UE reporting performace of an ai / ML model

Abstract

A User Equipment (UE) (100) for reporting performance of a first Artificial Intelligence / Machine-Learning (AI / ML) model (110) executed by the UE (100) is provided. The UE (100) comprises a communications interface (120) operative to communicate with a Radio Access Network (RAN) (200, 210) via one or more radio links (150), and processing circuitry (130) causing the UE (100) to be operative to receive, from the RAN (200, 210), first configuration information for configuring execution of the first AI / ML model (110) by the UE (100), for inference of one or more properties related to the one or more radio links (150), receive, from the RAN (200, 210), second configuration information for configuring reporting, by the UE (100), of a performance metric indicative of the performance of the first AI / ML model (110) when the first AI / ML model (110) is executed by the UE (100), and in response to determining that a link quality of the one or more radio links (150) is below a lower link-quality threshold and / or above an upper link-quality threshold, signal, to the RAN, a performance metric indicating a low performance of the first AI / ML model (110).

Description

[0001] P110678W001 2025-03-2807:36

[0002] 1

[0003] UE REPORTING PERFORMACE OF AN AI / ML MODEL

[0004] Technical field

[0005] The invention relates to a User Equipment (UE) for reporting performance of a first Artificial Intelligence / Machine-Learning (AI / ML) model executed by the UE, a method of reporting performance of a first AI / ML model executed by a UE, a corresponding computer program, a corresponding computer-readable data carrier, and a corresponding data carrier signal.

[0006] Background

[0007] Artificial Intelligence (Al) and Machine Learning (ML) have been investigated as promising tools to optimize the usage of the air interface in wireless communication networks, such as 3rd Generation Partnership Project (3GPP) mobile networks. Example use cases include, e.g., using autoencoders for Channel State Information (CSI) compression to reduce the feedback overhead and improve channel prediction accuracy, using deep neural networks for classifying Line-of-Sight (LOS) and Non-LOS (NLOS) conditions to enhance positioning accuracy, using reinforcement learning for beam selection by the Radio Access Network (RAN) and / or the User Equipment (UE) to reduce the signaling overhead and beam alignment latency, and using deep reinforcement learning to learn improved precoding policies for Multiple Input Multiple Output (MIMO) transmissions.

[0008] In the 3GPP New Radio (NR) standardization work for Release 18, a Study Item (SI) on AI / ML for the NR air interface was conducted (3GPP TR 38.843, V18.0.0, 2023-12, 3rd Generation Partnership Project; Technical Specification Group Radio Access Network; Study on Artificial Intelligence (AI)ZMachine Learning (ML) for NR air interface, Release 18). It has beenP110678W001 2025-03-2807:36

[0009] 2

[0010] agreed to continue the work in Release 19, to explore the benefits of augmenting the air interface with features enabling improved support for AI / ML-based algorithms for enhanced performance and / or reduced complexity / overhead. Through studying and specifying a few selected use cases (such as CSI feedback, beam management, and positioning), the aim is to design the mechanisms to accommodate AI / ML tools into the 3GPP standard.

[0011] The AI / ML models which are discussed in 3GPP for the NR air interface can be categorized into the following two types:

[0012] - One-sided AI / ML models, which can be AI / ML models provisioned at the UE and inference is performed exclusively at the UE, or AI / ML models provisioned at the network (e.g., in the RAN, and in particular at an access node of the RAN, such as a gNB or the like), and inference is performed exclusively in the network.

[0013] - Two-sided AI / ML models, which are paired AI / ML models, i.e. , one model is provisioned in the network (e.g., in the RAN) and another part of the model is provisioned in the UE, and joint inference is performed across the UE and the network. More specifically, a first part of the inference is performed by UE and then a remaining, second part is performed by the network, or vice versa. An example of a two-sided AI / ML model is the use case of autoencoder (AE)-based CSI feedback / reporting, where an encoder (the UE part of the two-sided AE model) is executed at the UE to compress the estimated wireless channel information, and the output of the encoder (the compressed wireless channel information estimates) is reported from the UE to the network, e.g., to a gNB. The gNB then uses a decoder (the network part of the two-sided AE model) to reconstruct the estimated wireless channel information. Here the two-sided AI / ML model for the AE is composed of the encoder at the UE and the decoder at the gNB.P110678W001 2025-03-2807:36

