Beyond Channel Data: Configuration-Enriched Radio Foundation Models

Shrihari Vasudevan; H. G. Ranjani; Raaghul Ranganathan; Satwik Sahu; Thirumaran Ekambaram

cs.AIOct 23, 2026

Beyond Channel Data: Configuration-Enriched Radio Foundation Models

Authors: Shrihari Vasudevan, H. G. Ranjani, Raaghul Ranganathan, Satwik Sahu, Thirumaran Ekambaram

Organizations: Ericsson R&D, India · BITS Pilani, India

Abstract

This paper proposes incorporating channel configuration (CFG) information into Radio Foundation Model (RFM) pre-training and demonstrates that the resulting CFG-enriched embeddings serve as effective inductive priors for downstream tasks under label scarcity. We compare the proposed embeddings with raw channel tokens and multiple baseline models, under three different data regimes, for multiple downstream tasks --- Line-of-Sight/Non-Line-of-Sight (LoS/NLoS) classification, timing advance (TA) estimation, and beam selection. Evaluation is done in both in-distribution and out-of-distribution configuration settings. Experimental outcomes show that the proposed embeddings consistently outperform and/or are competitive to the baselines, across experiments in both seen and unseen configuration scenarios.

Figures & tables

Fig. 1: Baseline channel RFM. The Encoder comprises multiple Transformer layers with MHA, Layer normalization and FFN (Feed-forward Network). The Decoder consists of one linear layer. NMSE - Normalized Mean Squared Error.

Fig. 2: Architecture of proposed work. In comparison to baseline, this paper introduces CFG tokens along with $H$ and $CLS$ tokens.

Parameter	Pretraining
Carrier frequency	3.5 GHz
Subcarriers (OFDM)	512
Selected subcarriers a	48, 96, 144, 192, 288, 384, or 480
BS antenna shape a	[4,2], [4,4], [6,4], [8,4], or [8,8]
UE antenna shape a	[2,1] or [2,2]
BS antenna spacing	0.25 $\lambda$

TABLE I: Pretraining scenario configurations resulting in

\sim

126K samples of

H

from 30 unique CFG variants.

Architecture		Pretraining
Embedding dim, $d$	64	Optimizer	AdamW
Layers	12	Learning rate	1e-4
Attention heads	12	Batch size	16
Activation	GELU	Weight decay	1e-5
Patch size, $ps$	16	Warm-up steps	50
Dropout	0.1	Clip multiplier	2.0

TABLE II: Architecture and pretraining Hyperparameters

Parameter	Task seen	Task unseen
Samples	$\sim$ 32K	$\sim$ 40K
Carrier frequency	3.5 GHz	3.5 GHz
Subcarriers (OFDM) a	512	512, 1024
Selected subcarriers a	48, 144	480, 512, 256
BS antenna shape a	[4,2] and [4,4]	[8,6], [4,8] and [4,6]
UE antenna shape a	[2,1]	[2,1] and [2,2]

TABLE III: Downstream task scenario configurations

			LoS/NLoS (F1 $\uparrow$ )			TA (NMSE $\times 10^{-6}$ $\downarrow$ )			Beam-8 (F1 $\uparrow$ )
Grp	Embedding	Dim	Few-shot	Moderate	Data-Rich	Few-shot	Moderate	Data-Rich	Few-shot	Moderate	Data-Rich
B1	$T_{H}$	27,648	$0.564{\pm}0.136$	$0.702{\pm}0.065$	$0.785{\pm}0.056$	$9.981{\pm}0.882$	$8.177{\pm}0.524$	$6.861{\pm}0.382$	$0.210{\pm}0.048$	$0.327{\pm}0.008$	$0.426{\pm}0.036$
B2	$T_{H,\text{CFG}}$	27,659	$0.564{\pm}0.126$	$0.654{\pm}0.068$	$0.826{\pm}0.029$	$10.526{\pm}1.365$	$8.156{\pm}0.569$	$6.913{\pm}0.268$	$0.142{\pm}0.017$	$0.283{\pm}0.037$	$0.426{\pm}0.031$
B3	$E_{\text{CLS}}$	64	$0.970{\pm}0.010$	$0.980{\pm}0.003$	$\mathbf{0.987{\pm}0.001}$	$5.833{\pm}2.840$	$0.880{\pm}0.051$	$0.696{\pm}0.101$	$0.170{\pm}0.024$	$0.379{\pm}0.037$	$\mathbf{0.558{\pm}0.041}$
B3	$E_{H_{\text{mean}}}$	64	$0.970{\pm}0.011$	$0.979{\pm}0.002$	$0.986{\pm}0.001$	$3.172{\pm}1.400$	$0.952{\pm}0.185$	$0.817{\pm}0.324$	$0.162{\pm}0.022$	$0.330{\pm}0.037$	$0.504{\pm}0.033$
B3	$E_{H_{\text{max}}}$	64	$0.967{\pm}0.023$	$\mathbf{0.980{\pm}0.002}$	$0.986{\pm}0.001$	$3.452{\pm}1.586$	$1.086{\pm}0.057$	$\mathbf{0.657{\pm}0.183}$	$0.169{\pm}0.011$	$0.272{\pm}0.038$	$0.489{\pm}0.046$

