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

Source-linked AI summary

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

arXiv:2608.17760v1Aug 18, 2026cs.ITcs.AIeess.SP

TL;DR

AI-based CSI feedback must balance broad model generalization with scenario-specific performance and deployment cost. The paper introduces a Learnware repository that retrieves scene-specific models using compact specifications, achieving approximately 95% LOS and 90% NLOS selection accuracy.

Problem
AI-based CSI feedback deployment remains challenging because unified models may perform suboptimally in specialized environments.
Method
A centralized repository uses semantic and codebook-fingerprint statistical specifications to retrieve suitable pre-trained CSI models without exposing raw training data.
Results
Approximately 95% accuracy was achieved for LOS and 90% for NLOS model selection with ϵ = 0.1.
Takeaways & Limitations
The framework supports model reuse across diverse deployment scenarios while reducing training overhead, raw CSI transmission, and exhaustive model evaluation.
Takeaways & Limitations
The observed relationship between codebook-fingerprint distributions and scenarios can be an experimental-design artifact that changes with sampling areas, antenna arrays, and obstacle layouts.

Abstract

from arXiv · show

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.

I. INTRODUCTION · II. SYSTEM MODEL · A. Massive MIMO FDD System

The paper motivates repository-based CSI feedback for massive MIMO FDD systems and introduces Learnware specifications to retrieve reusable, scene-specific models while avoiding raw CSI transmission. The system model establishes CSI feedback, compression needs, and massive-MIMO precoding foundations.

I. INTRODUCTION: Massive MIMO improves capacity and spectral efficiency, supporting 6G requirements for higher throughput, lower latency, reliability, connectivity, and spectrum utilization.These gains arise from spatial multiplexing in massive multiple-input multiple-output arrays.
I. INTRODUCTION: In FDD MIMO, the UE estimates downlink CSI and feeds it back to the BS, making efficient compression and reconstruction essential because full CSI transmission has substantial overhead.The feedback supports BS precoding while limiting transmission overhead.
I. INTRODUCTION: Unified CSI models offer broad applicability but can perform suboptimally in specialized environments while requiring high computational cost, long training times, and extensive expert tuning.Scene-specific small models can instead achieve near-optimal performance in target scenarios and support dynamic switching or rapid fine-tuning.
I. INTRODUCTION: The proposed repository maintains standardized metadata for scene-specific AI models and connects centralized storage with BS delivery requests for systematic model reuse across heterogeneous scenarios.This repository-centric architecture replaces isolated retrain-and-discard workflows with lifecycle-aware model orchestration.
I. INTRODUCTION: Each Learnware specification combines semantic network-architecture information with a statistical codebook-fingerprint representation that characterizes model capability without exposing training data.The codebook fingerprint captures angular-domain channel structure through DFT beamspace projection and avoids raw CSI exposure.
I. INTRODUCTION: Over 90% selection accuracy is achieved through pdf similarity matching, while multi-level LSH supports billion-model repositories with 0.25–1.15 ms retrieval latency.The retrieval framework provides a controllable accuracy-latency tradeoff for CSI feedback timing constraints.
A. Massive MIMO FDD System: In the downlink massive MIMO system, a BS with Nt ULA antennas serves a UE with Nr antennas, and the received signal depends on the channel matrix, precoding vector, transmitted symbol, and AWGN.The channel matrix is H ∈CNr×Nt, the precoder satisfies ∥v∥2 = 1, and the noise follows n ∼CN(0, σ2INr).
A. Massive MIMO FDD System: For single-stream transmission, the BS selects the first right-singular vector of the channel matrix as the precoding vector because it corresponds to the largest singular value.The SVD uses unitary matrices U and V and diagonal singular-value matrix Σ.

B. DL-Based CSI Feedback · C. Codeword Fingerprint

The CSI feedback method uses an autoencoder trained with a cosine-distance objective aligned to SGCS, while codeword fingerprints summarize dataset-specific angular characteristics through a DFT codebook. Aggregated dominant-beam directions distinguish concentrated LOS energy from broader NLOS multipath dispersion.