[0014] 3

[0015] When using AI / ML tools for air-interface use cases, different levels of collaboration between the UE and the network, e.g., one or more gNBs, can be distinguished:

[0016] - No collaboration or interaction between UEs and the network. In this case, a proprietary AI / ML model operating with the existing air interface is applied at one side of the air interface, e.g., at the UEs, and life cycle management of the AI / ML model (e.g., model selection / training, model monitoring, model retraining, model update) is exclusively done by the UEs, without any information or assistance provided by the network.

[0017] - Limited collaboration between the UEs and the network for one-sided models. In this case, an AI / ML model is operating at one side of the air interface, e.g., at the UEs, but the UEs receive assistance from the network life cycle management of the AI / ML model, e.g., from the gNBs.

[0018] - Joint ML inference between the UEs and the network for two-sided models. In this case, it is assumed that the AI / ML model is split with one part located at the UE and the other part located at the network, e.g., the gNBs. Because of the joint inference by the UE and the network, the AI / ML model life cycle management involves both sides of the air interface, i.e. , both the UEs and the network (e.g., the gNBs). While the use of AI / ML tools for the air interface has the potential to improve performance compared to legacy mechanisms, AI / ML typically requires an increased computational complexity. For example, it is reported in 3GPP TR 38.843 that the AI / ML models for two-sided CSI compression used for most evaluations have a computational complexity of more than 10 mega floating-point operations (flops). The legacy mechanism on the other hand requires less than 0.2 mega flops of computational complexity (“Views on Evaluation of AI / ML for CSI feedback enhancement”, 3GPP TSG RAN WG1 #114, R1 -2307668, Toulouse, France, August 21 st-25th, 2023).Because of the increased computational complexity of AI / ML models, monitoring the performance of the AI / ML model is substantial. This applies in particular to AI / ML models which are provisioned in the UE, due to the limited computational resources and energy budget imposed by the battery of the UE. Performance monitoring can be achieved by using an intermediate performance indicator, e.g., Normalized Mean Square Error (NMSE), Squared Generalized Cosine Similarity (SGCS), etc, an eventual performance indicator, e.g., User Perceived Throughput (UPT), an expected BLock Error Rate (BLER), etc, or by monitoring the data drift in the AI / ML input and / or in the AI / ML output.

[0019] Performance monitoring can be performed either on the UE side, on the network side, or both. As an example, performance monitoring for the UE-sided CSI prediction can be done by the UE as described in section 8.14.2 of “RAN1 Chair’s Notes”, 3GPP TSG RAN WG1 #115, Chicago, USA, November 13th - November 17th, 2023.

[0020] For the two-sided CSI compression case, performance monitoring may be performed as described in section 9.2.2.2 of “RAN1 Chair’s Notes”, 3GPP TSG RAN WG1 #112, Athens, Greece, February 27th - March 3rd, 2023.

[0021] While the use of AI / ML tools for the air interface may provide performance improvements, AI / ML operations are more complex and consume more UE power and resources (e.g., memory) compared to legacy mechanisms. Hence, the UE may need, or prefer, to use legacy mechanisms (i.e. , non-AI / ML mechanisms) under certain conditions, e.g., when the gain of using AI / ML mechanisms is non-substantial, negligible, or even detrimental.

[0022] One mechanism to reduce the UE’s power consumption is to report a lower Rank Indicator (Rl) to the network, in order to the UE to be configured with the lower rank (as a result, the UE can turn off one or more of its RX chains). However, the network cannot determine, based on the Rl report, if the channel experienced by the UE is more suitable for lower-rank transmission, or if the channel would support higher-rank transmissions butP110678W001 2025-03-2807:36

[0023] 5

[0024] the UE prefers a lower rank to reduce its power consumption. Therefore, the Rl report cannot be used by the UE to indicate its preference to use legacy mechanisms instead of AI / ML-based tools. Further, the User Assistance Information (UAI) is not available for AI / ML tools.