TABLE IV: Seen - LoS/NLoS (F1

\uparrow

), TA (NMSE

\times 10^{-6}

\downarrow

), Beam-8 (F1

\uparrow

			LoS/NLoS (F1 $\uparrow$ )			TA (NMSE $\times 10^{-6}$ $\downarrow$ )			Beam-8 (F1 $\uparrow$ )
Grp	Embedding	Dim	Few-shot	Moderate	Data-Rich	Few-shot	Moderate	Data-Rich	Few-shot	Moderate	Data-Rich
B1	$T_{H}$	27,648	$0.348{\pm}0.046$	$0.405{\pm}0.056$	$0.453{\pm}0.044$	$15.536{\pm}3.026$	$17.923{\pm}2.111$	$15.714{\pm}2.349$	$0.166{\pm}0.064$	$\mathbf{0.244{\pm}0.009}$	$0.076{\pm}0.013$
B2	$T_{H,\text{CFG}}$	27,659	$0.348{\pm}0.039$	$0.400{\pm}0.026$	$0.458{\pm}0.046$	$15.803{\pm}3.251$	$17.312{\pm}1.267$	$14.540{\pm}1.272$	$0.118{\pm}0.020$	$0.147{\pm}0.015$	$0.105{\pm}0.028$
B3	$E_{\text{CLS}}$	64	$0.887{\pm}0.051$	$0.849{\pm}0.031$	$0.882{\pm}0.033$	$2.496{\pm}2.619$	$1.811{\pm}0.095$	$2.620{\pm}0.279$	$0.182{\pm}0.013$	$0.201{\pm}0.004$	$0.215{\pm}0.004$
B3	$E_{H,\text{mean}}$	64	$0.918{\pm}0.068$	$0.860{\pm}0.057$	$0.756{\pm}0.110$	$\mathbf{2.456{\pm}1.754}$	$2.428{\pm}0.371$	$2.230{\pm}0.176$	$0.190{\pm}0.019$	$0.199{\pm}0.020$	$0.202{\pm}0.009$
B3	$E_{H,\text{max}}$	64	$0.911{\pm}0.068$	$0.836{\pm}0.072$	$0.762{\pm}0.103$	$3.161{\pm}3.311$	$4.093{\pm}1.376$	$2.052{\pm}0.222$	$0.185{\pm}0.023$	$0.175{\pm}0.026$	$0.220{\pm}0.008$

TABLE V: Unseen - LoS/NLoS (F1

\uparrow

), TA (NMSE

\times 10^{-6}

\downarrow

), Beam-8 (F1

\uparrow

			Beam-16 (F1 $\uparrow$ )			Beam-32 (F1 $\uparrow$ )
Grp	Embedding	Dim	Low	Med	Rich	Low	Med	Rich
B1	$T_{H}$	27,648	$0.076{\pm}0.017$	$0.095{\pm}0.024$	$0.256{\pm}0.091$	$0.019{\pm}0.006$	$0.052{\pm}0.022$	$0.212{\pm}0.053$
B2	$T_{H,\text{CFG}}$	27,659	$0.039{\pm}0.015$	$0.068{\pm}0.030$	$0.331{\pm}0.049$	$0.027{\pm}0.007$	$0.038{\pm}0.009$	$0.187{\pm}0.075$
B3	$E_{\text{CLS}}$	64	$0.067{\pm}0.029$	$0.176{\pm}0.027$	$0.446{\pm}0.047$	$0.018{\pm}0.014$	$0.182{\pm}0.041$	$\mathbf{0.356{\pm}0.045}$
B3	$E_{H_{\text{mean}}}$	64	$0.070{\pm}0.041$	$0.166{\pm}0.023$	$0.390{\pm}0.050$	$0.042{\pm}0.044$	$0.163{\pm}0.011$	$0.266{\pm}0.050$
B3	$E_{H_{\text{max}}}$	64	$0.072{\pm}0.048$	$0.158{\pm}0.022$	$0.360{\pm}0.063$	$0.065{\pm}0.027$	$0.178{\pm}0.015$	$0.237{\pm}0.025$

