Learnware for CSI Feedback: Scene-specific Small Models Can Do Big

Xiangyi Li; Jiajia Guo; Chao-Kai Wen; Xin Geng; Shi Jin; Zhi-Hua Zhou

cs.ITAug 18, 2026

Learnware for CSI Feedback: Scene-specific Small Models Can Do Big

Authors: Xiangyi Li, Jiajia Guo, Chao-Kai Wen, Xin Geng, Shi Jin, Zhi-Hua Zhou

Organizations: National Mobile Communications Research Laboratory, Southeast University, Nanjing, 210096, China · Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology · Institute of Communications Engineering, National Sun Yat-sen University, Kaohsiung 80424, Taiwan · School of Computer Science and Engineering and the Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, Southeast University, Nanjing 211189, China ·

Abstract

Intelligent channel state information (CSI) feedback is essential for realizing the high capacity and spectral efficiency goals of future 6G systems, yet existing deep learning solutions face a trade-off between model generalization and scenario-specific performance. Large neural networks generalize well but incur high computational and tuning costs, while small models excel in particular environments but require repetitive costly end-to-end training for each base station (BS). To address these challenges, we introduce a model repository-based deployment framework in which a centralized AI data center maintains a catalog of scene-specific CSI models. The repository is enhanced with a Learnware-based framework, where each model is associated with a specification including semantic part (network architecture parameters) and statistical part (codeboo-fingerprint embeddings of training-data distributions). A BS submits only its local statistical specifications to retrieve the most relevant pre-trained model, enhancing data privacy by avoiding raw CSI transmission and drastically reducing retrieval latency and communication overhead. We further develop a data-driven search strategy that matches codebook fingerprints to model performance, achieving over 90% selection accuracy. In simulations, our scheme yields 18.8% and 57.7% performance improvements over the General Model in LOS and NLOS scenarios, respectively while reducing local fine-tuning by up to 1000 samples and 100 epochs. This Learnware-based approach minimizes redundant training, maximizes model reuse, and supports rapid,privacy-enhancing deployment of CSI feedback models.

Figures & tables

Fig. 1: Model repository-based CSI feedback deployment, where the respective functions of the AI data center, BS and UE are marked in different colors.

Panel of Figure 2 — Fig. 2: Sample frequency distribution of codeword fingerprints. (a) Layout of QuaDRiGa-simulated scenarios with UEs randomly distributed in $5\,\rm{m}\times 5\,\rm{m}$ rectangle subregions. (b) Corresponding codeword fingerprint distributions $\mathbf{pdf}$ for LOS and NLOS conditions.

Fig. 3: CSI feedback model solution space. Curves indicate convergence paths from (1) a mismatched model, (2) a general model, and (3) the closest Learnware model.

Fig. 4: Learnware paradigm applied to CSI feedback neural networks.

Paradigm	Resource requirements	Transmission cost	Lifelong learning	Robustness to forgetting	Data privacy
Aggregate datasets	High	High	$\times$	$\times$	$\times$
Learnware Paradigm	Low	Low	✓	✓	✓

TABLE I: Comparison of Paradigms

Panel of Figure 6 — Fig. 5: Comparison between dataset-level distances and model generalization performance.

Panel of Figure 7 — Fig. 6: Online deployment schemes based on model repository, where the key differences are marked in orange.

Method		General Model	General Model with fine-tuning	Model Switch	Original Model Repository	Learnware Model Repository
Construction (Initialize)	AI data center	Train one model using large-scale data		Train several models using large-scale data	Train many models in parallel
Construction (Initialize)	BS	Store a single model		Store several models	Store a single model
Online deployment	AI data center	None	None	None	Model retrieval and return
	BS	None	Fine-tune and transmit encoder NN	Switch model and transmit encoder NN	Upload local CSI samples, fine-tune, transmit encoder NN	Upload specifications, fine-tune, transmit encoder NN
	UE	None	Upload local CSI samples to BS
Merits and Limitations	Accuracy	Low	High	Middle	High	High

TABLE II: Comparison of general model, model switch, and model repository-based schemes across key deployment considerations.

CSI_Number_Domain	spatial domain
Feedback_Mode	implicit feedback
Structure	two-sided model
Input_Dimension	$32\times 2$
Output_Dimension	$32\times 2$
Compression_Ratio	1/16
Quantization_Bits	4

TABLE III: Semantic specification labels used in the experiment.