[0025] Summary

[0026] It is an object of the invention to provide an improved alternative to the above techniques and prior art.

[0027] More specifically, it is an object of the invention to provide improved solutions for enabling a UE executing an AI / ML model which is configured by the network, in particular the RAN, to indicate its preference to use legacy mechanisms instead of the AI / ML model, or at least to use a less resource consuming configuration of the AI / ML model.

[0028] These and other objects of the invention are achieved by means of different aspects of the invention, as defined by the independent claims. Embodiments of the invention are characterized by the dependent claims.

[0029] According to a first aspect of the invention, a UE for reporting performance of a first AI / ML model is provided. The first AI / ML model is executed by the UE. The UE comprises a communications interface which is operative to communicate with a RAN via one or more radio links. The UE further comprises processing circuitry causing the UE to be operative to receive first configuration information for configuring execution of the first AI / ML model by the UE. The UE executes the first AI / ML model for inference of one or more properties related to the one or more radio links. The first configuration information is received from the RAN. The UE is further operative to receive second configuration information for configuring reporting of a performance metric indicative of the performance of the first AI / ML model. The performance metric is reported by the UE and is indicative of the performance of the first AI / ML model when the first AI / ML model isexecuted by the UE. The second configuration information is received from the RAN. The UE is further operative to signal a performance metric indicating a low performance of the first AI / ML model. The performance metric indicating a low performance of the first AI / ML model is signaled to the RAN in response to determining that a link quality of the one or more radio links is below a lower link-quality threshold and / or above an upper link-quality threshold.

[0030] According to a second aspect of the invention, a method of reporting performance of a first AI / ML model is provided. The first AI / ML model is executed by a UE. The UE communicates with a RAN via one or more radio links. The method is performed by the UE and comprises receiving first configuration information for configuring execution of the first AI / ML model by the UE. The UE executes the first AI / ML model for inference of one or more properties related to the one or more radio links. The first configuration information is received from the RAN. The method further comprises receiving second configuration information for configuring reporting of a performance metric indicative of the performance of the first AI / ML model. The performance metric is reported by the UE and is indicative of the performance of the first AI / ML model when the first AI / ML model is executed by the UE. The second configuration information is received from the RAN. The method further comprises signaling a performance metric indicating a low performance of the first AI / ML model. The performance metric indicating a low performance of the first AI / ML model is signaled to the RAN in response to determining that a link quality of the one or more radio links is below a lower link-quality threshold and / or above an upper link-quality threshold.

[0031] According to a third aspect of the invention, a computer program is provided. The computer program comprises instructions which, when the computer program is executed by one or more processors comprised in aP110678W001 2025-03-2807:36

[0032] 7

[0033] User Equipment, UE, cause the UE to carry out the method according to an embodiment of the second aspect of the invention.

[0034] According to a fourth aspect of the invention, a computer-readable data carrier is provided. The computer-readable data carrier has stored thereon the computer program according to the third aspect of the invention.

[0035] According to a fifth aspect of the invention, a data carrier signal is provided. The data carrier signal carries the computer program according to the third aspect of the invention.

[0036] The invention makes use of an understanding that the UE may indicate its preference to the RAN to use legacy mechanisms instead of an AI / ML model, or at least to use a less resource consuming configuration of the AI / ML model, by signaling a performance metric indicating a low performance of the AI / ML model.

[0037] Even though advantages of the invention have in some cases been described with reference to embodiments of the first aspect of the invention, corresponding reasoning applies to embodiments of other aspects of the invention.

[0038] Further objectives of, features of, and advantages with, the invention will become apparent when studying the following detailed disclosure, the drawings, and the appended claims. Those skilled in the art realize that different features of the invention can be combined to create embodiments other than those described in the following.

[0039] Brief description of the drawings

[0040] The above, as well as additional objects, features and advantages of the invention, will be better understood through the following illustrative and non-limiting detailed description of embodiments of the invention, with reference to the appended drawings, in which:P110678W001 2025-03-2807:36

[0041] 8

[0042] Fig. 1 schematically illustrates a UE operative to communicate with a RAN via one or more radio links, in accordance with embodiments of the invention.

