Apparatuses, methods, and computer programs for a network node and for user equipment and for maintaining a machine-learning model

Abstract

Provided are apparatuses, methods, and computer programs for a network node and for user equipment of a mobile communication system. The method (10) for the network node (300) of a mobile communication system (500) comprises providing (12) a baseline dataset to user equipment (400), UE, the base line dataset forming a baseline for a machine-learning model that is used at the UE (400) for parameter prediction. The method (10) also includes configuring (14) an accuracy threshold at the UE (400), the accuracy threshold allowing to determine a need for adaptation of the machine-learning model used at the UE (400) in case a prediction accuracy fails to meet an accuracy requirement, which is based on the accuracy threshold.

Description

[0001] 202500414 1

[0002] Apparatuses, Methods, and Computer Programs for a Network Node and for User Equipment and for Maintaining a Machine-Learning Model

[0003] Description

[0004] The present disclosure relates to apparatuses, methods, and computer programs for a network node and for user equipment of a mobile communication system, and more particularly, but not exclusively, to a concept for maintaining a machine-learning model at user equipment in a mobile communication system.

[0005] Machine-learning models deployed in wireless communication systems, particularly those operating at user equipment (UE) or network nodes in accordance with 3GPP (3rd Generation Partnership Project) specifications for the NR (New Radio) air interface, are subject to the development and accumulation of bias throughout their operational lifetime. Bias manifests as systematic performance degradation that evolves across distinct phases of the model lifecycle, presenting significant challenges for maintaining reliable AI / ML-enabled (Artificial Intelligence / Machine Learning) features in commercial deployments.

[0006] During the initial training phase, bias is first introduced into a machine-learning model through various mechanisms inherent to the development process. Training datasets may not comprehensively cover all real-world scenarios, configurations, or deployment sites that the model will encounter during operation. Ground truth labels used for supervised learning may contain systematic errors that become embedded in the model's learned representations. Furthermore, the choice of model architecture and hyperparameter optimization may inadvertently favor certain operating conditions over others. A particularly problematic form of initial bias arises from overfitting, wherein the model learns the training data too specifically, including noise and artifacts that do not generalize to operational environments. Models that are optimized for standardized test setups may demonstrate excellent performance during laboratory evaluation while exhibiting compromised robustness in real-world environments that do not align with the controlled training conditions.202500414 2

[0007] When a trained model transitions to the deployment phase in a live network, additional bias can emerge from mismatches between training and operational conditions.

[0008] Scenario mismatches occur when models are deployed in environments not adequately represented in the training data, such as a model trained on urban propagation data being deployed in rural or indoor factory settings. Configuration mismatches arise when the deployed network employs different antenna configurations, frequency bands, or network settings than those present during training. Site-specific differences introduce bias due to unique propagation characteristics of new deployment locations, while hardware variations across different UE implementations or chipsets can cause systematic prediction errors.

[0009] Throughout the operational phase, bias accumulates progressively due to phenomena commonly referred to as data drift and concept drift. Data drift, also known as covariate shift, occurs when the statistical distribution of input data changes over time relative to the training distribution. Environmental changes such as seasonal variations, construction of new buildings, and changes in foliage can alter radio propagation characteristics. Network evolution through the deployment of new cells or configuration updates modifies the operational context. Changes in user behavior patterns, including different mobility characteristics and usage scenarios, further contribute to input distribution shifts. Concept drift represents a more fundamental form of bias development wherein the underlying relationship between model input and desired output changes over time. Technology updates that introduce new network features that are not present during training, evolving interference patterns from new sources in the environment, and regulatory changes affecting power levels or frequency allocations all contribute to concept drift.

[0010] The temporal progression of bias accumulation follows a characteristic pattern in deployed machine-learning models. Initial performance levels achieved at deployment gradually degrade as data drift begins to affect prediction accuracy. This degradation accelerates as concept drift compounds the effects of data drift, eventually leading to performance levels that fall below acceptable thresholds. For specific use cases defined in 3GPP specifications, this bias evolution manifests in distinct ways. In channel state information prediction using UE-sided models, training data reflects specific channel conditions that become increasingly unrepresentative as seasonal changes affect202500414 3

[0011] propagation, new construction alters multipath characteristics, and network densification modifies interference patterns over months and years of operation. In beam management applications, models optimized for current beam configurations become biased when new beams are added, beam patterns are modified, or environmental changes such as the addition or removal of reflective surfaces alter the coverage area. For positioning applications, fingerprint databases that accurately reflect the environment at the time of training become progressively biased as facility layouts are modified, machinery is relocated, or new network infrastructure is deployed.

[0012] To address the development of bias over the model lifetime, lifecycle management mechanisms have been defined within the 3GPP framework for AI / ML-enabled features. Performance monitoring functions continuously evaluate model performance to detect when accumulated bias causes degradation below acceptable thresholds. Model switching techniques enable the selection among multiple scenario-specific, configuration-specific, or site-specific models to address bias arising from operational condition changes. Model update and fine-tuning processes enable flexible adaptation of model structures or parameters in response to detected drift. Fallback mechanisms ensure that when bias causes AI / ML-enabled features to underperform relative to legacy non-AI / ML-based operations, the system can revert to conventional methods to maintain reliable service. The functional framework for AI / ML lifecycle management encompasses data collection for ongoing operational monitoring, model training and adaptation for bias correction, model storage for maintaining multiple model versions, model inference for applying current models, and management functions for coordinating monitoring, switching, and update operations throughout the model lifetime.

[0013] Further details can be found in:

[0014] 1. US20220337487A1

[0015] 2. WO2022161624A1

[0016] 3. WO2022258149A1

[0017] 4. US20220108214A1

[0018] 5. US20220400373A1

[0019] 6. WO2022228666A1

[0020] 7. WO2023277780A1202500414 4

[0021] Various examples of the present disclosure are based on the finding that machinelearning models deployed at user equipment in mobile communication systems may experience degradation in prediction accuracy over time due to changing network conditions, user behavior patterns, or environmental factors. The present disclosure relates to a technique for managing machine-learning model accuracy at user equipment by establishing baseline datasets and accuracy thresholds that enable timely detection and adaptation of machine-learning models when their performance deteriorates.

