ContraWiMAE emphasize linear separability and data efficiency: for cross-frequency beam selection at codebook size $32$, top-1 linear-probe accuracy with $10\%$ labels reaches $77.0\%$ for ContraWiMAE versus $61.7\%$ for WiMAE, $26.6\%$ for LWM, and $37.6\%$ for RAW input [2505.09160].

Practical deployment results are also beginning to appear. The predictive estimator reports per-slot inference latency of $2.81$ ms and $24.0$M parameters on an RTX 4090 plus i9-14900K, with the claim that latency below $3$ ms is consistent with 3GPP timing budgets [2512.15562]. The NPI-suppression WCFM explicitly addresses the mismatch between perfect-CSI pretraining and degraded-CSI deployment by projecting received pilot signals into channel and orthogonal subspaces, estimating noise-plus-interference, subtracting it, and then completing CSI before foundation-model feature extraction. Its reported advantage is not only better NMSE across SINR and prediction span, but also lower downstream sample complexity: with $R_t=25\%$, it outperforms the other solutions even at $R_t=80\%$ on the channel-prediction task [2509.15993].

## 6. Limitations, open questions, and broader usage of the term

The strongest limitations recur across papers. Many wireless CFMs still rely on simulated or ray-traced data rather than measured data, making hardware impairments, calibration errors, phase noise, IQ imbalance, non-stationary interference, and severe domain shifts only partially represented [2505.09160][2512.15562][2601.03789]. Fixed or weakly variable input geometry also remains a constraint: CSI-CLIP explicitly notes poor generalization to unseen $N_t/N_r/N$ configurations because input reshaping changes spatial and temporal axes [2502.11965]. HeterCSI identifies scale heterogeneity itself as a barrier to generalized pretraining, requiring batch construction and masking strategies that ordinary mixed-scale training does not provide [2601.18200].

Several open questions are now well defined. The predictive-estimation work points to error propagation from previous-slot estimates, pilot-design co-optimization, lightweight variants via pruning or distillation, and the absence of formal bounds relating NMSE to pilot density under learned priors [2512.15562]. The CFM survey emphasizes physics-informed priors, robust tokenization for next-token prediction, low-complexity attention, realistic open datasets, federated and continual learning, explainability, privacy, security, and benchmarking standards as unresolved requirements for scalable deployment [2507.13637]. WiCo extends that list to multimodal alignment, physics-grounded regularization, and parameterized adapters across air–space–ground–sea environments, but also notes the cost of collecting spatially and temporally aligned multimodal corpora and the lack of paradigm-fair evaluation suites [2603.10343].

The term “Channel Foundation Model” is also not confined to wireless channel state information. In scientific computing, D-CHAG uses the phrase for channel-aware foundation models on multi-channel imagery such as hyperspectral data and weather fields, where the main systems challenge is scalable cross-channel tokenization and aggregation under large channel counts [2506.21411]. In multivariate time series, CHARM is presented as a channel-aware foundation embedding model that incorporates channel-level textual descriptions and is permutation-equivariant with respect to channel order, and Partial Channel Dependence introduces dataset-specific channel masks for channel-adaptive time-series foundation models [2505.14543][2410.23222]. This suggests that the phrase “CFM” is becoming polysemous: in wireless communications it denotes foundation models for propagation channels, while in adjacent fields it can denote foundation models that treat channels as a primary modeling axis.

Taken together, the cited literature establishes CFMs as a rapidly diversifying research program rather than a single architecture family. What unifies the field is the attempt to replace scenario-specific, label-intensive wireless intelligence with pretrained channel representations or priors that survive changes in environment, scale, modality, and task. Where the papers differ is in what they regard as the most faithful self-supervised signal: masked recovery of structure, contrastive alignment of physically equivalent views, predictive modeling of channel evolution, or multimodal grounding in environment observations [2507.13637][2606.25714].

Source: https://www.emergentmind.com/topics/channel-foundation-models-cfms