Antenna setting	32 ULA antennas at BS with spacing
	of 0.5 wavelength, centered at (0,0) and
	arranged in a north-south direction (Y-axis)
	Single antenna at UE
Center frequency	2.6 GHz
Bandwidth	20 MHz
Scene configuration	3GPP-38.901-UMi-LOS

TABLE IV: Basic parameter settings used in the simulation.

Layer	Type	Input	Output	Details
CNN-based Autoencoder [ 8 ]
Input	-	(32, 2)	(32, 2)	Real and imaginary parts
Encoder	Conv1D	(32, 2)	(32, 2)	Kernel: $7\times 1$ , BN, ReLU
	Reshape	(32, 2)	(64,)	Flatten
	Dense+Sigmoid	(64,)	(4,)	Compression ratio $CR=1/16$
	Quantization	(4,)	(16,)	$Q=4$ bits, feedback_bits=16

TABLE V: Autoencoder Structures for CNN and Transformer-based CSI Feedback

Sample Range ( $\mathrm{m}^{2}$ )	LOS								NLOS
	Scene-specific		Learnw. Model Repository	General Model (various complexity) 1					Scene-specific		Learnw. Model Repository	General Model (various complexity)
	CNN 2	TF-1	Learnw. Model Repository	CNN	TF-1	TF-2	TF-3	TF-4 3	CNN	TF-1	Learnw. Model Repository	CNN	TF-1	TF-2	TF-3	TF-4
$5\times 5$	0.915	0.915	0.889	0.332	0.628	0.725	0.755	0.777	0.350	0.357	0.317	0.181	0.208	0.207	0.214	0.230
$5\times 10$	0.881	0.883	0.855	0.340	0.701	0.719	0.746	0.715	0.347	0.343	0.321	0.225	0.213	0.220	0.245	0.263
$5\times 20$	0.879	0.886	0.808	0.297	0.703	0.710	0.713	0.715	0.330	0.347	0.314	0.162	0.205	0.205	0.216	0.221
$5\times 30$	0.868	0.881	0.721	0.292	0.651	0.684	0.691	0.660	0.324	0.339	0.303	0.187	0.213	0.207	0.214	0.213

TABLE VI: Average SGCS performance of scene-specific small models and large general model with various complexity. CR=1/16 and feedback quantization bits

Q

=4.

Configurations		LOS scenes					NLOS scenes
CR	Quantization bits ( $Q$ )	General Model	Model Switch	Learnware Model Repository	Original Model Repository	Upper Bound	General Model	Model Switch	Learnware Model Repository	Original Model Repository	Upper Bound
1/16	2	0.681	0.371	0.832	0.841	0.878	0.219	0.197	0.296	0.309	0.336
	3	0.715	0.453	0.850	0.859	0.881	0.230	0.222	0.313	0.328	0.356
	4	0.777	0.477	0.889	0.891	0.915	0.230	0.230	0.317	0.333	0.350
	5	0.781	0.482	0.899	0.907	0.921	0.232	0.233	0.320	0.335	0.356
1/32	4	0.348	0.249	0.794	0.816	0.852	0.134	0.144	0.215	0.239	0.251

TABLE VII: Average SGCS performance of methods with various CR and feedback quantization bits. Sampling range:

5\times 5\,m^{2}

Panel of Figure 14 — Fig. 7: SGCS performance without fine-tuning and codebook fingerprint cosine similarity. The LOS and NLOS target testing scenes are presented in (a) and (b). In each subfigure, the LOS and NLOS Learnwares are displayed in blue and orange, respectively. The Learnware selected by the codebook fingerprint (i.e., the one with the maximum cosine similarity) is marked in green. The CSI sampling area for the test set is set to $5\,\text{m}\times 5\,\text{m}$ , the same as the Learnware training set. CR=1/16 and feedback quantization bits $Q$ =4.

Panel of Figure 15 — Fig. 8: Performance improvements on the two cases from Fig. 7 post-fine-tuning. The LOS and NLOS target test scenarios are shown in (a) and (b), respectively. In each subfigure, the upper and lower sections represent the number of samples used for fine-tuning and the number of fine-tuning epochs, respectively. Tested CSI sampling area: $5\,\mathrm{m}\times 5\,\mathrm{m}$ . The corresponding fine-tuning results are summarized in Tab. VIII . (c) presents the average SGCS performance over multiple sampling areas, where each bar represents the average across 10 test scenarios. Fine-tuning overhead: 500 samples and 100 epochs.