[0043] Fig. 2 schematically illustrates the processing circuitry comprised in the UE, in accordance with embodiments of the invention.

[0044] Fig. 3 is a signaling diagram illustrating reporting of a performance metric indicative of a performance of a first AI / ML model executed by the UE, in accordance with embodiments of the invention.

[0045] Fig. 4 is a flowchart illustrating a method performed by a UE communicating with a RAN via one or more radio links, in accordance with embodiments of the invention.

[0046] All the figures are schematic, not necessarily to scale, and generally only show parts which are necessary in order to elucidate the invention, wherein other parts may be omitted or merely suggested.

[0047] Detailed description

[0048] The invention will now be described more fully herein after with reference to the accompanying drawings, in which certain embodiments of the invention are shown. This invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided by way of example so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0049] Although the use of AI / ML tools may be beneficial to optimize the usage of the air interface in wireless communication networks, such as 3GPP mobile networks, the comparatively large computational complexity of AI / ML models increases the requirements for computational resources and power consumption. This applies in particular to AI / ML models which are executed by the UE, either as one-sided AI / ML models which are provisioned on theP110678W001 2025-03-2807:36

[0050] 9

[0051] UE, or two-sided AI / ML models for joint inference with one part of the AI / ML model being provisioned on the UE and a corresponding network-side part of the AI / ML model which is provisioned in the RAN, e.g., in a gNB.

[0052] Embodiments of the invention enable the UE to signal its preference to not be configured for executing an AI / ML model, or to be configured with an AI / ML model which requires less computational and / or power recourses, to the network (i.e. , the RAN 200). This may be achieved by relying on signaling which is intended for AI / ML performance monitoring, e.g., as described in “RAN1 Chair’s Notes”, 3GPP TSG RAN WG1 #115, or“RAN1 Chair’s Notes”, 3GPP TSG RAN WG1 #112.

[0053] In the following, embodiments of a UE 100 for reporting performance of a first AI / ML model 110 executed by the UE 100 are described with reference to Fig. 1. The UE 100 comprises a communications interface 120, described in further detail below, operative to communicate with a RAN 200 via one or more radio links 150. The UE 100 further comprises processing circuitry 130, described in further detail below, causing the UE 100 to be operative in accordance with embodiments of the invention described herein.

[0054] The one or more radio links 150 may be established, configured, and / or used, for signaling (i.e., exchange of messages and / or data) between the UE 100 and one or more access nodes 210 of the RAN 200, e.g., gNBs or the like, in accordance with a 3GPP standard. Throughout this disclosure, any communication which is described to commence between the UE 100 and the RAN 200 is to be understood to be between the UE 100 and one or more access nodes 210 of the RAN 200.

[0055] More specifically, and with further reference to Fig. 3, which is a signaling diagram illustrating reporting of a performance metric indicative of a performance of the first AI / ML model 110 executed by the UE 100, the UE 100 is operative to receive first configuration 301 information for configuring execution of the first AI / ML model 110 by the UE 100. The execution of the first AI / ML model 110 is for inference of one or moreproperties related to the one or more radio links 150. The one or more properties related to the one or more radio links 150 may, e.g., comprise one or more of: CSI reporting, beam management, and positioning of the UE 100. The first AI / ML model 110 may either be executed by the processing circuitry 130 or by a separate processing circuitry based on hardware, software, or a combination thereof.

[0056] The first configuration information 301 is received from the RAN 200, i.e. , from an access node 210. The first AI / ML model 110 may, e.g., be configured as a one-sided AI / ML model. Optionally, the first AI / ML model 110 may be configured to perform joint inference with a second AI / ML model 211 executed by the RAN 200. For example, the first AI / ML model 110 may be configured as the UE-side part of a two-sided AI / ML model, e.g., for AE-based CSI feedback / reporting together with the corresponding second AI / ML model 211. In the case of two-sided CSI compression, the first configuration information may include at least one of CSI-RS configurations and CSI-report configurations (including payload-related configurations such as compression ratio, encoder output size, quantization size, puncturing rate, etc). For the case of UE-side CSI prediction, the first configuration may include CSI-RS measurement occasions, CSI prediction window, etc.