B. DL-Based CSI Feedback: The DL-based CSI feedback framework preprocesses the precoding vector v by separating real and imaginary parts into a (2, Nt) matrix normalized to [0, 1].This normalized matrix is used as the network input.
B. DL-Based CSI Feedback: An encoder at the UE compresses v, and a decoder at the BS reconstructs it after uniform quantization with level Q.The encoder and decoder have parameters Ωenc and Ωdec, respectively.
B. DL-Based CSI Feedback: The feedback bitstream has compression ratio CR = M/(2Q × Nt), linking its length to the quantization level and antenna dimension.The bitstream is denoted sQ and has length Mlength.
B. DL-Based CSI Feedback: Training minimizes cosine distance between original and reconstructed precoding vectors, directly aligning the objective with the squared generalized cosine similarity (SGCS) evaluation metric.The loss is computed over dataset D with cardinality |D|.
C. Codeword Fingerprint: The statistical fingerprint uses a DFT codebook of size Ncode = 2B, suited to uniform linear arrays because it captures long-term angular channel characteristics.The codebook is W = [w0, w1, . . . , wNcode−1].
C. Codeword Fingerprint: For each precoding vector, the angular fingerprint is obtained by projecting v onto the DFT codebook and identifying the dominant beam direction.The projection produces f = v^H W = [f1, f2, . . . , fNcode].
C. Codeword Fingerprint: Aggregating dominant directions yields an angular probability-density approximation whose energy spread reflects environmental characteristics: concentrated in LOS and broader in NLOS.In LOS, the dominant direction corresponds to the LOS path’s AoD; in NLOS, it corresponds to the strongest scattering path’s AoD. QuaDRiGa 3GPP-38.901-UMi-LOS/NLOS data show this concentration-versus-dispersion contrast across 5 m × 5 m subregions.

III. LEARNWARE FRAMEWORK · A. Motivation

The Learnware framework introduces a centralized CSI-feedback model repository that reuses scene-specific small models through statistical specifications. It targets the trade-off between per-scene performance and generalization while reducing the computational and deployment burdens associated with large models.

III. LEARNWARE FRAMEWORK: The section presents a Learnware-based repository for CSI-feedback models.The repository is integrated into the CSI-feedback workflow and supports model reuse.
III. LEARNWARE FRAMEWORK: The repository is centralized to organize scene-specific CSI-feedback models.Centralized repository integration is identified as part of the framework’s motivation.
A. Motivation: Scene-specific small models are motivated by their effectiveness in particular environments.The section identifies the efficacy of small models as a key insight.
A. Motivation: AI-driven CSI feedback must balance generalization against per-scene performance.This is identified as one of two primary deployment challenges.
A. Motivation: Large models create computational and deployment-cost challenges.Managing these costs is the second primary challenge identified for AI-driven CSI-feedback deployment.
A. Motivation: Statistical specifications enable efficient reuse of scene-specific models.The repository identifies reusable models through statistical specifications.
A. Motivation: The framework highlights statistical specifications as a source of deployment efficiency.Statistical specifications are listed among the three key insights motivating the repository.
A. Motivation: Accelerated fine-tuning convergence is another key insight of the proposed repository.The motivation explicitly identifies accelerated fine-tuning convergence alongside small-model efficacy and statistical specifications.

1) Scene-Specific Small Models vs. General-Purpose Models: … 2) Specification Design:

The paper favors reusable scene-specific small models over broad general-purpose models and organizes them through a Learnware repository. Its specification design combines semantic descriptors for functional matching with privacy-preserving statistical descriptors for distribution-aware retrieval and adaptation.

1) Scene-Specific Small Models vs. General-Purpose Models:: Scene-specific small models capture localized propagation characteristics, use fewer resources, and outperform general models in their target environments.Large general-purpose models simplify management but require extensive compute and memory and can perform suboptimally in specific environments.
2) Statistical Specifications vs. Semantic Specifications:: Statistical specifications, such as codeword fingerprint histograms, distinguish training-data distributions that coarse structural or semantic descriptors cannot capture.They index models in a continuous specification space that can generalize to unseen environments and avoid inefficient retrieval.
3) Rapid Fine-Tuning Convergence:: Closer alignment between a model’s training distribution and the target domain reduces fine-tuning iterations and samples, while a dense repository can enable direct deployment.Convergence is illustrated for mismatched, general, and nearest Learnware models.
B. Learnware Paradigm and Specification Design: The Learnware-based repository promotes scene-specific model reuse, statistical-specification retrieval, and efficient fine-tuning adaptation.Its framework is organized around model submission and deployment stages.
B. Learnware Paradigm and Specification Design: The paradigm lets contributors keep training data local while users reuse pre-trained models through specification-based matching.It provides a heterogeneous, evolving alternative to training generalized models from aggregated datasets.
1) Model Submission:: The system supports large-scale model contribution through a diverse repository, with users able to reuse or refine identified or reassembled learnwares.Most CSI-feedback models are trained centrally in an AI resource data center, which maintains a broad collection.
2) Specification Design:: The specification world groups functionally similar models into specification islands, coupling a discrete functional space with a continuous statistical space.Users first identify an island through semantic requirements, then select the closest model by comparing statistical similarity.
2) Specification Design:: Semantic specifications describe task, architecture, and system configuration, whereas statistical specifications characterize training-data distributions without sharing original data.Statistical specifications are generated offline by developers and packaged with models and semantic tags; the framework uses the codeword fingerprint distribution pdf.