Methods	LOS Test Scene		NLOS Test Scene
Methods	Without fine-tuning	With fine-tuning	Without fine-tuning	With fine-tuning
General Model	0.781	0.936	0.293	0.408
Model Switch	0.506	None	0.413	None
Original Model Repository	0.943	0.951	0.492	0.499
Learnware Model Repository	0.928	0.946	0.462	0.472
Upper Bound	0.951		0.507

TABLE VIII: SGCS performance before and after fine-tuning (500 samples, 100 epochs) of individual representative case, corresponding to Fig. 8 .

Level	k	L	Input Candidate	LSH Score	Brute-force	Relative	Per-Level	Cumulative	Speedup
			Repository Size		Upper Bound	Accuracy	Time (ms)	Time (ms)	Factor
L1	4	25	$1.0\times 10^{8}$	0.85	0.99	85.9%	0.25	0.25	20,000 $\times$
L2	5	18	$6.25\times 10^{6}$	0.91	0.99	91.9%	0.35	0.60	8,333 $\times$
L3	6	12	$1.95\times 10^{5}$	0.96	0.99	97.0%	0.30	0.90	5,556 $\times$
L4	7	6	$3.05\times 10^{3}$	0.98	0.99	99.0%	0.20	1.10	4,545 $\times$
Final	-	-	48	0.99	0.99	100%	0.05	1.15	4,348 $\times$

TABLE IX: Performance of Multi-Level LSH with Adjustable Search Depth

Deployment considerations	Communication overhead		Time complexity		Communication overhead [KB] 1		Time complexity [s] 2
Deployment considerations	Data center-BS	BS-UE	Fine-tuning	Model retrieval	Data center-BS	BS-UE	Fine-tuning	Model retrieval
General Model	None	None	None	None	0	0	0	0
General Model + fine-tuning	None	500 samples for fine-tuning; return 1 encoder	100 epochs, 500 samples	None	0	240 + 3.37 $\times 10^{4}$	373.5	0
Model Switch	None	500 samples for evaluation; return 1 encoder	None	Evaluate 4 models	0	240 + 5.8	0	0.71
Original model repository	500 samples for evaluation; return 1 model	500 samples for fine-tuning; return 1 encoder	100 epochs, 500 samples	Evaluate 144 models	240 + 31.7	240 + 5.8	61.6	25.7
Learnware model repository	Specification; return 1 model	500 samples for fine-tuning; return 1 encoder	100 epochs, 500 samples	Retrieve from 144 models	2.6 + 31.7	240 + 5.8	61.6	1.48e-5

TABLE X: Online deployment cost comparison for the general model, model switch, and proposed model repository-based schemes.

Panel of Figure 19 — Fig. 9: We simulated 1,000 LOS and 1,000 NLOS scenarios in each sampling area, with each scenario comprising 500 samples. (a) Performance ranking of the learnware selected by $\mathbf{pdf}$ within the model repository (comprising 144 models), i.e., the probability of being ranked in the top x positions; (b) box-plot of the performance gap between the $\mathbf{pdf}$ -selected learnware and the top-performing model (i.e., Original model repository); (c) Selection accuracy: probability that the performance gap of the $\mathbf{pdf}$ -selected learnware meets the expected threshold.

Panel of Figure 20 — Fig. 10: (a) Average SGCS performance of Learnware Model repository, with each group of 10 cases. CR=1/16 and feedback quantization bits $Q$ =4; (b) Test scenarios: 10 cases for LOS/NLOS, $5\,m\times 5\,m$ sampling range, For each sample size, the estimated $\mathbf{pdf}$ is compared against the benchmark (full 10,000 samples), Performance loss is defined as SGCS ${}_{\text{benchmark}}-$ SGCS ${}_{\text{other}}$ . PDF similarity measures the cosine similarity between the estimated $\mathbf{pdf}$ and the benchmark $\mathbf{pdf}$ . (c) Multi-carrier OFDM scenario without fine-tuning. The configuration and legend follow the same format as Fig. 7 .