[0057] The UE 100 is further operative to receive second configuration information 302 for configuring reporting, by the UE 100, of a performance metric. The performance metric is indicative of the performance of the first AI / ML model 110 when the first AI / ML model 110 is executed by the UE 100. The first AI / ML model 110 is preferably executed in accordance with first configuration information 301 received by the UE 100.

[0058] The second configuration information 302 may include, e.g., the monitored rank, layer, configurations, models, etc. Note that some configuration aspects of AI / ML performance monitoring may be configured independently, or be derived, from other configurations which relate to the execution of the first AI / ML model 110. For example, in the case of two-sidedCSI compression, the UE may be configured to transmit the CSI report with rank = 2. In this case, rank = 2 can then be assumed as the configuration for performance monitoring.

[0059] Another configuration related the AI / ML performance monitoring may be a performance threshold or a performance range. For the case of two-sided CSI compression or UE-side CSI prediction, the performance can be measured in terms of SGCS, NMSE, etc. The UE 100 may be configured with one or more threshold values for AI / ML performance reporting. The second configuration information 302 may also reference a pre-configured configuration information, e.g., threshold values or ranges, which are preconfigured in the UE 100, e.g., standardized values which optionally may depend on UE capabilities.

[0060] The first configuration information 301 and the second configuration information 302 may be received by the UE 100 in separate messages or signals, or combined into a single message or signal.

[0061] The UE 100 is further operative to signal, to the RAN 200, a performance metric 312 indicating a low performance of the first AI / ML model 110. The UE 100 is operative to signal the low performance metric in response to determining 311 that a link quality of the one or more radio links 150, between the UE 100 and the RAN 200, is below a lower link-quality threshold and / or above an upper link-quality threshold.

[0062] In the present context, the performance metric may be a configured set of discrete values, or a range of discrete values, which the UE 100 can use to indicate a performance of the first AI / ML model 110, when executed by the UE 100, to the RAN 200. The performance metric 312 indicating a low performance of the first AI / ML model 110 may correspond to a lowest value of a configured set of values for the performance metric. Alternatively, the performance metric 312 indicating a low performance of the first AI / ML model 110 may correspond to a performance which is lower, i.e. , worse, thana measured performance, i.e., the actual performance, of the first AI / ML model 110 when executed by the UE 100.

[0063] The UE 100 may be operative to signal the performance metric 312 indicating a low performance of the first AI / ML model 110 as a performance monitoring report, e.g., in the Physical Uplink Control Channel (PUCCH) or in the Physical Uplink Shared Channel (PUSCH). Alternatively, the performance monitoring report may be signaled together with the result of the inference by the first AI / ML model 110. For example, in the case of AI / ML-based CSI compression or prediction, the performance monitoring report may be signaled together with the AI / ML-based CSI report.

[0064] Embodiments of the invention enable the UE 100 to signal a performance metric 312 indicating a low performance of the first AI / ML model 110, rather than the model’s measured, i.e., actual, performance, to indicate the UE 100’s preference to the RAN 200 to be configured with legacy mechanisms. Herein, the term “legacy mechanisms” encompasses mechanisms which typically do not rely on AI / ML models and are therefore computationally less complex, but may also include AI / ML models with less resource-demanding configurations. Advantageously, the UE 100 can thereby reduce its power consumption and / or make computational resources available for other computational tasks. This may be used in situations when there is little or no benefit for the UE 100 to use the first AI / ML model 110 for inference of one or more properties related to the one or more radio links 150, or in situations where the first AI / ML model performs worse than corresponding legacy mechanisms.

[0065] Signaling the low-performance metric in response to determining that a link quality of the one or more radio links 150, between the UE 100 and the RAN 200, is below a lower link-quality threshold and / or above an upper linkquality threshold is based on the understanding that the gain of executing the first AI / ML model 110, for inference of one or more properties related to the one or more radio links 150, may depend on radio conditions which arecurrently experienced by the UE 100. For example, the gain of AI / ML tools is typically limited if the link quality of the one or more radio links 150 is relatively low, e.g., below a lower link-quality threshold. This may happen if the UE 100 is not able to obtain appropriate input for the AE-encoder. In this situation, the performance of non-AI / ML-based (legacy) mechanisms and AI / ML-based mechanism is comparable, and executing the first AI / MOL model 110 by the UE 100 is therefore not beneficial in view of the increased power consumption and / or the increased usage of computational resources. Similarly, in situations where the link quality of the one or more radio links 150 is relatively high, e.g., above an upper link-quality threshold, the first AI / ML model 110 may not perform considerably better than legacy mechanisms which do not rely on AI / ML models, and which consume less power and computational resources.