3) Validation of Codeword Fingerprint PDF:

The validation compares raw CSI distances, codeword-fingerprint PDF similarities, and normalized cross-dataset SGCS performance across 144 LOS/NLOS datasets. PDF similarity aligns more closely with transfer performance, supporting its use as a Learnware retrieval specification, while showing greater sensitivity to LOS scenes.

Validation results: PDF similarity aligns more closely with performance than raw CSI distance, confirming PDF as a more effective proxy for model transferability and Learnware retrieval.The comparison supports the principle that statistically similar datasets should produce similar transferred-model performance.
Dataset and matrix construction: 144 scenario-specific datasets—72 LOS and 72 NLOS—provide full 360-degree azimuthal coverage for evaluating dataset similarity and model generalization.The datasets are arranged counterclockwise by azimuth angle relative to a central BS.
Dataset and matrix construction: The evaluation compares raw-CSI Euclidean distances, PDF-vector cosine similarities, and normalized SGCS transfer-performance matrices.Each performance entry represents a model trained on one dataset and tested on another, normalized by the test dataset’s self-performance.
Validation results: PDF is more sensitive to LOS-scene variations than NLOS variations, reflected in sharper LOS matrix structure and direct-path AoD concentration.NLOS PDFs instead reflect broader scattering profiles caused by multipath propagation.
Repository initialization: Initial configuration labels and simulation-trained CSI models can expand repository coverage when the available model diversity is limited.As models accumulate, both semantic and statistical specification coverage are expected to broaden.

IV. DEPLOYMENT STRATEGIES · A. Deploying Stage · B. Post-Retrieval Adaptation

The Learnware deployment strategy retrieves CSI models through compact specifications instead of raw local data, then adapts retrieved models when repository coverage or performance is insufficient. This reduces privacy exposure, communication overhead, and retrieval latency while supporting near-real-time updates.

IV. DEPLOYMENT STRATEGIES: The deployment stage combines online model retrieval, comparison with alternative methods, and an optimized search strategy to accelerate model lookup.The framework addresses both model-selection effectiveness and post-retrieval adaptation overhead.
A. Deploying Stage: The BS monitors for model mismatch and submits a Model Delivery Request so the Learnware marketplace can identify and deliver suitable models.Deployment must retrieve models that satisfy task requirements while maximizing returned-model utility and minimizing adaptation overhead.
A. Deploying Stage: The Original Model Repository uploads local CSI validation data, evaluates every candidate model, and returns the highest-performing model.This baseline achieves high retrieval accuracy but requires sequential evaluation of all candidates.
A. Deploying Stage: The baseline compromises privacy and intellectual property, incurs substantial communication overhead, and scales poorly, making it impractical for real-time CSI deployment.Its limitations arise from uploading local CSI data and sequentially evaluating candidate models.
A. Deploying Stage: Learnware generates specifications locally, sends them as requests, uses semantic specifications to locate a specification island, and indexes similar models with statistical specifications.An exact statistical match is not required, and raw CSI remains at the BS.
A. Deploying Stage: Sparse repository coverage and uncertain ideal local performance can produce suboptimal matches, so user-defined thresholds govern fine-tuning based on samples, iterations, and resources.Fine-tuning addresses the gap between retrieved-model performance and an optimally trained model.
B. Post-Retrieval Adaptation: Online CSI deployment involves the AI data center, BS, and UE: the center trains and deploys a two-sided network, while the BS requests updates after performance degradation.The UE collects new CSI samples, which the BS can use either to fine-tune the current model or pursue another adaptation path.
B. Post-Retrieval Adaptation: Compared with the Original Model Repository, Learnware sends only compact statistical specifications, reducing raw CSI exposure and retrieval time while enabling near-real-time updates.The trade-off is a possible slight reduction in selection accuracy relative to exhaustive raw-data evaluation.