[0022] The proposed concept provides a coordinated approach between network nodes and user equipment for monitoring and maintaining machine-learning model performance. By configuring accuracy thresholds at the user equipment and providing baseline datasets from the network, the system enables autonomous evaluation of model performance and triggers adaptation procedures when accuracy requirements are not met. This improves the reliability and effectiveness of machine-learning-based parameter prediction in mobile communication systems, ensuring that predictions remain accurate and useful for network operations even as conditions change. The proposed concept results in more efficient use of network resources by allowing user equipment to identify when model updates are needed rather than requiring continuous monitoring by the network or accepting degraded performance.

[0023] Some aspects of the present disclosure relate to a method for a network node of a mobile communication system. The method comprises providing a baseline dataset to user equipment, UE, the baseline dataset forming a baseline for a machine-learning model that is used at the UE for parameter prediction. The method further comprises configuring an accuracy threshold at the UE, the accuracy threshold allowing determination of a need for adaptation of the machine-learning model used at the UE in case a prediction accuracy fails to meet an accuracy requirement, which is based on the accuracy threshold. By providing both the baseline dataset and the accuracy threshold, the network node establishes a framework for quality control of machine-learning predictions at the user equipment, enabling consistent performance standards across the mobile communication system.202500414 5

[0024] To enable responsive adaptation when model performance degrades, the method may further comprise receiving a request for machine-learning-model adaptation from the UE. This allows the user equipment to initiate corrective actions when its evaluation indicates that the machine-learning model is no longer meeting accuracy requirements.

[0025] For scenarios where the machine-learning model requires updating to restore accuracy, in various examples, the request may be a request for re-training of the machinelearning model. The method may further comprise re-training the machine-learning model at the UE and / or at a training node. This enables the machine-learning model to be updated with new training data that better reflects current conditions, thereby restoring prediction accuracy.

[0026] To address situations where a pre-trained or updated model is available at the network, alternatively, the request may be a request for downloading the machine-learning model or an adapted machine-learning model. The method may further comprise forwarding the machine-learning model or the adapted (e.g. trained or re-trained) machine-learning model to the UE. This provides a mechanism for rapidly deploying improved models without requiring re-training at the UE.

[0027] To accommodate changing network conditions and varying performance requirements, in some examples, the method may further comprise dynamically adjusting the accuracy threshold. The configuring may comprise dynamically signaling the accuracy threshold to the UE. By dynamically adjusting the accuracy threshold, the network node can adapt quality requirements based on current network conditions, traffic patterns, or service requirements, thereby optimizing the balance between prediction accuracy and adaptation overhead.

[0028] In various examples, the machine-learning model may be a federated machine-learning model. This allows multiple user equipment devices to collaboratively train and improve the machine-learning model while maintaining data privacy, as training data remains distributed across the participating devices.

[0029] Some aspects of the present disclosure relate to a method for user equipment, UE, of a mobile communication system. The method comprises receiving a baseline dataset202500414 6

[0030] from a network node of the mobile communication system, the baseline dataset forming a baseline for a machine-learning model that is used at the UE for parameter prediction.

[0031] The method further comprises receiving a configuration for an accuracy threshold at the UE from the network node, the accuracy threshold allowing determination of a need for adaptation of the machine-learning model used at the UE in case a prediction accuracy fails to meet an accuracy requirement based on the accuracy threshold. The method further comprises evaluating the machine-learning model using the accuracy threshold. The method further comprises requesting machine-learning model adaptation if the evaluating indicates that an accuracy of the machine-learning model fails to meet the accuracy requirement based on the accuracy threshold. By autonomously evaluating model performance and requesting adaptation when needed, the user equipment ensures that parameter predictions remain reliable without requiring continuous network supervision, thereby improving both prediction quality and network efficiency.

[0032] For scenarios where model retraining is the preferred adaptation approach, in some examples, the requesting may comprise requesting a retraining of the machine-learning model. The method may further comprise receiving retraining data from the network node and retraining the machine-learning model based on the retraining data. This enables the user equipment to update its machine-learning model with fresh training data that reflects current conditions.

[0033] To enable rapid model updates using pre-trained models, alternatively, the requesting may comprise requesting a download of the machine-learning model or an adapted machine-learning model. The method may further comprise downloading the machinelearning model or the adapted (e.g. trained or re-trained) machine-learning model via the network node and installing the downloaded machine-learning model. This provides a mechanism for quickly deploying improved models that have been trained or adapted elsewhere in the network.

[0034] To systematically assess model performance, in various examples, the evaluating may comprise calculating a prediction accuracy of the machine-learning model and further evaluating if the prediction accuracy does not meet the accuracy requirement. This provides a quantitative basis for determining whether model adaptation is necessary.202500414 7

[0035] To distinguish between inherent model limitations and potential data-related issues, in some examples, the evaluating may further comprise testing the machine-learning model to obtain a tested machine-learning model based on the baseline data in case the accuracy of the machine-learning model does not meet the accuracy requirement. By testing the machine-learning model against the baseline data, the user equipment can determine whether performance degradation is due to model drift or other factors.

[0036] For comprehensive performance assessment of the tested model, in various examples, the evaluating may further comprise calculating a prediction accuracy of the tested machine-learning model and comparing the prediction accuracy to the accuracy requirement. This enables the user equipment to verify whether testing with baseline data improves performance or whether model adaptation is genuinely needed.

[0037] To ensure that adaptation requests are made only when truly necessary, in some examples, the requesting of the machine-learning model adaptation may be carried out if the prediction accuracy of the tested machine-learning model fails to meet the accuracy requirement based on the accuracy threshold. This prevents unnecessary adaptation requests when performance issues can be resolved through testing with baseline data.

[0038] Further aspects of the present disclosure relate to a computer program having a program code for performing the methods described above, when the computer program is executed on a computer, a processor, or a programmable hardware component. This enables the proposed concept to be implemented in software on various computing platforms.

