Methods and devices for evaluating ai / ML inference model applicability

A prediction model at the UE assesses inference model applicability, preventing wastage by ensuring only applicable models are executed, thus optimizing resource use and data accuracy in wireless communication systems.

WO2026234156A1PCT designated stage Publication Date: 2026-11-12CONTINENTAL AUTOMOTIVE TECHNOLOGIES GMBH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/EP2026/063218
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-08
Filing Date
2026-05-07
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing wireless communication systems face inefficiencies due to AI/ML models being applied in contexts where they are not applicable, leading to resource wastage and inaccurate data reporting, which can trigger unnecessary network adjustments.

Method used

Implementing a prediction model at the UE to assess the applicability of inference models based on UE context, preventing execution of inapplicable models and reducing resource consumption.

Benefits of technology

Enables the network to avoid wasting resources by knowing whether an inference model is applicable at the UE, preserving hardware and energy resources, and ensuring accurate data reporting.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure EP2026063218_12112026_PF_FP_ABST
    Figure EP2026063218_12112026_PF_FP_ABST
Patent Text Reader

Abstract

The present disclosure relates to a method (40) for communicating in a wireless communication system, the method being implemented by a user equipment, UE (20), of the wireless communication system, wherein the UE comprises a communication unit configured to exchange data with a radio access network, RAN, of the wireless communication system, wherein the exchanged data comprises data related to an inference model to be executed by the UE, the method comprises: - (S40) obtaining a prediction model, wherein the prediction model is configured to output, based on input data which includes information representative of a UE context, a prediction result indicating whether the inference model is applicable, - (S41) executing the prediction model to obtain a prediction result, - in response to the obtained prediction result indicating that the inference model is not applicable: (S42) transmitting all or part of the obtained prediction result to the RAN.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] 202502592 1

[0002] DESCRIPTION TITLE: Methods and devices for evaluating AI / ML inference model applicability

[0003] TECHNICAL FIELD

[0004] The present disclosure relates to wireless communication systems and relates more specifically to methods and devices for enabling efficient usage of artificial intelligence (Al) / machine learning (ML) techniques in wireless communication systems.

[0005] BACKGROUND

[0006] Wireless communication systems are widely deployed to provide various telecommunication services such as telephony, video, data, messaging, broadcasts, or other similar types of services. Although wireless communication systems have made great technological advancements over many years, challenges still exist, and there still exists a need for further improvements in wireless communications systems.

[0007] For example, 3GPP (Third Generation Partnership Project) is currently investigating AI / ML techniques in order to improve the performance of the air interface, among others.

[0008] More specifically, 3GPP is currently investigating several use cases in which AI / ML techniques are expected to improve the performance of the air interface. For example, the technical report TR 38.843 V18.0.0 discusses the following use cases:

[0009] channel state information (CSI) feedback enhancement: for example, spatial-frequency domain CSI compression using two-sided Al model (or possibly only a UE-sided Al model), time domain CSI prediction using UE-sided model, etc.,

[0010] beam management: for example, spatial-domain downlink (DL) beam prediction for Set A of beams based on measurement results of Set B of beams, temporal DL beam prediction for Set A of beams based on the historic measurement results of Set B of beams, etc.,

[0011] positioning accuracy enhancements: for example, direct AI / ML positioning, AI / ML assisted positioning, etc.

[0012] In some cases, a UE-sided AI / ML model may be used. When a UE performs model inference (i.e. , it executes the UE-sided AI / ML model), it can report the inference data to the network. This process is part of the broader framework for AI / ML model lifecycle management, which includes model training, deployment, inference, monitoring, and updating. In the technical report TR 38.843 V18.0.0, the signaling procedures for model inference reporting involve several key phases:

[0013] model inference initiation: the UE initiates the model inference process based on predefined triggers or conditions,

[0014] inference execution: the UE performs the inference using the deployed AI / ML model, reporting results: the UE reports the inference data to the network; this can include various metrics and outcomes derived from the AI / ML model,202502592 2

[0015] network actions: based on the reported inference data, the network may take specific actions, such as adjusting resource allocations or optimizing network parameters.

[0016] However, AI / ML models for a given use case may be tailored towards and applicable to specific scenarios, locations, configuration, deployments, among other factors. In this regard, it is acknowledged that the execution of a given AI / ML model may fail to produce inference data, or it may produce inaccurate inference data, if performed in a context for which the AI / ML model is not applicable. In such a case, the network may fail to receive inference data, or it may receive inaccurate inference data. In any case, this will trigger actions from the network which will waste resources by repeatedly pushing inapplicable AI / ML models.

[0017] SUMMARY

[0018] The present disclosure aims at improving the situation. In particular, the present disclosure aims to address at least some of the limitations of the prior art discussed above. In particular, the present disclosure aims to propose a solution for enabling the network to know whether a given inference model is applicable at the UE, in order to avoid wasting resources at the network and / or at the UE.