TABLE VI: Seen - Beam-16 and Beam-32 (F1

\uparrow

			Beam-16 (F1 $\uparrow$ )			Beam-32 (F1 $\uparrow$ )
Grp	Embedding	Dim	Low	Med	Rich	Low	Med	Rich
B1	$T_{H}$	27,648	$0.049{\pm}0.013$	$0.077{\pm}0.056$	$0.117{\pm}0.054$	$0.010{\pm}0.005$	$0.019{\pm}0.003$	$0.018{\pm}0.003$
B2	$T_{H,\text{CFG}}$	27,659	$0.032{\pm}0.009$	$0.034{\pm}0.011$	$0.102{\pm}0.040$	$0.012{\pm}0.004$	$0.041{\pm}0.041$	$0.028{\pm}0.013$
B3	$E_{\text{CLS}}$	64	$0.038{\pm}0.014$	$0.140{\pm}0.027$	$\mathbf{0.130{\pm}0.013}$	$0.015{\pm}0.001$	$0.016{\pm}0.001$	$0.026{\pm}0.005$
B3	$E_{H_{\text{mean}}}$	64	$0.049{\pm}0.036$	$0.130{\pm}0.039$	$0.110{\pm}0.012$	$0.014{\pm}0.001$	$0.017{\pm}0.011$	$0.010{\pm}0.004$
B3	$E_{H_{\text{max}}}$	64	$\mathbf{0.103{\pm}0.021}$	$\mathbf{0.155{\pm}0.023}$	$0.129{\pm}0.011$	$\mathbf{0.036{\pm}0.011}$	$\mathbf{0.053{\pm}0.003}$	$\mathbf{0.057{\pm}0.011}$

TABLE VII: Unseen - Beam-16 and Beam-32 (F1

\uparrow

Panel of Figure 10 — Fig. 3: A t-SNE visualization of 64-dimensional (a) $E_{\text{CLS}}$ embeddings from baseline 3 (B3) model (b) $E_{\text{CLS}}^{\dagger}$ embeddings from proposed approach from the seen configurations.

Wireless Foundation Models

Wireless foundation models learn reusable representations of radio channels, signals, network traces, and telemetry for transfer across wireless tasks and settings.

See all 33 papers

CFM-Bench: A Unified Multi-Domain, Multi-Task Benchmark for Channel Foundation Models
Jul 16, 2026Yuan Gao, Wenjun Yu, Jun Jiang +3Wireless Foundation Models Benchmark Design
School of Communication and Information Engineering, Shanghai University, Shanghai, China · Xi’an Jiaotong-Liverpool University, Suzhou, China
OpenRFM: Dissecting Relational In-Context Learning
Jun 3, 2026Zhikai Chen, Junyu Yin, Jialiang Gu +5Relational Foundation Models Synthetic Data Pretraining
Michigan State University · George Mason University · Georgia Institute of Technology +1
LatentWave: JEPA Pretraining for Wireless Foundation Models
Jun 4, 2026Ahmed Mohamed, Ahmed Aboulfotouh, Hatem Abou-ZeidWireless Foundation Models Self-Supervised Pre-Training
Department of Electrical and Software Engineering, University of Calgary, Canada

Jul 16, 2026J/K move · Enter open · S save

CFM-Bench: A Unified Multi-Domain, Multi-Task Benchmark for Channel Foundation Models

Yuan Gao, Wenjun Yu, Jun Jiang, Yunfan Li+2

School of Communication and Information Engineering, Shanghai University, Shanghai, China · Xi’an Jiaotong-Liverpool University, Suzhou, China

Wireless Foundation Models Benchmark Design Wireless Communications Radio Propagation Modeling

Open paper AskPDF arXiv

Channel foundation models (CFMs) are developing rapidly, with recent studies reporting benefits from pretraining across downstream wireless tasks. Yet CFMs are commonly evaluated in model-specific pipelines with different data, radio configurations, partitions, adaptation procedures, task definitions, and metrics. Reported comparisons therefore tend to show that pretraining improves over supervised training from scratch within one pipeline, but neither rank CFMs nor compare them fairly with task-specific models. We release CFM-Bench, a unified multi-domain, multi-task benchmark designed to address this gap. It curates six channel configurations spanning 3GPP statistical simulation, two independent ray-tracing pipelines, industrial and aerial measurements, and synchronized vehicular multimodal simulation. Official partitions isolate complete trajectories, measurement sessions, vehicle links, simulation realizations, or buffered spatial regions. CFM-Bench does not prescribe an external pretraining corpus or strategy; no benchmark split may be used for foundation-model pretraining, and the official training split is reserved exclusively for downstream fine-tuning. The benchmark additionally requires disclosure of all data used during model development and prohibits training-stage use of official test units. Six task groups are organized along three CFM application dimensions: physical-layer (PHY) channel intelligence, radio-access-network (RAN) decision intelligence, and integrated sensing and communication (ISAC). They cover CSI feedback, frequency and temporal channel extrapolation, propagation-state classification, current- and future-beam prediction, and single-frame and temporal localization. CFM-Bench provides a common substrate for comparing the transferability of channel representations across models, domains, and tasks.

Abstract

Wireless Foundation Models

Explore similar work

CFM-Bench: A Unified Multi-Domain, Multi-Task Benchmark for Channel Foundation Models

OpenRFM: Dissecting Relational In-Context Learning

LatentWave: JEPA Pretraining for Wireless Foundation Models