[0039] Additional aspects of the present disclosure relate to an apparatus for a network node of a mobile communication system. The apparatus comprises one or more interfaces configured to communicate in the mobile communication system. The apparatus further comprises one or more processing devices configured to perform the methods for the network node as described above. This provides a hardware implementation of the network node functionality for managing machine-learning model accuracy.202500414 8

[0040] Further aspects of the present disclosure relate to an apparatus for UE of a mobile communication system. The apparatus comprises one or more interfaces configured to communicate in the mobile communication system. The apparatus comprises one or more processing devices configured to perform the methods for the UE as described above. This provides a hardware implementation of the UE functionality for evaluating and adapting machine-learning models.

[0041] Some aspects of the present disclosure relate to a network node of a mobile communication system comprising the apparatus for the network node as described above. This provides a complete network node implementation incorporating the proposed machine-learning model management capabilities.

[0042] Additional aspects of the present disclosure relate to a UE of a mobile communication system comprising the apparatus for the UE as described above. This provides a complete UE implementation incorporating the proposed machine-learning model evaluation and adaptation capabilities.

[0043] Further aspects of the present disclosure relate to a mobile communication system comprising the network node and the UE as described above. This provides a complete system implementation that enables coordinated machine-learning model management across network infrastructure and user devices.

[0044] Some examples of apparatuses, computer programs, and / or methods will be described in the following by way of example only, and with reference to the accompanying figures, in which

[0045] Fig. 1 shows an example of a flowchart of a method for a network node;

[0046] Fig. 2 depicts an example of a flowchart of a method for user equipment;

[0047] Fig. 3 illustrates block diagrams of examples of apparatuses for a network node, user equipment, and a communication network;

[0048] Fig. 4 shows an evaluation procedure at user equipment in an example;202500414 9

[0049] Fig. 5 depicts another example of a flow chart of a method for a network node; and

[0050] Fig. 6 illustrates a flowchart of another example of a method for user equipment.

[0051] Some examples are now described in more detail with reference to the enclosed figures. However, other possible examples are not limited to the features of these embodiments described in detail. Other examples may include modifications of the features, as well as equivalents and alternatives to the features. Furthermore, the terminology used herein to describe certain examples should not be restrictive of further possible examples.

[0052] Throughout the description of the figures, same or similar reference numerals refer to same or similar elements and / or features, which may be identical or implemented in a modified form while providing the same or a similar function. The thickness of lines, layers, and / or areas in the figures may also be exaggerated for the sake of clarification.

[0053] When two elements A and B are combined using an "or," this is to be understood as disclosing all possible combinations, i.e. , only A, only B, as well as A and B, unless expressly defined otherwise in the individual case. As an alternative wording for the same combinations, "at least one of A and B" or "A and / or B" may be used. This applies equivalently to combinations of more than two elements.

[0054] If a singular form, such as "a", "an", and "the" is used and the use of only a single element is not defined as mandatory either explicitly or implicitly, further examples may also use several elements to implement the same function. If a function is described below as implemented using multiple elements, further examples may implement the same function using a single element or a single processing entity. It is further understood that the terms "include", "including", "comprise", and / or "comprising", when used, describe the presence of the specified features, integers, steps, operations, processes, elements, components, and / or a group thereof, but do not exclude the presence or addition of one or more other features, integers, steps, operations, processes, elements, components, and / or a group thereof.202500414 10

[0055] Fig. 1 shows an example of a flowchart of method 10 for a network node. The method 10 for the network node of a mobile communication system comprises providing 12 a baseline dataset to user equipment, UE. The baseline dataset forms a baseline for a machine-learning model that is used at the UE for parameter prediction. The method 10 further comprises configuring 14 an accuracy threshold at the UE. The accuracy threshold allows determining a need for adaptation of the machine-learning model used at the UE in case a prediction accuracy fails to meet an accuracy requirement, which is based on the accuracy threshold.

[0056] Fig. 2 depicts an example of a flowchart of a method 20 for a UE. The method 20 for the UE of a mobile communication system comprises receiving 22 a baseline dataset from a network node of the mobile communication system, the baseline dataset forming a baseline for a machine-learning model that is used at the UE for parameter prediction. The method 20 further comprises receiving 24 a configuration for an accuracy threshold at the UE from the network node. The accuracy threshold allows determining a need for adaptation of the machine-learning model used at the UE in case a prediction accuracy fails to meet an accuracy requirement. The accuracy requirement is based on the accuracy threshold. The method 20 further comprises evaluating 26 the machinelearning model using the accuracy threshold and requesting 28 machine-learning model adaptation if the evaluating indicates that an accuracy of the machine-learning model fails to meet the accuracy requirement based on the accuracy threshold.

[0057] Fig. 3 illustrates block diagrams of examples of apparatuses for a network node, user equipment, and a communication network. Fig. 3 illustrates an apparatus 30 for a network node 300 of a mobile communication system 500. The apparatus 30 comprises one or more interfaces 32 configured to communicate in the mobile communication system 500. The one or more interfaces 32 are coupled to one or more processing devices 34, which are configured to perform one of the methods 10 for the network node 300 as described herein.

[0058] Fig. 3 also shows an apparatus 40 for a UE 400 of the mobile communication system 500. The apparatus 40 comprises one or more interfaces 42 configured to communicate in the mobile communication system 500. The one or more interfaces 42 are coupled to202500414 11

[0059] one or more processing devices 44, which are configured to perform one of the methods 20 for the UE 400 as described herein.

[0060] As optional components (dashed lines), optional from the perspective of the apparatuses 30, 40, Fig. 3 also shows a network node 300 of the mobile communication system 500 comprising the apparatus 30 and a UE 400 of the mobile communication system 500 comprising the apparatus 40. Fig. 3 also illustrates an example of the mobile communication system 500 comprising the network node 300 and the UE 400.

[0061] As illustrated in Fig. 3, the respective one or more signal processing devices 34, 44 are coupled to the respective one or more interfaces 32, 42. The one or more interfaces 32, 42 may serve as an interface for communicating in the communication system 500. The one or more interfaces 32, 42 may correspond to one or more inputs and / or outputs for receiving and / or transmitting information, which may be in digital (bit) values or analog according to a specified code or protocol, within a module, between modules, or between modules of different entities. For example, an interface may comprise interface circuitry configured to receive and / or transmit information. In examples, an interface may comprise any means for obtaining, receiving, transmitting, or providing analog or digital signals or information, e.g., any connector, contact, pin, register, input port, output port, conductor, lane, etc., which allows providing or obtaining a signal or information.