[0019] Also, the present disclosure aims at enabling, in some embodiments at least, the network to get knowledge of the reasons why a given inference model is not applicable.

[0020] For that purpose, it is proposed to use a prediction model at the UE, associated with an inference model, which can be executed to predict applicability of the inference model. By “prediction”, we mean determining for a future time, referred to as “prediction time”, which corresponds to a future inference model execution time. Hence, assuming that the inference model is to be executed at this prediction time, the prediction model evaluates whether the inference model will be applicable, given at least the expected UE context at the prediction time. By “inference model not applicable” at the prediction time, we mean that:

[0021] the inference model cannot be executed (i.e., the inference cannot be performed) at the prediction time due to, e.g., a lack of hardware and / or energy resources, a lack of input data to be processed, etc.,

[0022] the inference model can be executed at the prediction time, but the inference data is expected to be irrelevant (e.g., low confidence score) due to, e.g., the inference model not being adapted to the expected environment of the UE at the prediction time, a lack of input data to be processed at the prediction time, etc.

[0023] When the inference model is considered inapplicable by the prediction model, this is reported by the UE to the network. Hence, the network is made aware of the fact that the inference model is not applicable, such that it will not waste resources to repeatedly push for the execution of the inference model. The UE may also benefit from the determination made by the prediction model, since the execution of the inapplicable inference model may be prevented, thereby preserving hardware and energy resources of the UE.

[0024] According to a first aspect, the present disclosure relates to a method for202502592 3

[0025] communicating in a wireless communication system, the method being implemented by a user equipment, UE, of the wireless communication system, wherein the UE comprises a communication unit configured to exchange data with a radio access network, RAN, of the wireless communication system, wherein the exchanged data comprises data related to an inference model to be executed by the UE, the method comprises:

[0026] obtaining a prediction model, wherein the prediction model is configured to output, based on input data which includes information representative of a UE context, a prediction result indicating whether the inference model is applicable,

[0027] executing the prediction model to obtain a prediction result,

[0028] in response to the obtained prediction result indicating that the inference model is not applicable: transmitting the obtained prediction result to the RAN.

[0029] In some embodiments, the method according to the first aspect can further comprise one or more of the following optional features, considered either alone or in any technically possible combination.

[0030] In some embodiments of the method according to the first aspect, in response to the prediction result indicating that the inference model is not applicable, the inference model is not executed and / or inference data obtained by executing the inference model is not transmitted to the RAN.

[0031] In some embodiments of the method according to the first aspect, the information representative of the UE context includes information on usage of hardware and / or energy resources of the UE and / or information on data generation at the UE and / or information on a variation of an environment of the UE.

[0032] In some embodiments of the method according to the first aspect, the information on the usage of hardware and / or energy resources of the UE includes at least one among:

[0033] information on computing resources usage at the UE,

[0034] information on battery usage at the UE,

[0035] information on memory usage at the UE.

[0036] In some embodiments of the method according to the first aspect, the input data of the prediction model further includes information on the inference model.

[0037] In some embodiments of the method according to the first aspect, the information on the inference model includes at least one among:

[0038] information on usage policy of the inference model,

[0039] information on a type of traffic associated with the inference model,

[0040] information on a size of the inference model,

[0041] information on expiration of the inference model,

[0042] information on an execution time of the inference model,

[0043] information on an amount of input data to be processed by the inference model.

[0044] In some embodiments of the method according to the first aspect, all or part of the202502592 4

[0045] information on the inference model is provided by the RAN.

[0046] In some embodiments of the method according to the first aspect, in response to the prediction result indicating that the inference model is applicable, the inference model is executed and / or inference data obtained by executing the inference model is transmitted to the RAN.

[0047] In some embodiments of the method according to the first aspect, the information representative of the prediction result is also transmitted to the RAN when the prediction result indicates that the inference model is applicable.

[0048] In some embodiments of the method according to the first aspect, the prediction model is a machine learning model.

[0049] In some embodiments of the method according to the first aspect, the inference model is a machine learning model.

[0050] In some embodiments of the method according to the first aspect, the execution of the prediction model uses fewer hardware resources than the execution of the inference model.

[0051] In some embodiments of the method according to the first aspect, the inference model is used for at least one among:

[0052] channel state information, CSI, compression,

[0053] CSI prediction,

[0054] downlink, DL, beam prediction,

[0055] UE positioning.

[0056] In some embodiments of the method according to the first aspect, the obtained prediction result is transmitted to the RAN by using layer 1, layer 2 or layer 3 signaling.

[0057] In some embodiments of the method according to the first aspect, the prediction model is executed before using the inference model, or recurrently.

[0058] In some embodiments of the method according to the first aspect, when the prediction result indicates that the inference model is not applicable, the UE transmits to the RAN an indication of inference model inapplicability reasons.

[0059] According to a second aspect, the present disclosure relates to a user equipment comprising at least one memory and at least one processor configured to carry out a method according to any one of the embodiments of the first aspect.