C. Comparison with Alternative Methods · D. Adjustable-Speed Model Search Strategy

The paper compares repository-based CSI feedback deployment alternatives and introduces a Learnware repository that matches statistical specifications to retrieve suitable pre-trained models. It further proposes an adjustable-speed hierarchical LSH search strategy that balances retrieval latency, accuracy, and scalability.

C. Comparison with Alternative Methods: The evaluation compares the General Model, fine-tuned General Model, Model Switch, Original Model Repository, and Learnware Model Repository.These alternatives differ in model selection, local fine-tuning, repository interaction, and specification-based retrieval.
C. Comparison with Alternative Methods: The Learnware repository retrieves a pre-trained model using uploaded statistical specifications and semantic and statistical matching, avoiding local-data upload.The comparison considers initialization and deployment trade-offs in accuracy, communication cost, computation overhead, and data privacy.
D. Adjustable-Speed Model Search Strategy: The search strategy accelerates retrieval by replacing impractical brute-force cosine-similarity computation across all repository candidates.Models with similar statistical specifications tend to be functionally interoperable, supporting faster candidate retrieval.
D. Adjustable-Speed Model Search Strategy: The proposed adjustable-speed multi-level LSH strategy combines coarse-grained anchor selection, hierarchical indexing, and early termination in a K-level cascade.The method is specified as Hierarchical LSH Retrieval in Algorithm 1.
D. Adjustable-Speed Model Search Strategy: Offline indexing builds K LSH levels with increasing hash precision, using k1 < k2 < · · · < kK and L1 > L2 > · · · > LK.Coarse levels use more tables for high recall, while finer levels use fewer tables with longer hash keys for precise filtering.
D. Adjustable-Speed Model Search Strategy: Small kstop enables sub-millisecond latency for urgent updates, whereas large kstop uses the full cascade for near-optimal accuracy.Confidence-based early exit stops retrieval when anchors satisfy performance requirements, avoiding unnecessary computation.
D. Adjustable-Speed Model Search Strategy: The per-level retrieval complexity is O(Lℓkℓd+Lℓd), independent of repository size N, enabling scalable deployment.The adjustable-speed design therefore exposes a speed–accuracy trade-off through level selection and confidence-based termination.

V. SIMULATION RESULTS AND DISCUSSIONS … 2) NN Training and Evaluation:

The simulations establish a Learnware repository with diverse CSI-trained models and evaluate statistical-specification retrieval using lightweight autoencoders against a transformer-based General Model. Testing covers LOS/NLOS environments and compares both initial and optional fine-tuned performance.

A. Learnware Model Repository Initiation: The repository was initialized with an implicit-feedback CSI two-sided network to construct a specification island.Its corresponding semantic labels are reported in Table III.
1) CSI Simulation:: 144 Learnwares were created under 3GPP-38.901-UMi, comprising 72 LOS and 72 NLOS models with 10,000 samples per dataset.CSI sampling covered full 360° azimuth diversity, with independently restarted environment layouts for each dataset.
2) NN Training and Evaluation:: Each Learnware uses a lightweight 1D-convolutional CsiNet autoencoder, while the General Model uses the transformer-based EVCsiNet-T.The CsiNet design replaces 2D convolutions with 1D convolutions and expands the kernel size from 3 to 7.
2) NN Training and Evaluation:: The datasets use 85% training, 10% validation, and 5% test splits, with Adam optimization at a 1e-4 learning rate for 1,000 epochs.Models were saved at the point of optimal validation performance.
2) NN Training and Evaluation:: Targeted tests generate CSI over randomly centered and rotated rectangular areas of 5 m × [20] m under LOS/NLOS configurations.The test environments match the semantic configurations used by the repository models.
2) NN Training and Evaluation:: The proposed method retrieves the best Learnware using statistical-specification similarity and returns it for optional fine-tuning.Evaluation compares both the initial retrieved-model performance and the fine-tuned performance.
2) NN Training and Evaluation:: Table V presents the CNN and transformer-based autoencoder structures used for CSI feedback.The table documents the architectures underlying the Learnware models and General Model.

B. Performance Gain · 1) Initial Performance Verification:

The paper evaluates repository-based CSI models against generalized and scene-specific alternatives, showing that small scene-specific models provide stronger initial performance while Learnware retrieval remains robust across feedback settings. Model selection uses codebook-fingerprint similarity, but NLOS scenarios may still require fine-tuning or increased feedback overhead.