[0062] The one or more interfaces 32, 42 may be configured to communicate (transmit, receive, or both) in a wireless and / or wired manner, and they may be configured to communicate, i.e., transmit and / or receive signals or information with further internal or external components. The one or more interfaces 32, 42 or the apparatuses 30, 40 may comprise further components to enable communication in a (mobile) communication system or network; such components may include transceiver (transmitter and / or receiver) components, such as one or more Low-Noise Amplifiers (LNAs), one or more Power-Amplifiers (PAs), one or more duplexers, one or more diplexers, one or more filters or filter circuitry, one or more converters, one or more mixers, accordingly adapted radio frequency components, one or more antennas, etc. For example, the respective one or more interfaces 32, 42 may enable radio communication with UEs202500414 12

[0063] and communication between base stations, which can be direct and / or indirect, wired and / or wireless, respectively.

[0064] The one or more (signal) processing devices 34, 44 may be implemented using one or more processing units, one or more circuitries, or any means for processing, such as a processor, a computer, or a programmable hardware component being operable with accordingly adapted software. In other words, the described function of the one or more processing devices 34, 44 may as well be implemented in software, which is then executed on one or more programmable hardware components. Such hardware components may comprise a general-purpose processor, a Digital Signal Processor (DSP), a microcontroller, etc.

[0065] In examples, network node 300 may also be referred to as a base station, server, network device, etc. It may belong to an access network or to a core network. A network entity / node may correspond to a remote radio head, a transmission point, an access point, a macro cell, a small cell, a micro cell, a pico cell, a femto cell, or a metro cell. The term small cell may refer to any cell smaller than a macro cell, e.g., a micro cell, a pico cell, a femto cell, or a metro cell. A network node / base station can be a wireless interface of a wired network, which enables transmission and reception of radio signals to a communication device. Such a radio signal may comply with radio signals as, for example, standardized by 3GPP or, generally, in line with one or more of the systems listed herein. Thus, a network entity / node may be a base station and may correspond to a NodeB, an eNodeB, an ngNB, a gNB, a BTS (Base Transceiver Station), or an access point, all of which may be implemented in a satellite, plane, HAPS (High Altitude Platform), etc. In case of a moving implementation in a satellite, an airplane, etc., the link towards a core network of the communication system may also be implemented in a wireless manner.

[0066] The mobile communication system 500 may hence be cellular. The term cell refers to a coverage area of radio services provided by a transmission point, a remote unit, a remote head, a remote radio head, communication device, network entity, or a NodeB, an eNodeB, an ngNB, a gNB, a beam, or a satellite, respectively. The terms cell and base station may be used synonymously; a base station may generate multiple cells and it may be implemented in a terrestrial device, a high-altitude platform, a plane, a202500414 13

[0067] drone, a satellite, etc. A UE or wireless communication device can be registered or associated with at least one cell (e.g., the network entity / node); e.g., it can be associated with a cell such that data can be exchanged between the network and the mobile in the coverage area of the associated cell using a dedicated channel, connection, or link.

[0068] In general, the UE is capable of communicating wirelessly. In particular, however, the UE / communication device may be a mobile communication device, e.g., a communication device that is suitable for being carried around by a user. For example, the communication device may be a User Terminal (UT) or UE within the meaning of the respective communication standards being used for mobile communication. For example, the communication device may be a mobile phone, such as a smartphone, a network access device embedded in a vehicle, ship, or airplane, or another type of mobile communication device, such as a computer, a laptop computer, a tablet computer, and so on.

[0069] For example, the UE / communication device and the network entity / node may be configured to communicate in a cellular mobile communication system. Accordingly, the UE and the network node may be configured to communicate in a cellular mobile communication system, for example in a Sub-6GHz-based cellular mobile communication system (covering frequency bands between 400 MHz and, in the meantime, 7 GHz), in a mmWave-based cellular mobile communication system (covering frequency bands between 24 GHz and 71 GHz), or in the so-called mid-bands (covering frequency bands between 7 GHz and 24 GHz). For example, the UE and the network node may be configured to communicate in a mobile communication system / cellular mobile communication system.

[0070] In general, the mobile communication system may, for example, correspond to one of the 3GPP-standardized mobile communication networks, where the term mobile communication system is used synonymously with mobile communication network. The mobile communication system may correspond to, for example, a 6th Generation system (6G), a 5th Generation system (5G), a New Radio (NR) system, a Long-Term Evolution (LTE) system, an LTE-Advanced (LTE-A) system, a High Speed Packet Access (HSPA) system, a Universal Mobile Telecommunication System (UMTS), a202500414 14

[0071] UMTS Terrestrial Radio Access Network (UTRAN), an evolved-UTRAN (e-UTRAN), a Global System for Mobile communication (GSM) or Enhanced Data rates for GSM Evolution (EDGE) network, a GSM / EDGE Radio Access Network (GERAN), or mobile communication networks with different standards, for example, generally an Orthogonal Frequency Division Multiple Access (OFDMA) network, a Time Division Multiple Access (TDMA) network, a Code Division Multiple Access (CDMA) network, a Wideband-CDMA (WCDMA) network, a Frequency Division Multiple Access (FDMA) network, a Spatial Division Multiple Access (SDMA) network, etc.

[0072] As further illustrated in Fig. 3, the apparatus 30 for the network node 300 provides / transmits information on a baseline dataset and an accuracy threshold to the apparatus 40 for the UE 400. At the apparatus 40, the machine-learning model is then evaluated using the accuracy threshold. Machine-learning model adaptation is then requested if the evaluation indicates that the accuracy of the machine-learning model fails to meet the accuracy requirement based on the accuracy threshold.

[0073] Machine-learning and artificial intelligence models deployed at user equipment, in examples, may be in accordance with 3GPP specifications for the NR air interface. They may be categorized according to several distinguishing characteristics that determine their operational behavior, deployment requirements, and lifecycle management procedures. There are one-sided models and two-sided models, which are distinguished based on the location of inference operations. In a one-sided UE-sided model, the complete inference operation is performed locally at the user equipment, with the model receiving input data derived from measurements or observations made by the device and producing output predictions or decisions without requiring real-time collaboration with the network during inference.