[0060] According to a third aspect, the present disclosure relates to a method for communicating in a wireless communication system, the method being implemented by a base station, BS, of a radio access network, RAN, of the wireless communication system, wherein the BS is configured to exchange data with a user equipment, UE, wherein the exchanged data comprises data related to an inference model to be executed by the UE, the method comprises:

[0061] receiving from the UE a prediction result indicating whether the inference model is applicable,

[0062] configuring the inference model at the UE based on the received prediction results.202502592 5

[0063] In some embodiments, the method according to the third aspect can further comprise one or more of the following optional features, considered either alone or in any technically possible combination.

[0064] In some embodiments of the method according to the third aspect, the BS stores successive prediction results received from the UE to predict future applicability of the inference model.

[0065] In some embodiments, the method according to the third aspect comprises a prior transmission, to the UE, of configuration data for configuring a prediction model at the UE, wherein the prediction model is used by the UE to obtain prediction results.

[0066] According to a fourth aspect, the present disclosure relates to a base station comprising at least one memory and at least one processor configured to carry out a method according to any one of the embodiments of the third aspect.

[0067] According to a fifth aspect, the present disclosure relates to a wireless communication system comprising at least one base station according to any one of the embodiments of the present disclosure and at least one user equipment according to any one of the embodiments of the present disclosure.

[0068] According to a sixth aspect, the present disclosure relates to a computer program product comprising instructions which, when executed by at least one processor, configure said at least one processor to carry out a method for communicating according to any one of the embodiments of the present disclosure. The computer program product can use any programming language, and can be in the form of source code, object code, or in any intermediate form between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0069] According to a seventh aspect, the present disclosure relates to a (non-transitory) computer-readable storage medium comprising instructions which, when executed by at least one processor, configure said at least one processor to carry out a method for communicating according to any one of the embodiments of the present disclosure.

[0070] BRIEF DESCRIPTION OF FIGURES

[0071] The invention will be better understood upon reading the following description, given as an example that is in no way limiting, and made in reference to the figures which show:

[0072] Figure 1: a schematic representation of a wireless communication system,

[0073] Figure 2: a schematic representation of an example of a UE,

[0074] Figure 3: a schematic representation of an example of a BS,

[0075] Figures 4 and 5: flow charts illustrating examples of methods for communicating implemented by a UE and a BS, respectively.

[0076] In these figures, references identical from one figure to another designate identical or analogous elements. For reasons of clarity, the elements shown are not to scale, unless explicitly stated otherwise.202502592 6

[0077] Also, the order of steps or phases of any methods disclosed herein, in particular in the figures, is provided for illustration purposes only and is not meant to limit the present disclosure which may be applied with the same steps or phases executed in a different order and / or with all or part of the steps or phases executed in parallel and / or jointly.

[0078] DETAILED DESCRIPTION

[0079] The detailed description set forth below, with reference to the figures, is intended as a description of various configurations and is not intended to represent the only configurations in which the concepts described herein may be practiced. The detailed description includes specific details for the purpose of providing a thorough understanding of the various concepts. However, it will be apparent to those skilled in the art that these concepts may be practiced without these specific details. For instance, although 3GPP terminology, from e.g., 5G NR, may be used in this disclosure to exemplify embodiments herein, this should not be seen as limiting the scope of the present disclosure. In particular, the present disclosure can be applied to future wireless communication systems such as 6G wireless communication systems.

[0080] Generally, all terms used herein are to be interpreted according to their ordinary meaning in the relevant technical field, unless a different meaning is clearly given and / or is implied from the context in which it is used. All references to a / an / the element, apparatus, component, means, step, etc. are to be interpreted openly as referring to at least one instance of the element, apparatus, component, means, step, etc., unless explicitly stated otherwise. Also, the order of steps of any methods disclosed herein, in particular in the figures, is provided only for illustration purposes and is not meant to limit the present disclosure which may be applied with the same steps executed in a different order and / or with all or part of the steps executed in parallel or jointly, unless a step is explicitly described as necessarily following or preceding another step and / or where it is implicit that a step must follow or precede another step. Also, in a figure, steps represented surrounded by a dashed line are to be considered as optional for the embodiment represented in this figure. Any feature of any of the embodiments disclosed herein may be applied to any other embodiment, wherever appropriate. Likewise, any advantage of any of the embodiments may apply to any other embodiments, and vice versa.

[0081] Figure 1 represents schematically an example of wireless communication system, which may be for example a 5G NR wireless communication system. More specifically, figure 1 represents a RAN of the wireless communication system, which is used to exchange data with UEs 20 via radio signals. For example, the RAN may send data to the UEs 20 (downlink, DL), for instance data received from a core network (CN, not represented in the figures). The RAN may also receive data from the UEs 20 (uplink, UL), which data may be forwarded to the CN.

[0082] In the example illustrated by figure 1, the RAN comprises one base station, BS, 30. Of course, the RAN may comprise more than one BS 30 to increase the radio coverage of the wireless communication system. Each of these BSs may be referred to as NB, eNodeB (or eNB), gNodeB (or gNB, in the case of a 5G NR wireless communication system), an access point or202502592 7

[0083] the like, depending on the wireless communication standard(s) implemented.