[0066] The UE 100 may further be operative to signal, to the RAN 200, a performance metric 314 indicating a measured performance of the first AI / ML model 110. The performance metric 314 indicating a measured performance of the first AI / ML model 110 is signaled in response to determining 313 that the link quality of the one or more radio links 150 is in a range between the lower link-quality threshold and the upper link-quality threshold. In other words, in situations where the performance of legacy mechanisms not relying on AI / ML models is inferior to AI / ML tools, the UE 100 reports the actual, measured performance of the first AI / ML model 110 to the RAN 200 so as to enable the RAN 200 to configure the UE 100 with an appropriate configuration for the execution of the first AI / ML model 110.

[0067] The UE 100 may be operative to determine the link quality of the one or more radio links 150 based on a set of consecutive Channel Quality Indicator (CQI) indices signaled by the UE 100 to the RAN 200. For instance, the UE 100 may be operative to determine that the link quality of the one or more radio links 150 is below the lower link-quality threshold after a threshold number of consecutive CQIs indicating the lowest channel quality (e.g., A.)P110678W001 2025-03-2807:36

[0068] 14

[0069] have been transmitted to the RAN 200. The threshold number of consecutive CQIs indicating the lowest channel quality may, e.g., be equal to 4 or 5. Similarly, the UE 100 may be operative to determine that the link quality of the one or more radio links 150 is above the upper link-quality threshold after a threshold number of consecutive CQIs indicating the highest channel quality (e.g., / H) have been transmitted to the RAN 200. The threshold number of consecutive CQIs indicating the highest channel quality may, e.g., be equal to 4 or 5.

[0070] Alternatively, the UE 100 may be operative to determine the link quality of the one or more radio links 150 based on a set of consecutive ACKs and / or NACKs received by the UE 100 from the RAN 200 in response to transmissions by the UE 100 to the RAN 200, or based on a set of consecutive ACKs and / or NACKs transmitted by the UE 100 to the RAN 200 in response to transmissions by the RAN 200 to the UE 100. In practice, the link quality of the one or more radio links 150 may be determined derived from one or more of: a number of consecutive ACKs, a number of consecutive NACKs, a ratio between ACKs and NACKs, a ratio between ACKs and a total number of ACKs and NACKs, and a ratio between NACKs and a total number of ACKs and NACKs. For instance, the UE 100 may be operative to determine that the link quality of the one or more radio links 150 is below the lower link-quality threshold after a threshold number of consecutive NACKs have been received from, or transmitted to, the RAN 200. The threshold number of consecutive NACKs may, e.g., be equal to 3. Similarly, the UE 100 may be operative to determine that the link quality of the one or more radio links 150 is above the upper link-quality threshold after a threshold number of consecutive ACKs have been received from, or transmitted to, the RAN 200. The threshold number of consecutive ACKs may, e.g., be equal to 10.

[0071] As yet another alternative, the UE 100 may be operative to determine the link quality of the one or more radio links 150 based on a ReferenceP110678W001 2025-03-2807:36

[0072] 15

[0073] Signal Received Power (RSRP) of a reference signal which is received by the UE 100. For instance, the UE 100 may be operative to determine that the link quality of the one or more radio links 150 is below the lower link-quality threshold if the RSRP is less than a lower RSRP threshold value, e.g., -100 dBm. Similarly, the UE 100 may be operative to determine that the link quality of the one or more radio links 150 is above the upper link-quality threshold if the RSRP is larger than a higher RSRP threshold value, e.g., -70 dBm.