B. Performance Gain: The evaluation compares models with the same architecture but different training processes, including General Model, Model Switch, Original Model Repository, and Learnware Model Repository.The General Model uses a transformer trained on 144 datasets, while Model Switch uses four CNN models trained on groups of 36 datasets.
B. Performance Gain: The Learnware Model Repository selects the model whose codebook-fingerprint sample-frequency pdf has the highest cosine similarity to the target scene, using N=64.Similarity is computed among Learnwares within the specified island.
1) Initial Performance Verification:: Specific CNN models with 2k parameters and 0.11M FLOPs consistently outperform transformer-based General Models TF-1 through TF-4 without fine-tuning.The comparison reports SGCS performance averaged over 10 cases in each sampling-area range.
1) Initial Performance Verification:: 0.05–0.45 (6.4%–128.2%) SGCS gains occur in LOS, while 0.05–0.09 (12.4%–60.4%) gains occur in NLOS versus the General Model across compression and quantization settings.The Learnware Model Repository maintains these gains across all configurations analyzed in Tab. VII.
1) Initial Performance Verification:: 0.794 (LOS) and 0.215 (NLOS) are achieved by the Learnware at CR=1/32 and Q=4, exceeding the General Model values of 0.348 LOS and 0.134 NLOS.These results demonstrate robustness under highly constrained feedback.
1) Initial Performance Verification:: The fingerprint method selects among LOS Learnwares for LOS scenarios and among NLOS Learnwares for NLOS scenarios, so LOS/NLOS is excluded from semantic specifications.The selected model is the one with the highest cosine similarity between its training-data pdf and the target scene.
1) Initial Performance Verification:: A predefined fine-tuning threshold is unreliable because the Upper Bound is relatively high in LOS but relatively low in NLOS, where fine-tuning becomes necessary.Increasing feedback overhead by switching to another specification island is proposed as one way to improve NLOS performance.

2) Fine-Tuning Performance: · C. Adjustable-Speed LSH · D. Complexity Analysis

Fine-tuning the selected Learnware improves performance with limited overhead, while adjustable-depth LSH provides controllable retrieval accuracy and latency. The repository-based approach also reduces communication and deployment costs relative to conventional model schemes.

2) Fine-Tuning Performance:: Minimal fine-tuning boosts selected Learnware performance to parity with the Original model repository and sometimes the Upper bound.The selected model provides a better initialization point, reducing fine-tuning overhead.
2) Fine-Tuning Performance:: Across 10 LOS and 10 NLOS cases in multiple sampling areas, repository-based approaches provide superior initialization and sometimes outperform fine-tuned General Model.Fine-tuning used 500 samples for 100 epochs, with yellow bars denoting the resulting gain.
2) Fine-Tuning Performance:: Learnware improves more slowly than General Model because weaker generalization and proximity to the upper bound make optimal convergence more difficult.The closer starting point to the upper bound slows gradient descent toward the optimum.
C. Adjustable-Speed LSH: The adjustable-speed LSH evaluation constructs a 100-million-vector statistical specification space from 1,000 QuaDRiGa CSI datasets and GMM-based distribution representations.The datasets cover diverse urban microcell scenarios, and their probability density functions are computed before GMM fitting.
C. Adjustable-Speed LSH: Deeper LSH search increases accuracy and cumulative latency, allowing retrieval depth to match deployment requirements.The strategy is evaluated specifically for accuracy-latency trade-offs in real-time CSI feedback deployment.
D. Complexity Analysis: 373.5 seconds of TF-4 fine-tuning exceeds the CNN model’s 61.6 seconds, whereas Learnware retrieval takes 1.48e-5 seconds and LSH retrieval remains within 0.25–1.15 ms.The TF-4 Transformer has 2.1 million parameters; the LSH latency remains below the 10–80 ms CSI feedback cycle even for billion-model repositories.
C. Adjustable-Speed LSH: 0.25–0.60 ms retrieval latency and 85.9–91.9% accuracy characterize shallow L1-L2 search, whereas L3-L4 reaches 97.0–99.0% accuracy at 0.90–1.15 ms.The deeper levels provide near-optimal accuracy at moderately higher latency.
D. Complexity Analysis: 240 KB of local-data transmission is reduced to 2.6 KB by Learnware, while the TF-4 encoder is 33.7 MB versus 5.83 KB for the CNN encoder.The TF-4 encoder is three orders of magnitude larger than the CNN encoder, making frequent updates impractical in bandwidth-constrained settings.