[0084] Figure 2 represents schematically an example of a wireless device 25 suitable for implementing any method, discussed in the present disclosure, performed at a UE 20. Basically, the wireless device 25 corresponds to an apparatus that provides wireless connectivity with the RAN of the wireless communication system, and that can be used to exchange data with said RAN. Such a wireless device 25 may be included in a UE 20, as illustrated by figure 2. The UE 20 may for instance be a cellular phone, a wireless modem, a handheld device, a laptop computer, a headset, smart glasses, a vehicle (manned or unmanned), etc., or any other equipment that may execute applications that need to exchange data with remote recipients, via the wireless device 25.

[0085] As illustrated by figure 2, the wireless device 25 comprises one or more processors 250 and one or more memories 251. The one or more processors 250 may include for instance a central processing unit (CPU), a digital signal processor (DSP), a graphics processing unit (GPU), a neural processing unit (NPU), a field-programmable gate array (FPGA), an application specific integrated circuit (ASIC), etc. The one or more memories 251 may include any type of computer readable volatile and non-volatile memories (magnetic hard disk, solid-state disk, optical disk, electronic memory, etc.). The one or more memories 251 may store a computer program product 252, in the form of a set of program-code instructions to be executed by the one or more processors 250 to implement all or part of the steps of a method 40 for communicating, performed at a UE 20, according to any one of the embodiments disclosed herein.

[0086] As illustrated by figure 2, the wireless device 25 comprises also a wireless communication unit 253 adapted to exchange data with BSs 30 of the RAN using radio signals. The wireless communication unit 253 may implement one or more wireless communication protocols, and may for instance be a 3G, 4G, 5G, 6G, WiFi, WiMax, etc. transceiver or the like. In some examples, the wireless communication unit 253 comprises a 5G NR wireless communication unit.

[0087] Figure 3 represents schematically an example of a BS 30 suitable to implement any method, discussed in the present disclosure, performed by the RAN.

[0088] As illustrated by figure 3, the BS 30 comprises one or more processors 300 and one or more memories 301. The one or more processors 300 may include for instance a CPU, a DSP, a GPU, an NPU, an FPGA, an ASIC, etc. The one or more memories 301 may include any type of computer readable volatile and non-volatile memories (magnetic hard disk, solid-state disk, optical disk, electronic memory, etc.). The one or more memories 301 may store a computer program product 302, in the form of a set of program-code instructions to be executed by the one or more processors 300 to implement all or part of the steps of a method 50 for communicating, performed at the RAN’s side, according to any one of the embodiments disclosed herein.202502592 8

[0089] As illustrated by figure 3, the BS 30 comprises also a wireless communication unit 303, configured to exchange data with UEs 20 using radio signals, and more specifically with wireless communication units 253 of wireless devices 25 included in these UEs 20. The wireless communication unit 303 may for instance be a 3G, 4G, 5G, 6G, WiFi, WiMax, etc. transceiver or the like. In some examples, the wireless communication unit 303 of the BS 30 comprises a 5G NR transceiver.

[0090] As illustrated by figure 3, the BS 30 may comprise also, in some examples, a network communication unit 304, configured to exchange data with other base stations of the RAN and / or with the CN. The network communication unit 304 may support one or more suitable communication protocols, which may be wired (including optical) and / or wireless.

[0091] As discussed above, the present disclosure aims at proposing a solution for enabling the RAN to know whether a given inference model is applicable at the UE 20, in order to avoid wasting resources at the RAN and / or at the UE 20.

[0092] For that purpose, it is proposed to use a prediction model at the UE 20, associated with an inference model, which can be executed to predict applicability of the inference model. By “prediction”, we mean determining for a future time, referred to as “prediction time”, which corresponds to a future inference model execution time. Hence, assuming that the inference model is to be executed at this prediction time, the prediction model evaluates whether the inference model will be applicable, given at least the expected UE context at the prediction time. By “inference model not applicable” at the prediction time, we mean that:

[0093] the inference model cannot be executed (i.e., the inference cannot be performed) at the prediction time due to, e.g., a lack of hardware and / or energy resources, a lack of input data to be processed, etc.,

[0094] the inference model can be executed at the prediction time, but the inference data is expected to be irrelevant (e.g., low confidence score) due to, e.g., the inference model not being adapted to the expected environment of the UE 20 at the prediction time, a lack of input data to be processed at the prediction time, etc.

[0095] When the inference model is considered inapplicable by the prediction model, this is reported by the UE 20 to the RAN. Hence, the network is made aware of the fact that the inference model is not applicable, such that it will not waste resources to repeatedly push for the execution of the inference model. The UE 20 may also benefit from the determination made by the prediction model, since the execution of the inapplicable inference model may be prevented, thereby preserving hardware and energy resources of the UE 20.