[0074] One-sided UE-sided models are applicable to use cases such as time-domain channel state information prediction, wherein the user equipment predicts future channel conditions based on historical measurements, beam management applications wherein the user equipment predicts optimal beam selections based on reference signal measurements of a subset of available beams, and positioning / handover prediction applications wherein the user equipment determines its location based on radio measurements using techniques such as fingerprinting. In contrast, two-sided models202500414 15

[0075] distribute the inference operation across both the user equipment and the network, with each entity performing a portion of the overall inference task. The primary application of two-sided models in 3GPP specifications is spatial-frequency domain channel state information compression, wherein the user equipment operates an encoder model to compress channel information into a reduced representation for transmission over the air interface, and the network operates a corresponding decoder model to reconstruct the full channel state information from the received compressed representation.

[0076] Machine-learning models at the user equipment may further be distinguished according to their identification and lifecycle management approach. Example definitions can be found in 3GPP specifications. Under functionality-based lifecycle management, the network and user equipment maintain mutual understanding of AI / ML-enabled features or feature groups, referred to as functionalities, without necessarily identifying the specific underlying AI / ML model or models that implement each functionality. The user equipment reports its supported functionalities through capability signaling mechanisms, and the network controls the selection, activation, deactivation, switching, and fallback of functionalities through standardized signaling, while the specific model implementations remain proprietary and potentially unknown to the network. Under model-identification-based lifecycle management, each AI / ML model is assigned a unique model identifier that enables both the network and user equipment to reference and manage specific models explicitly.

[0077] The model identifier may be accompanied by additional conditions specifying the scenarios, sites, configurations, or datasets for which the model is applicable, enabling fine-grained control over model selection based on current operating conditions. Models may further be characterized according to their generalization properties, ranging from models designed to generalize across multiple scenarios, configurations, and sites to scenario-specific, configuration-specific, or site-specific models optimized for particular operating conditions. The selection among multiple available models, the switching between models in response to changing conditions, and the updating or fine-tuning of models to address performance degradation constitute key lifecycle management operations that may be performed by the user equipment autonomously, under network control, or through collaborative procedures between the user equipment and the202500414 16

[0078] network depending on the deployment configuration and collaboration level defined for the AI / ML-enabled feature.

[0079] A baseline dataset, in the context of example machine-learning models deployed at UEs in wireless communication systems, comprises a defined set of parameters and associated reference values that enable the evaluation of model performance under controlled and reproducible conditions. The baseline dataset serves as a standardized benchmark against which the inference outputs of a machine-learning model can be compared to assess prediction accuracy, detect performance degradation, and determine whether the model continues to meet specified performance thresholds. The parameters constituting the baseline dataset may include input features representative of expected operational conditions, corresponding ground truth values or labels that represent correct model outputs, and metadata describing the scenarios, configurations, or environmental conditions under which the baseline data was collected or generated.

[0080] By maintaining a baseline dataset at the UE, the device can periodically evaluate its deployed machine-learning model by providing baseline inputs to the inference function and comparing the resulting outputs against the known reference values contained in the dataset, thereby enabling autonomous performance assessment without requiring real-time ground truth measurements from the network.

[0081] The composition and characteristics of the baseline dataset are tailored to the specific use case for which the machine-learning model is deployed. For channel state information prediction models, the baseline dataset may comprise historical channel measurements paired with corresponding actual channel realizations, enabling the user equipment to assess prediction accuracy by computing error metrics between model outputs and reference values. For beam management models, the baseline dataset may include reference signal measurements for a subset of beams together with known optimal beam selections, allowing evaluation of beam prediction accuracy. For positioning models, the baseline dataset may contain radio measurements associated with known reference positions, enabling assessment of positioning error statistics. The baseline dataset may be provisioned to the UE during initial model deployment, updated periodically by the network to reflect changing operational conditions, or generated locally by the user equipment during periods when ground truth information is available.202500414 17

[0082] The use of a baseline dataset for model evaluation provides a computationally efficient and overhead-reducing alternative to continuous ground-truth-based monitoring, enabling the user equipment to detect model performance degradation, trigger lifecycle management actions such as model switching or update requests, and ensure that AI / ML-enabled features maintain performance levels equal to or exceeding those of legacy non-AI / ML-based operations.

[0083] Fig. 4 shows an evaluation procedure at a UE in an example. Fig. 4 shows a UE 400, which evaluates an installed AI / ML model based on a baseline dataset. The baseline dataset comprises a set of input parameters and the AI / ML model provides a prediction, e.g. channel state information. The baseline dataset also provides output values, e.g. for a given environment and network conditions (defined by the input parameters) the baseline dataset also provides the output, i.e. the channel state information that should result. In Fig. 4 this process is illustrated as prediction. The predicted value can then be compared to the desired output from the baseline dataset. As a result, an inaccuracy can be determined as shown on the right of Fig. 4. Without bias the upper graph would result and with the actual model the lower graph results. The difference can be compared to the accuracy threshold obtained from the network node. If the threshold is exceeded as exemplified in the viewgraph of Fig. 4 with the value on the right (marked with “X”), model adaptation can be triggered. The bias may, for example, occur because of changing environmental data, disadvantageous re-training of the model, etc.

[0084] In examples, the evaluating 26 may comprise calculating a prediction accuracy of the machine-learning model. For example, this may be carried out by comparing predicted values and measured values. If the outcome of the comparison is that the prediction accuracy is better than demanded by the accuracy threshold (threshold not exceeded) the machine-learning model may be kept unchanged. In case the accuracy requirement is not met, further evaluating may be carried out (if the prediction accuracy does not meet the accuracy requirement). The accuracy requirement may be implemented by a threshold comparison. The accuracy threshold may be defined as an upper or lower threshold and the requirement may be defined relative to the threshold, e.g. the determined accuracy is higher or lower than the threshold. In examples equality to the threshold may be defined as meeting or not meeting the accuracy requirement.202500414 18

[0085] The further evaluating may comprise testing the machine-learning model to obtain a tested machine-learning model based on the baseline dataset in case the accuracy of the machine-learning model does not meet the accuracy requirement. This is shown in Fig. 4. The further evaluating comprises calculating a prediction accuracy of the tested machine-learning model and comparing the prediction accuracy to the accuracy requirement.