Metric	Original Formula	Original Range	Normalized Similarity	Complexity 1
Cosine Similarity	$\displaystyle s_{\cos}(p,q)=\frac{\sum_{i}p_{i}q_{i}}{\\|p\\|\cdot\\|q\\|}$	$[0,1]$	$S_{\cos}=s_{\cos}$	Low
Bhattacharyya Coefficient [ 47 ]	$\displaystyle s_{\text{Bhatt}}(p,q)=\sum_{i}\sqrt{p_{i}q_{i}}$	$[0,1]$	$S_{\text{Bhatt}}=s_{\text{Bhatt}}$	Medium
Hellinger Distance [ 48 ]	$\displaystyle d_{\text{Hell}}(p,q)=\frac{1}{\sqrt{2}}\left\\|\sqrt{p}-\sqrt{q}\right\\|_{2}$	$[0,1]$	$S_{\text{Hell}}=1-d_{\text{Hell}}$	Medium
Total Variation	$\displaystyle d_{\text{TV}}(p,q)=\frac{1}{2}\sum_{i}\|p_{i}-q_{i}\|$	$[0,1]$	$S_{\text{TV}}=1-d_{\text{TV}}$	Low
Jensen-Shannon Divergence [ 49 ]	$\displaystyle d_{\text{JS}}(p,q)=\frac{1}{2}D_{\text{KL}}(p\\|m)+\frac{1}{2}D_{\text{KL}}(q\\|m)$	$[0,1]$	$S_{\text{JS}}=1-d_{\text{JS}}$	High
KL Divergence [ 50 ]	$\displaystyle d_{\text{KL}}(p,q)=\sum_{i}p_{i}\log\frac{p_{i}}{q_{i}}$	$[0,+\infty)$	$S_{\text{KL}}=\exp(-d_{\text{KL}})$	High

TABLE XI: Similarity and Distance Metrics for Probability Distributions

Similarity Metrics	LOS						NLOS
Sampling Range ( $\mathrm{m}^{2}$ )	$5\times 5$	$5\times 10$	$5\times 20$	$5\times 30$	$5\times 40$	$5\times 50$	$5\times 5$	$5\times 10$	$5\times 20$	$5\times 30$	$5\times 40$	$5\times 50$
Cosine ( $s_{\cos}$ )	0.886	0.855	0.839	0.721	0.659	0.599	0.324	0.318	0.319	0.303	0.285	0.233
Bhattacharyya [ 47 ] ( $s_{\text{Bhatt}}$ )	0.883	0.857	0.855	0.746	0.666	0.612	0.327	0.309	0.312	0.287	0.266	0.209
Hellinger [ 48 ] ( $S_{\text{Hell}}$ )	0.883	0.857	0.855	0.746	0.666	0.612	0.327	0.309	0.312	0.287	0.266	0.209
Total Variation ( $S_{\text{TV}}$ )	0.884	0.856	0.854	0.721	0.669	0.625	0.324	0.307	0.302	0.300	0.283	0.226
KL [ 50 ] ( $S_{\text{KL}}$ )	0.881	0.847	0.855	0.786	0.719	0.657	0.327	0.296	0.318	0.286	0.270	0.209

TABLE XII: Average SGCS performance of Learnware Model Repository with various distributional similarity metrics. CR=1/16 and feedback quantization bits

Q

=4.

Wireless Communications

Wireless communications studies how information is sent over radio links and how networks manage signals, spectrum, and resources reliably.

See all 400 papers

Jun 12, 2026

Towards CSI-Native Foundation Models: A Channel-Adaptive Roadmap for 6G

Wireless foundation models offer a path toward reusable channel state information (CSI) intelligence for sixth-generation (6G) systems. However, existing generic-backbone adaptation and CSI pretraining methods often treat CSI as task tensors rather than propagation-conditioned channel responses, thereby failing to capture the intrinsic time-frequency-spatial geometry of wireless environments. This paper presents a channel-adaptive roadmap toward CSI-native foundation models, proposing a unified framework that aligns pretraining, positional modeling, and attention control with three channel requirements: scale-aware heterogeneous exposure, physical time-frequency-antenna coordinates, and correlation-bounded token interaction. Extensive experiments demonstrate the superiority of the proposed framework across three dimensions: zero-shot generalization, reducing NMSE by more than 4 dB across spatial-temporal-frequency tasks; scale extrapolation, yielding up to a 5.4 dB gain under 8 times unseen antenna scaling; and inference efficiency, accelerating mobility-aware processing by up to 18.8%. A system-level evaluation with Sionna SYS further shows that the proposed framework uses only 7.01% of dense-pilot overhead, reaches -18.64 dB average NMSE, and improves average net spectral efficiency by 36.6% over dense LMMSE and 15.5% over WiFo, indicating that CSI-native representation learning can support pilot-efficient radio access.