[0096] It should be noted that the present disclosure may be applied to any type of inference model used at a UE 20 to perform inference (i.e., to produce inference data that may be transmitted to the RAN), and the choice of a specific type of inference model corresponds to a specific but non-limitative embodiment of the present disclosure. In some embodiments, the inference model is a machine learning model, for example previously trained to perform a202502592 9

[0097] specific type of inference. Again, the present disclosure is not limited to a specific type of inference and may be applied to many different inference types. For example, the inference model may be used to perform inference for any one of the use cases discussed in the technical report TR 38.843 V18.0.0, e.g.:

[0098] CSI compression,

[0099] CSI prediction,

[0100] DL beam prediction,

[0101] UE positioning, etc.

[0102] We now present non-limitative examples of methods for communicating in a wireless communication system, and more specifically for allocating UL communication resources.

[0103] Figure 4 represents a diagram showing the main steps of an exemplary embodiment of a method 40 for communicating, which is implemented by a UE 20. Figure 5 represents a diagram showing the main steps of an exemplary embodiment of a method 50 for communicating, which may be implemented by a BS 30 of the RAN when the UE 20 implements the method 40 for communicating illustrated by figure 4.

[0104] As illustrated by figure 4, the method 40 for communicating comprises a step S40 of obtaining a prediction model.

[0105] For example, the prediction model may be retrieved from a memory of the UE 20, where it has been previously stored, for example during manufacturing of the UE 20 (e.g., hard-coded prediction model), or via over the air (OTA) configuration, etc. In some examples, the prediction model may be provided and / or configured by the RAN. For example, the prediction model may be configured by the RAN together with the inference model.

[0106] The prediction model is configured to output a prediction result indicating whether the inference model is expected to be applicable at a prediction time which corresponds to a future (inference model execution) time.

[0107] In order to perform the prediction, the prediction model receives as input data at least information representative of a UE context. Generally, different types of parameters may be considered for describing the UE context, and the choice of a specific type of UE context parameter corresponds to a specific but non-limitative embodiment of the present disclosure. Also, for a given UE context parameter, it is possible to provide one or more values associated with one or more past and / or present times. In other words, for a given UE context parameter, it is possible to consider a single value (e.g., a current value of the UE context parameter) or a plurality of successive past and present values of the UE context parameter (e.g., temporal profile of the UE context parameter).

[0108] For example, the information representative of the UE context may include information on usage of hardware and / or energy resources of the UE 20.

[0109] For example, the information on usage of hardware resources may include information on the usage of computing resources of the UE 20. For example, the information on the usage202502592 10

[0110] of computing resources may correspond to a CPU load, or to a temporal profile of the CPU load, which may be used by the prediction model to predict the CPU load at the prediction time (future inference model execution time). If the CPU load predicted for the prediction time is too high for allowing the inference model to be executed, said inference model can be considered to be not applicable at the prediction time. Alternatively, or in combination thereof, the information on the usage of computing resources may correspond to temporal execution profiles of other applications executed by the UE 20, which may be used by the prediction model to predict whether many other applications are scheduled to be executed at the prediction time and may prevent the execution of the inference model, etc.

[0111] Alternatively, or in combination thereof, the information on usage of hardware resources may include information on the usage of memory of the UE 20. As for the usage of computing resources discussed above, the inference model may be inapplicable if the predicted amount of memory available at the prediction time is insufficient to execute said inference model.

[0112] For example, the information on usage of energy resources may include information on battery usage at the UE 20. For example, the information on battery usage may correspond to a battery level (state of charge, SOC), which may be used by the prediction model to predict the battery level at the prediction time (future inference model execution time). If the battery level predicted for the prediction time is too low for allowing the inference model to be executed, said inference model can be considered to be not applicable at the prediction time.

[0113] Alternatively, or in combination thereof, the information representative of the UE context may include information on data generation at the UE 20. For example, the execution of the inference model may require that a specific type of data (for example associated with a specific application or traffic type or network slice), required as input data by the inference model, is available at the prediction time in a buffer of the UE 20. Alternatively, or in combination thereof, the execution of the inference model may require that a specific amount of data, required as input data by the inference model, is available at the prediction time in a buffer of the UE 20. In such cases, the information on data generation at the UE 20 may include one or more temporal data generation profiles (for example one for each specific type of data generated at the UE 20), which may be used by the prediction model to predict whether the required input data will be available at the prediction time (future inference model execution time). If the required input data is not available at the prediction time (e.g., specific type of data missing or insufficient), said inference model can be considered to be not applicable at the prediction time.

[0114] Alternatively, or in combination thereof, the information representative of the UE context may include information on the variation of the environment of the UE 20. For example, the information on the variation of the environment of the UE 20 may correspond to the speed of the UE 20, or to the speed of objects relative to the UE 20, which may be used by the prediction model to predict whether the environment will have changed at the prediction time (future202502592 11

[0115] inference model execution time). Indeed, the inference model may be tailored to a specific environment of the UE 20 such that, if the environment changes significantly, said inference model may become inapplicable to the environment experienced at the prediction time.