[0086] The requesting 28 of the machine-learning model adaptation is carried out if the prediction accuracy of the tested machine-learning model fails to meet the accuracy requirement based on the accuracy threshold. On the network side, the method 10 then further comprises receiving the request for machine-learning-model adaptation from the UE.

[0087] In examples different options for the model adaptation are conceivable. For example, the request is a request for retraining of the machine learning model and the method further comprises retraining the machine learning model at the UE and / or at a training node. Hence, the retraining may be carried out at the UE and the network node would then provide respective training data to the UE. At the UE side the requesting 28 may comprise requesting a retraining of the machine learning model and the method 20 may further comprise receiving retraining data from the network node and retraining the machine learning model based on the retraining data.

[0088] Another option is that the retraining of the model is carried out at the network node or any other node in the network also referred to as retraining node. For example, there may be a dedicated maintenance server in the network that maintains, monitors and controls the machine-learning models and that may also be used for retraining. From the UE perspective the adaptation of the machine-learning model may then be a download of a retrained model. In method 20 for the UE apparatus 40 the requesting 28 then comprises requesting a download of the machine-learning model or an adapted machine-learning model. Method 20 then further comprises downloading the machinelearning model or the adapted machine-learning model via the network node and installing the downloaded machine-learning model.202500414 19

[0089] The request may be a request for downloading the machine-learning model or an adapted machine-learning model. The method 10 at the network node apparatus may then further comprise forwarding the machine-learning model or the adapted machinelearning model to the UE. As outlined above, in some examples the model may just be re-downloaded, e.g. when the model at the UE was developed in a disadvantageous direction, e.g. by learning / training / adaptation algorithms at the UE.

[0090] In some examples, the network node method 10 may further comprise dynamically adjusting the accuracy threshold. The configuring then comprises dynamically signaling the accuracy threshold to the UE. Dynamic adaptation may result in dynamic model adaptation. The network may also use the accuracy threshold to coordinate model updates. By adjusting the threshold, the network may trigger model updates.

[0091] At least in some examples, the machine-learning model is a federated machine-learning model. Federated learning represents a distributed machine-learning paradigm, also adopted within 3GPP specifications, that enables the collaborative training of AI / ML models across multiple decentralized entities without requiring the exchange or centralized aggregation of raw training data. In the context of 3GPP wireless communication systems, federated learning addresses fundamental challenges related to data privacy, communication efficiency, and the practical constraints of collecting training data from geographically distributed user equipment and network nodes. The 3GPP architecture supports federated learning through the Network Data Analytics Function (NWDAF), which can coordinate training operations across multiple decentralized NWDAF instances or other participating entities.

[0092] In a federated learning configuration, user equipment performs partial training operations based on locally available data, generating local model updates that are transmitted to a central aggregation entity, which combines the contributions from multiple participants to produce an updated global model without accessing the underlying local datasets. The 5G system provides assistance for application layer federated learning operations, including candidate federated learning member selection according to specific criteria such as user equipment performance, location, trajectory, and network resource availability. The Network Exposure Function (NEF) may assist AI / ML application servers in scheduling available user equipment to participate in202500414 20

[0093] federated learning operations, enabling dynamic coordination of distributed training tasks across the network.

[0094] The 3GPP specifications distinguish between two primary variants of federated learning that differ in how data is partitioned across participating entities. Horizontal federated learning, defined in Release 18, applies to scenarios where multiple participating entities possess datasets that share the same feature space but contain different data samples. In horizontal federated learning, each participant trains a local model using its own data samples, and the resulting model updates are aggregated to produce a global model that benefits from the collective knowledge across all participants without exposing individual data samples. Vertical federated learning (VFL), which became a focus of Release 19 studies, addresses scenarios where participating entities possess datasets with different feature spaces that must be combined to jointly train a global model. In vertical federated learning, different entities contribute complementary features or measurements related to the same set of samples, enabling the training of models that leverage diverse data sources while preserving the privacy of each entity's proprietary features. The 3GPP architecture supports vertical federated learning through the NWDAF or Application Functions (AF), which may assume the role of VFL server or VFL clients. Security aspects of federated learning operations are addressed through authorization procedures for entities assuming VFL server roles and protection of information exchange during the federated learning process using Transport Layer Security (TLS). The 5G system further enables dynamic addition or removal of specific user equipment to or from AI / ML federated learning tasks, including scenarios where user equipment communicates via direct device connections, providing flexibility in the composition of federated learning participants based on evolving network conditions and application requirements.

[0095] Fig. 5 depicts another example of a flow chart of a method 10 for a network node, which is a gNB in this example. The gNB configures the threshold and provides the baseline dataset to a UE in step 502. In step 504 the gNB provides the new model parameters to the UE after request.

[0096] Fig. 6 illustrates a flowchart of another example of method 20 for a UE. The UE computes the prediction based on the AI / ML model and compares it with the configured202500414 21

[0097] threshold in step 602. If the accuracy of the prediction is better than required by the threshold in step 604 the UE proceeds in step 606 and keeps using the AI / ML model as it is. If the accuracy requirement is not met the UE re-computes the prediction with the baseline dataset in step 608. Another threshold comparison is done in step 610. If the accuracy requirement is met in step 610 the UE proceeds with using the AI / ML as it is in step 612. Otherwise, a request for model re-training / re-download is sent to the network node in step 614.

[0098] Performance monitoring constitutes a critical component of the AI / ML lifecycle management framework, serving as the primary mechanism for detecting when accumulated bias causes model performance to degrade below acceptable thresholds. In examples, this is done by evaluating the accuracy requirement defined by the accuracy threshold. In the context of 3GPP specifications for AI / ML-enabled features in the NR air interface, performance monitoring may encompass the continuous evaluation of model outputs against expected performance criteria, the computation of relevant performance metrics, the reporting of monitoring results to management entities, and the triggering of corrective actions when performance degradation is detected. The implementation of such monitoring mechanisms may require the definition of appropriate key performance indicators (KPIs) that quantify model accuracy and reliability under operational conditions.