E. Further Discussion … 3) Impact of Sample Size on Model Selection Performance:

The discussion validates PDF-based Learnware selection, examines repository size, and shows that reliable model selection remains robust with limited CSI samples. Across these analyses, selection performance is strong relative to repository and benchmark references.

E. Further Discussion: The study validates PDF-based Learnware selection using 1,000 LOS and 1,000 NLOS scenarios per sampling area, each containing 500 samples.The samples generate PDFs and evaluate performance.
1) Validation for pdf Selection:: Over 90% of selected Learnwares rank within repository positions 25 to 27 among 144 models.Top-three ranking probability is 45% to 55% for LOS and 25% to 35% for NLOS scenarios.
1) Validation for pdf Selection:: Performance-gap distributions are more concentrated for smaller LOS sampling areas but become more concentrated as sampling area increases in NLOS scenarios.The boxplot summarizes the first quartile, median, third quartile, and interquartile range.
1) Validation for pdf Selection:: Approximately 95% LOS and 90% NLOS selections satisfy the SGCS performance-gap threshold ϵ = 0.1, exceeding random selection.Acceptable cases have gap ≤0.1, while failures have gap > 0.1.
2) Impact of Learnware Quantity:: Selected-Learnware performance improves as the model library grows, with the 144-model repository providing the best-Learnware reference.Library sizes were randomly formed from a pool of 144 and evaluated over 100 repetitions.
3) Impact of Sample Size on Model Selection Performance:: The sample-size analysis evaluates how CSI samples affect PDF similarity and SGCS performance against a full-sample benchmark.PDF similarity uses estimated-versus-benchmark PDFs, while SGCS performance is compared with benchmark SGCS.
3) Impact of Sample Size on Model Selection Performance:: BSs can collect as few as 500 samples for reliable model selection regardless of propagation conditions, avoiding extensive practical data collection.This conclusion follows from the observed robustness to sample-size variations.

4) Comparison of Distributional Similarity Metrics: · 5) Multi-carrier OFDM Scenario: · VI. CONCLUSION

The paper selects Cosine similarity for robust, efficient learnware retrieval across LOS and NLOS conditions, and validates framework extensibility in multi-carrier OFDM while reducing deployment overhead through model reuse.

4) Comparison of Distributional Similarity Metrics:: Seven distributional similarity metrics are evaluated for learnware retrieval from codebook-fingerprint probability distributions.The metrics are Cosine similarity, Bhattacharyya coefficient, Hellinger distance, Total Variation, KL divergence, Jensen-Shannon divergence, and Euclidean distance.
4) Comparison of Distributional Similarity Metrics:: KL divergence matches the repository upper bound in LOS large-range environments but falls to 0.209 at 5×50 versus 0.261 upper bound in NLOS diffuse conditions.Cosine similarity instead remains stable across NLOS scenarios, ranging from 0.324–0.233.
4) Comparison of Distributional Similarity Metrics:: Cosine similarity is selected because L2-normalization emphasizes relative angular patterns, avoids sensitivity to absolute energy concentration, and supports computationally efficient real-time retrieval.Its complexity is described as low compared with the high complexity of KL divergence.
5) Multi-carrier OFDM Scenario:: The multi-carrier OFDM experiment transforms a 512-subcarrier channel into the angle-delay domain using 2D-DFT and prunes it to the first 32 delay taps.Angle and delay distributions are formed by aggregating dominant angle and delay indices across the dataset.
5) Multi-carrier OFDM Scenario:: 200 Learnware models are constructed with TF-1 architecture and trained on 10,000 samples each, while a TF-4 General Model uses a mixture of 2 million samples.The repository contains 100 LOS and 100 NLOS models; the General Model uses LTF = 12 and Embedding dim = 192.
5) Multi-carrier OFDM Scenario:: The selected Learnware consistently outperforms the General Model and approaches the Original Model Repository on representative 5m × 5m LOS/NLOS test cases.This result confirms the framework’s effectiveness in the multi-carrier OFDM scenario.
VI. CONCLUSION: The proposed centralized AI-data-center framework enables model reuse across diverse deployment scenarios, reducing training overhead and improving adaptability.It targets data privacy concerns, communication costs, and inefficient retrieval in conventional model repositories.