[0116] In some examples, the input data of the prediction model further includes information (e.g., metadata) on the inference model. All or part of such information on the inference model may be provided to the UE 20 by the RAN.

[0117] In some examples, the information on the inference model may comprise information on the size of the inference model (e.g., number of parameters or memory footprint of the inference model) and / or information on the execution time (i.e. , duration) of the inference model. Such size information and / or execution time information may be used by the prediction model, in combination with e.g. information on usage of hardware and / or energy resources of the UE 20, in order to predict whether the hardware and / or energy resources will be sufficient to execute the inference model at the prediction time. If the hardware and / or energy resources are not sufficient, said inference model can be considered to be not applicable at the prediction time.

[0118] Alternatively, or in combination thereof, the information on the inference model may comprise information on a type of traffic (e.g., application type, network slice, etc.) associated with the inference model and / or information on an amount of input data required to be processed by the inference model. Such traffic type information and / or amount of input data information may be used by the prediction model, in combination with e.g. information on data generation at the UE 20, in order to predict whether the data generated at the UE 20 will satisfy the traffic type and / or amount of input data requirements at the prediction time. If the data generated until the prediction time is not expected to satisfy the traffic type and / or amount of input data requirements, said inference model can be considered to be not applicable at the prediction time.

[0119] Alternatively, or in combination thereof, the information on the inference model may comprise information on the expiration of the inference model. Indeed, an inference model may be configured to expire at some point, and the expiration may be defined by e.g. an expiration date, a maximum duration during which it may be executed, a maximum number of times it may be executed, etc. Such expiration information may be used by the prediction model to predict whether the inference model will have expired at the prediction time. If the inference model is set to expire before the prediction time, said inference model can be considered to be not applicable at the prediction time.

[0120] Alternatively, or in combination thereof, the information on the inference model may comprise information on usage policy of the inference model. For example, the usage policy may define privacy requirements. Hence, if the privacy requirements are not expected to be satisfied at the prediction time, the inference model can be considered to be not applicable.

[0121] It should be noted that the present disclosure may be applied to any type of prediction model suitable for performing predictions based on past and present input data, and the choice202502592 12

[0122] of a specific type of prediction model corresponds to a specific but non-limitative embodiment of the present disclosure. In some embodiments, the prediction model is a regression model or a time-series model. For example, the prediction model corresponds to a machine learning model, for example previously trained to predict inference model applicability. Such a machine learning model may be trained via supervised learning by using a training dataset which comprises reference (ground truth) prediction results associated with respective training input data.

[0123] In preferred embodiments, the prediction model is configured such that the execution of the prediction model uses fewer hardware resources than the execution of the inference model. In other words, it is computationally less expensive to execute the prediction model than the inference model, and the prediction model is a lightweight model compared to the inference model. This is advantageous in that the applicability of the inference model can be evaluated at a lower computational cost than the execution of the inference model, such that resources can be saved when the inference model is predicted to be not applicable.

[0124] As illustrated by figure 4, the method 40 for communicating comprises a step S41 of executing the prediction model to obtain a prediction result for a given prediction time.

[0125] If the obtained prediction result indicates that the inference model is not applicable at the prediction time (reference S41a in figure 4), the method 40 for communicating comprises a step S42 of transmitting all or part of the obtained prediction result to the RAN. For example, the UE 20 may notify the prediction result to the RAN as a single bit value indicating whether the inference model is applicable or not. In some examples, the UE 20 may also transmit to the RAN an indication of the inapplicability reasons (e.g., insufficient high CPU load, low battery level, inference model expired, change of environment, etc.), by using any suitable format (for example as a bitmap, wherein each bit of the bitmap represents a specific inapplicability reason).

[0126] The obtained prediction result may be transmitted by using any type of signaling. In particular, the obtained prediction result may be transmitted by using layer 1 (L1), layer 2 (L2) or layer 3 (L3) signaling.

[0127] It should be noted that, when the obtained prediction result indicates that the inference model is not applicable at the prediction time (reference S41a in figure 4), the UE 20 preferably does not execute the inference model at the prediction time (to save UE resources), and / or the inference data obtained by executing the inference model is not transmitted to the RAN (to save UE and RAN resources). However, the RAN may also benefit from inference data obtained while the inference model is considered not applicable, in order to enable the RAN to learn to detect inference model inapplicability based on inference data. Hence, it is also possible, in some cases, to transmit to the RAN the inference data obtained by executing the inference model deemed inapplicable by the prediction model.

[0128] If the obtained prediction result indicates that the inference model is applicable at the prediction time (reference S41b in figure 4), the method 40 for communicating comprises a step S43 of executing the inference model at the prediction time, thereby producing inference data202502592 13

[0129] which may be transmitted to the RAN during an optional step S44. It should be noted that the prediction result may also be transmitted to the RAN, in some cases.