[0099] For channel state information (CSI) prediction, relevant metrics (accuracy criteria that may be compared to the accuracy threshold) may include the mean squared error between predicted and actual CSI values, the correlation coefficient between predicted and measured channel matrices, or the resulting throughput performance when predictions are used for link adaptation. For beam management applications, monitoring metrics may encompass beam prediction accuracy, the frequency of beam misalignment events, or the overhead reduction achieved relative to exhaustive beam sweeping.

[0100] For positioning use cases, metrics such as horizontal and vertical positioning error, the cumulative distribution function of positioning accuracy, or the percentage of estimates meeting specified accuracy requirements provide quantifiable measures of model performance.202500414 22

[0101] The practical implementation of performance monitoring mechanisms in wireless communication systems may follow several architectural approaches. In a ground-truth-based monitoring approach (baseline dataset), the UE or network periodically obtains actual measurements or outcomes that can be compared against model predictions to compute error metrics directly. For example, in CSI prediction, the UE may periodically measure actual channel conditions and compare these measurements against previously generated predictions to calculate prediction error statistics. This approach provides accurate performance assessment but incurs overhead from the additional measurements required to obtain ground truth. In a reference-signal-based monitoring approach, dedicated reference signals or pilot transmissions enable the evaluation of model performance without requiring full ground-truth measurements. The network may configure specific monitoring occasions during which reduced-complexity measurements are performed to assess prediction quality. In an inference-consistency-based monitoring approach, the temporal or spatial consistency of model outputs is evaluated without explicit ground truth, wherein sudden changes in prediction patterns or statistical anomalies in output distributions may indicate performance degradation. A hybrid monitoring approach may combine elements of these methods, using lightweight consistency checks during normal operation while periodically invoking more comprehensive ground-truth-based evaluation during dedicated monitoring intervals.

[0102] The monitoring framework may further require mechanisms for aggregating performance metrics over appropriate time windows, comparing aggregated metrics against predefined thresholds, and communicating monitoring outcomes to trigger lifecycle management actions. When monitored performance falls below configured thresholds, the monitoring mechanism may initiate model switching to select an alternative model better suited to current conditions, trigger model update or fine-tuning procedures to adapt the current model, or activate fallback to legacy non-AI / ML-based operations to ensure continued reliable service. The signaling infrastructure supporting performance monitoring may utilize existing 3GPP reporting mechanisms enhanced with AI / ML-specific information elements, or may employ dedicated signaling procedures defined specifically for AI / ML lifecycle management. The balance between monitoring accuracy, signaling overhead, computational complexity, and power consumption at the UE represents a key design consideration in the implementation of202500414 23

[0103] practical performance monitoring mechanisms for AI / ML-enabled features in wireless communication systems.

[0104] Accuracy bias is a problem when it comes to Al algorithms; the best of the models may be prone to dataset biases, model training errors, etc. This may lead to an improper prediction and hence may lead to an overall operation failure or loss of service for the user.

[0105] According to some examples, a gNB configures an accuracy threshold and provides a baseline dataset to request model re-training or model re-download based on applicable conditions. The gNB configures the threshold based on the functionality support indicated by the UE. In case of federated models, the gNB may dynamically adjust the threshold for accuracy by dynamically signaling it to the UE, e.g. in a RRC (Radio Resource Control) Reconfiguration message. The UE upon receiving the accuracy threshold shall compare the computed prediction results with the computed result.

[0106] The same procedure can be followed for gNB side model. Also, for a UE side of a two-sided model, the UE shall compare the result of Al model with the configured threshold. If the prediction accuracy is lower than the configured threshold, the UE shall recompute the model with the baseline dataset. Else, it continues the operation with the prediction result. If the model accuracy is still below threshold after computing the prediction with baseline dataset, the UE shall initiate the model retraining request or model re-download request.

[0107] The UE shall indicate to the gNB as follows:

[0108] The UE may report the un-usability of the model to the gNB and request model re-training / re-download. For models to be downloaded via an OTT (Over-The-Top) server, the UE may initiate the download after reporting model unavailability and disabling the functionality.

[0109] The UE may signal this via measurement report, e.g., as specified for RRC etc. After receiving the UE indication for model unavailability, the gNB shall update the model ID and re-configure the UE model parameters.202500414 24

[0110] For example, OTT in the 3GPP context refers to applications and services that are delivered over the internet, bypassing traditional network operator services and infrastructure. An OTT server is the application server that hosts and delivers these services.

[0111] The aspects and features described in relation to a particular one of the previous examples may also be combined with one or more of the further examples to replace an identical or similar feature of that further example or to additionally introduce the features into the further example.

[0112] Examples may further be or relate to a (computer) program including a program code to execute one or more of the above methods when the program is executed on a computer, processor or other programmable hardware component. Thus, steps, operations or processes of different ones of the methods described above may also be executed by programmed computers, processors or other programmable hardware components. Examples may also cover program storage devices, such as digital data storage media, which are machine-, processor- or computer-readable and encode and / or contain machine-executable, processor-executable or computer-executable programs and instructions. Program storage devices may include or be digital storage devices, magnetic storage media such as magnetic disks and magnetic tapes, hard disk drives, or optically readable digital data storage media, for example. Other examples may also include computers, processors, control units, (field) programmable logic arrays ((F)PLAs), (field) programmable gate arrays ((F)PGAs), graphics processor units (GPUs), application-specific integrated circuits (ASICs), integrated circuits (ICs) or system-on-a-chip (SoC) systems programmed to execute the steps of the methods described above.

[0113] It is further understood that the disclosure of several steps, processes, operations or functions disclosed in the description or claims shall not be construed to imply that these operations are necessarily dependent on the order described, unless explicitly stated in the individual case or necessary for technical reasons. Therefore, the previous description does not limit the execution of several steps or functions to a certain order. Furthermore, in further examples, a single step, function, process or operation may202500414 25

[0114] include and / or be broken up into several sub-steps, -functions, -processes or -operations.