Chenyu Zhang, Xinchen Lyu, Chenshan Ren +2

National Engineering Research Center for Mobile Network Technologies, Beijing University of Posts and Telecommunications, Beijing 100876, China · Key Laboratory of Ethnic Language Intelligent Analysis and Security Governance of MOE, Minzu University of China, Beijing 100081, China · China Telecom Corporation Limited Gansu Branch, Gansu 730000, China

Wireless Foundation Models Representation Learning 6G Networks Wireless Communications Channel Estimation

Ask PDF

Jun 26, 2026

Contrastive Predictive Coding with Compression for Enhanced Channel State Feedback in Wireless Networks

Accurate and timely channel state information (CSI) is essential for next-generation wireless systems, yet existing works treat CSI compression and CSI prediction as separate problems, both in academia and in current 3GPP studies. Consequently, channel aging remains insufficiently addressed within standardized CSI feedback pipelines. In this article, we propose a unified compression-prediction framework that integrates Contrastive Predictive Coding (CPC) directly into the 3GPP-compliant CSI compression architecture. Instead of predicting high-dimensional CSI matrices, our approach forecasts future latent representations and jointly optimizes reconstruction fidelity and temporal predictive coherence via a combined 1-SGCS and InfoNCE objective. This design enables temporal representation learning without increasing feedback overhead. We present two variants: CPC-before-Compression, which performs autoregressive modeling on encoded features prior to quantization, and CPC-after-Compression, which shifts temporal modeling to the base-station to reduce the complexity of users' devices. Evaluations on 3GPP-compliant datasets from Nokia, Oppo, and CATT show that CPC-before-Compression achieves over 90% reconstruction accuracy with 32x lower decoder GFLOPs than the 3GPP baseline, while CPC-after-Compression preserves an identical encoder footprint and the same 64-bit feedback overhead. By unifying compression and prediction within a standardized pipeline, the proposed framework provides an age-aware, computationally efficient CSI feedback solution. The source code is publicly available at: https://github.com/AhmedRadwan02/cpc-3gpp

Ahmed Y. Radwan, Fahad Syed Muhammad, Matthew Baker +1

Department of Electrical Engineering and Computer Science, York University, Toronto, ON M3J 1P3, Canada · Nokia Networks, 12 rue Jean Bart, Paris-Saclay, Massy, 91300, France · Nokia UK, Broers Building, 21 JJ Thomson Avenue, Cambridge, CB3 0FA, United Kingdom

Channel State Prediction Wireless Communications Predictive Representation Learning

Ask PDF

Oct 27, 2026

RAG-CSI: Global Local Retrieval via Joint State and Dynamics Matching for CSI Prediction

Channel state information (CSI) prediction forecasts future channel states from historical observations, supporting transceiver design and resource allocation with reduced measurement overhead. However, existing deep learning predictors mainly encode channel knowledge in model parameters, which can limit their ability to represent diverse patterns of channel dynamics across scenes. In this paper, we propose RAG-CSI, a retrieval-enhanced Transformer framework that augments prediction with relevant historical channel knowledge. Specifically, we construct a global--local CSI knowledge store containing representative observations across scenes and historical observations from the current serving scene. We further develop joint state--dynamics retrieval (JSDR), which jointly characterizes channel states, phase dynamics, and amplitude dynamics to retrieve relevant CSI sequences. A Transformer predictor then uses cross-attention to fuse the retrieved knowledge with the current CSI history and predict future CSI with limited additional model parameters. Experiments on a multi-scene CSI dataset show overall improvements exceeding 10% across all five evaluation metrics over the best baseline for each metric, including a 15.42% reduction in normalized mean squared error.

Xinyue Jiang, Tong Li, Shiceng Zhang +2

College of Computer Science and Electronic Engineering, Hunan University, Changsha, China. · Xidian University, Xian, China. · Peng Cheng Laboratory, Shenzhen 518066, China.

Channel State Prediction Wireless Communications Retrieval-Augmented Time Series Forecasting Retrieval-Augmented Generation

Ask PDF

Abstract

Wireless Communications

Explore similar work

Towards CSI-Native Foundation Models: A Channel-Adaptive Roadmap for 6G

Contrastive Predictive Coding with Compression for Enhanced Channel State Feedback in Wireless Networks

RAG-CSI: Global Local Retrieval via Joint State and Dynamics Matching for CSI Prediction