[0130] It should be noted that the execution of the prediction model, during step S41, can be performed reactively (e.g., in response to a request from the RAN, or in response to the configuration or activation of the inference model by the RAN, before using said inference model) and / or proactively (e.g., the UE 20 may execute the prediction model recurrently, for example periodically, to dynamically update the applicability evaluation according to the real-conditions of the UE context and / or RAN state, and to detect when the inference model may become inapplicable). Also, the prediction model may be executed when the execution of the inference model has failed, in order to determine the reasons for this execution failure, in order to transmit said reasons to the RAN.

[0131] As discussed above, figure 5 represents a diagram showing the main steps of an exemplary embodiment of a method 50 for communicating, which may be implemented by a BS 30 when the UE 20 implements the method 40 for communicating illustrated by figure 4.

[0132] As illustrated by figure 5, the method 50 for exchanging data comprises a step S50 of receiving from the UE 20 the prediction result (obtained by the UE 20 by executing the prediction model) indicating whether the inference model is applicable.

[0133] As illustrated by figure 5, the method 50 for communicating comprises a step S51 of configuring the inference model at the UE 20 based on the received prediction result.

[0134] For example, if the received prediction result indicates that the inference model is not applicable, then the BS 30 may transmit configuration data to the UE 20, for reconfiguring the inference model at the UE 20 or for deactivating execution of the inference model at the UE. The configuration data may be determined based on the received prediction result, in particular when it receives inapplicability reasons. For example, if the inapplicability reasons indicate that the hardware and / energy resources at the UE 20 are insufficient, the BS 30 may configure another inference model the execution of which is less computationally expensive, or it may deactivate the execution of the inference model at the UE 20. If the inapplicability reasons indicate that the environment of the UE 20 has changed (e.g., mobility scenario), the BS 30 may update the inference model to account for the environment change, etc.

[0135] In cases where the BS 30 may receive a prediction result indicating that the inference model is applicable, no configuration data needs to be determined and transmitted to the UE 20 during the step S51 of configuring the inference model.

[0136] As indicated above, in some examples, the prediction model may be configured at the UE 20 by the RAN, in which case the method 50 for communicating may comprise a prior step (not represented in the figures) of transmitting configuration data to the UE 20, for configuring the prediction model at the UE 20.

[0137] As indicated above, the BS 30 may reconfigure the inference model based on the received prediction result. In some examples, the BS 30 may also store successive prediction202502592 14

[0138] results received from the UE, in order to predict future applicability of the inference model.

[0139] It is emphasized that the present disclosure is not limited to the above exemplary embodiments. Variants of the above exemplary embodiments are also within the scope of the present disclosure.

[0140] It should be noted that the following exemplary embodiments are also included in the present disclosure.

[0141] EMBODIMENT 1: A method for pre-empting AI / ML model applicability in a wireless communication network, comprising:

[0142] standardizing a predictive module on UE to pre-evaluate model applicability based on device and network context;

[0143] the predictive module considering:

[0144] o UE context such as CPU load, memory usage, battery level;

[0145] o network state such as Al model policy, gNB configuration, slice profile; o model metadata such as size, execution time estimate, hardware dependency; reporting the inference to the network for configuring the UE-side model parameters. EMBODIMENT 2: The method according to EMBODIMENT 1, wherein the predictive module is a lightweight machine learning model standardized on the UE.

[0146] EMBODIMENT 3: The method according to one of EMBODIMENTS 1 to 2, wherein the UE reports the inference before or during the model application, or reactively after model applicability failure.

[0147] EMBODIMENT 4: The method according to one of EMBODIMENTS 1 to 3, wherein the inference is signaled to the network via Layer 1 (L1), Layer 2 (L2), or Layer 3 (L3) signaling.

[0148] EMBODIMENT 5: The method according to one of EMBODIMENTS 1 to 4, wherein the UE initiates the inference upon receiving model activation signaling from the network, periodically, or as a failure fallback.

[0149] EMBODIMENT 6: The method according to one of EMBODIMENTS 1 to 5, wherein the predictive module dynamically updates its evaluation based on real-time conditions of the UE and network.

[0150] EMBODIMENT 7: The method according to one of EMBODIMENTS 1 to 6, wherein the network configures the UE-side model parameters to optimize resource allocation based on the inference received from the UE.

[0151] EMBODIMENT 8: The method according to one of EMBODIMENTS 1 to 7, further comprising the step of the network entity storing historical inference data to predict future applicability conditions.

[0152] EMBODIMENT 9: The method according to one of EMBODIMENTS 1 to 8, wherein the predictive module uses a combination of device context, network state, and model metadata to provide a comprehensive applicability evaluation.

[0153] EMBODIMENT 10: A system for pre-empting AI / ML model applicability in a wireless202502592 15

[0154] communication network, comprising:

[0155] a UE configured with a predictive module to pre-evaluate model applicability based on device and network context;

[0156] a network entity configured to receive the inference and adjust the model parameters accordingly.