[0115] If some aspects have been described in relation to a device or system, these aspects should also be understood as a description of the corresponding method. For example, a block, device or functional aspect of the device or system may correspond to a feature, such as a method step, of the corresponding method. Accordingly, aspects described in relation to a method shall also be understood as a description of a corresponding block, a corresponding element, a property or a functional feature of a corresponding device or a corresponding system.

[0116] The following claims are hereby incorporated in the detailed description, wherein each claim may stand on its own as a separate example. It should also be noted that although in the claims a dependent claim refers to a particular combination with one or more other claims, other examples may also include a combination of the dependent claim with the subject matter of any other dependent or independent claim. Such combinations are hereby explicitly proposed, unless it is stated in the individual case that a particular combination is not intended. Furthermore, features of a claim should also be included for any other independent claim, even if that claim is not directly defined as dependent on that independent claim.202500414 26

[0117] Reference numerals

[0118] 10 method for network node

[0119] 12 providing baseline dataset

[0120] 14 configuring accuracy threshold

[0121] 20 method for UE

[0122] 22 receiving baseline data set an indication 24 receiving information about a duration

[0123] 26 evaluating machine-learning model

[0124] 28 requesting adaptation

[0125] 30 apparatus for network node

[0126] 32 one or more interfaces

[0127] 34 one or more processing devices

[0128] 40 apparatus for UE

[0129] 42 one or more interfaces

[0130] 44 one or more processing devices

[0131] 300 network node

[0132] 400 UE

[0133] 500 mobile communication system

[0134] 502 configuration

[0135] 504 provision

[0136] 602 computing prediction

[0137] 604 evaluation

[0138] 606 proceed without adaptation

[0139] 608 re-computing

[0140] 610 evaluations

[0141] 612 proceed without adaptation

[0142] 614 request for model re-training

Claims

202500414 1ClaimsWhat is claimed is:

1. A method (10) for a network node (300) of a mobile communication system (500), the method (10) comprisingproviding (12) a baseline dataset to user equipment (400), UE, the base line dataset forming a baseline for a machine-learning model that is used at the UE (400) for parameter prediction;configuring (14) an accuracy threshold at the UE (400), the accuracy threshold allowing to determine a need for adaptation of the machine-learning model used at the UE (400) in case a prediction accuracy fails to meet an accuracy requirement, which is based on the accuracy threshold.

2. The method (10) of claim 1 , further comprising receiving a request for machine-learning-model adaptation from the UE (400).

3. The method (10) of claim 2, wherein the request is a request for re-training of the machine learning model and wherein the method (10) further comprises re-training the machine-learning model at the UE (400) and / or at a training node.

4. The method (10) of claim 2, wherein the request is a request for downloading the machine-learning model or an adapted machine learning model and wherein the method further comprises forwarding the machine-learning model or the adapted machine learning model to the UE (400).

5. The method (10) of one of the claims 1 to 4, further comprising dynamically adjusting the accuracy threshold and wherein the configuring comprises dynamically signaling the accuracy threshold to the UE (400).

6. The method (10) of one of the claims 1 to 5, wherein the machine-learning model is a federated machine-learning model.202500414 27. A method (20) for user equipment(400), UE, of a mobile communication system (500), the method (20) comprisingreceiving (22) a baseline dataset from a network node of the mobile communication system (500), the baseline dataset forming a baseline for a machine-learning model that is used at the UE (400) for parameter prediction;receiving (24) a configuration for an accuracy threshold at the UE (400) from the network node (300), the accuracy threshold allowing to determine a need for adaptation of the machine-learning model used at the UE (400) in case a prediction accuracy fails to meet an accuracy requirement based on the accuracy threshold;evaluating (26) the machine-learning model using the accuracy threshold; andrequesting (28) machine-learning model adaptation if the evaluating indicates that an accuracy of the machine-learning model fails to meet the accuracy requirement based on the accuracy threshold.

8. The method (20) of claim 7, wherein the requesting (28) comprises requesting a re-training of the machine-learning model and wherein the method (20) further comprises receiving re-training data from the network node (300) and re-training the machine-learning model based on the re-training data.

9. The method (20) of claim 7, wherein the requesting (28) comprises requesting a download of the machine-learning model or an adapted machine learning model, and wherein the method (209 further comprises downloading the machine-learning model or the adapted machine learning model via the network node (300) and installing the downloaded machine-learning model.

10. The method (20) of one of the claims 7 to 9, wherein the evaluating (26) comprises calculating a prediction accuracy of the machine-learning model and further evaluating if the prediction accuracy does not meet the accuracyrequirement.202500414 311. The method (20) of claim 10, wherein the further evaluating comprises testing the machine-learning model to obtain a tested machine-learning model based on the baseline data in case the accuracy of the machine-learning model does not meet the accuracy requirement.

12. The method (20) of claim 11 , wherein the further evaluating comprises calculating a prediction accuracy of the tested machine-learning model and comparing the prediction accuracy to the accuracy requirement.

13. The method (20) of claim 12, wherein the requesting (28) of the machinelearning model adaptation is carried out if the prediction accuracy of the tested machine-learning model fails to meet the accuracy requirement based on the accuracy threshold.

14. A computer program having a program code for performing one of the methods (10; 20) of claims 1 to 13, when the computer program is executed on a computer, a processor, or a programmable hardware component.

15. An apparatus (30) for a network node (300) of a mobile communication system (500), the apparatus (30) comprisesone or more interfaces (32) configured to communicate in the mobile communication system (500); andone or more processing devices (34) configured to perform one of the methods (10) of claims 1 to 7.

16. An apparatus (40) for user equipment (400), UE, of a mobile communication system (500), the apparatus (40) comprisesone or more interfaces (42) configured to communicate in the mobile communication system (500); and202500414 4one or more processing devices (44) configured to perform one of the methods (20) of claims 8 to 13.

17. A network node (300) of a mobile communication system (500) comprising the apparatus (30) of claim 15.

18. User equipment (400), UE, of a mobile communication system (500) comprising the apparatus (40) of claim 16.

19. A mobile communication system (500) comprising the network node (300) of claim 17 and the UE (400) of claim 18.

Images