[0157] EMBODIMENT 11: The system according to EMBODIMENT 10, wherein the predictive module on the UE is a lightweight machine learning model.

[0158] EMBODIMENT 12: The system according to EMBODIMENTS 10 to 11, wherein the network entity is a gNB configured to configure the UE-side model parameters based on the inference provided by the UE.

[0159] EMBODIMENT 13: The system according to EMBODIMENTS 10 to 12, wherein the UE and network entity collaborate to enhance AI / ML performance for wireless communication.

[0160] EMBODIMENT 14: The system according to EMBODIMENTS 10 to 13, wherein the UE periodically updates the network with applicability evaluations to ensure optimal model performance.

[0161] EMBODIMENT 15: A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform the method of any of EMBODIMENTS 1 to 9.

Claims

202502592 16CLAIMS1. A method (40) for communicating in a wireless communication system, the method being implemented by a user equipment, UE (20), of the wireless communication system, wherein the UE comprises a communication unit configured to exchange data with a radio access network, RAN, of the wireless communication system, wherein the exchanged data comprises data related to an inference model to be executed by the UE, the method comprises:(540) obtaining a prediction model, wherein the prediction model is configured to output, based on input data which includes information representative of a UE context, a prediction result indicating whether the inference model is applicable,(541) executing the prediction model to obtain a prediction result,in response to the obtained prediction result indicating that the inference model is not applicable: (S42) transmitting all or part of the obtained prediction result to the RAN.

2. The method (40) according to claim 1, wherein, in response to the prediction result indicating that the inference model is not applicable, the inference model is not executed and / or inference data obtained by executing the inference model is not transmitted to the RAN.

3. The method (40) according to any one of the preceding claims, wherein the information representative of the UE context includes information on usage of hardware and / or energy resources of the UE and / or information on data generation at the UE and / or information on a variation of an environment of the UE.

4. The method (40) according to claim 3, wherein the information on the usage of hardware and / or energy resources of the UE includes at least one among:information on computing resources usage at the UE,information on battery usage at the UE,information on memory usage at the UE.

5. The method (40) according to any one of the preceding claims, wherein the input data of the prediction model further includes information on the inference model.

6. The method (40) according to claim 5, wherein the information on the inference model includes at least one among:information on usage policy of the inference model,information on a type of traffic associated with the inference model,information on a size of the inference model,information on expiration of the inference model,information on an execution time of the inference model,information on an amount of input data to be processed by the inference model.

7. The method (40) according to any one of claims 5 to 6, wherein all or part of the information on the inference model is provided by the RAN.

8. The method (40) according to any one of the preceding claims, wherein, in response to202502592 17the prediction result indicating that the inference model is applicable, the inference model is executed and / or inference data obtained by executing the inference model is transmitted to the RAN.

9. The method (40) according to any one of the preceding claims, wherein the prediction model is a machine learning model.

10. The method (40) according to any one of the preceding claims, wherein the inference model is a machine learning model.

11. The method (40) according to any one of the preceding claims, wherein the execution of the prediction model uses fewer hardware resources than the execution of the inference model.

12. The method (40) according to any one of the preceding claims, wherein the inference model is used for at least one among:channel state information, CSI, compression,CSI prediction,downlink, DL, beam prediction,UE positioning.

13. The method (40) according to one of the preceding claims, wherein the obtained prediction result is transmitted to the RAN by using layer 1, layer 2 or layer 3 signaling.

14. The method (40) according to one of the preceding claims, wherein the prediction model is executed before using the inference model, or recurrently.

15. The method (40) according to one of the preceding claims, wherein, when the prediction result indicates that the inference model is not applicable, the UE transmits to the RAN an indication of inference model inapplicability reasons.

16. A user equipment (20) comprising at least one memory and at least one processor configured to carry out a method (40) according to any one of the preceding claims.

17. A method (50) for communicating in a wireless communication system, the method being implemented by a base station, BS (30), of a radio access network, RAN, of the wireless communication system, wherein the BS is configured to exchange data with a user equipment, UE (20), wherein the exchanged data comprises data related to an inference model to be executed by the UE, the method comprises:(550) receiving from the UE a prediction result indicating whether the inference model is applicable,(551) configuring the inference model at the UE based on the received prediction result.

18. The method (50) according to claim 17, wherein the BS stores successive prediction results received from the UE to predict future applicability of the inference model.

19. The method (50) according to any one of claims 17 to 18, comprising a prior transmission, to the UE, of configuration data for configuring a prediction model at the UE, wherein the prediction model is used by the UE to obtain prediction results.202502592 1820. A base station (30) comprising at least one memory and at least one processor configured to carry out a method (50) according to any one of claims 17 to 19.

21. A wireless communication system comprising at least one base station (30) according to claim 20 and at least one user equipment (20) according to claim 16.

22. A computer program product (252, 302) comprising instructions which, when executed by at least one processor, configure said at least one processor to carry out a method (40) according to any one of claims 1 to 15 or a method (50) according to any one of claims 17 to 19.