[0074] As yet a further alternative, the UE 100 may be operative to determine the link quality of the one or more radio links 150 based on a BLER. For instance, the UE 100 may be operative to determine that the link quality of the one or more radio links 150 is below the lower link-quality threshold if the BLER is larger than an upper BLER threshold value, e.g., 20%. Similarly, the UE 100 may be operative to determine that the link quality of the one or more radio links 150 is above the upper link-quality threshold if the BLER is less than a lower BLER threshold value, e.g., 5%.

[0075] It will be appreciated that the UE 100 may be operative to signal, to the RAN 200, a performance metric 312 indicating a low performance of the first AI / ML model 110 in response to additional, or alternative, conditions other than determining 311 that a link quality of the one or more radio links 150 is below than a lower link-quality threshold and / or above an upper link-quality threshold.

[0076] For example, the UE 100 may be operative to signal, to the RAN 200, a performance metric 312 indicating a low performance of the first AI / ML model 110 in response to determining 321 that a battery level of a battery comprised in the UE 100 is below a battery-level threshold. This may be the case when the UE 100 is in a low-battery condition and needs to preserve its battery for more critical or prioritized usage. In such situation, the UE 110 may prefer to be configured with a legacy mechanism which consumes lesspower compared to the AI / ML-based mechanism, i.e. , the first AI / ML model 110.

[0077] Optionally, the UE 100 may further be operative to signal, to the RAN 200, UE Assistance Information (UAI) indicative of the UE’s preference for power saving. In practice, the UAI indicative of the UE’s preference for power saving is signaled shortly before, or shortly after, signaling the performance metric 312 indicating a low performance of the first AI / ML model 110. Signaling the UAI indicative of the UE’s preference for power saving is indicative of the UE 100 entering a lower-power mode.

[0078] The UE 100 may further be operative to signal, to the RAN 200, a performance metric 314 indicating a measured performance of the first AI / ML model 110. The actual, measured performance of the first AI / ML model 110 is signaled in response to determining 323 that the battery level of the battery comprised in the UE 100 is above, or equal to, the battery-level threshold.

[0079] A further example for a condition which may trigger the UE 100 to signal a performance metric 312 indicating a low performance of the first AI / ML model 110 to the RAN 200 is the speed of the UE 100. More specifically, the UE 100 may be operative to signal, to the RAN 200, a performance metric indicating a low performance 312 of the first AI / ML model in response to determining 331 that a speed of the UE 100 is below a speed threshold. This is advantageous in situations when the UE 100 is stationary or moving at relatively low speed. For example, although AI / ML based prediction may perform much better than non-AI / ML based mechanisms for CSI prediction, for CSI compression in the temporal-spatial-frequency domain the correlation between CSI report occasions may no longer hold, and non-AI / ML legacy mechanisms (such as auto-regression based prediction) may be sufficient. The speed threshold may, e.g., be equal to 30 km / h.

[0080] The UE 100 may be operative to determine its speed using a positioning sensor, such as a Global Navigation Satellite System (GNSS) or an Inertial Measurement Unit (IMU). Alternatively, or additionally, the UE 100may be operative to determine its speed based on measurements of the Doppler spread / shift or the Timing Advance by the communication interface 120.

[0081] The UE 100 may further be operative to signal, to the RAN 200, a performance metric 314 indicating a measured performance of the first AI / ML model 110 in response to determining 333 that the speed of the UE 100 is above, or equal to, the speed threshold. That is, the UE 100 reports its actual, measured performance of the first AI / ML model 110 in situations when the use of the first AI / ML model 110 for inference of one or more properties related to the one or more radio links 150 is advantageous.

[0082] The UE 100 may further be operative to signal, to the RAN 200, a performance metric 312 indicating a measured performance of the first AI / ML model 110. This performance metric 312 indicating the actual, measured performance of the first AI / ML model 110 is signaled in response to determining 341 that a reporting condition according to the second configuration information 302 is met. The reporting condition may, e.g., relate to periodic reporting, or reporting triggered by a change in the measured performance of the first AI / ML model 110 by more than a threshold value.

[0083] As an alternative to signaling, to the RAN 200, a performance metric 312 indicating a low performance of the first AI / ML model 110 triggered by the conditions related to link quality, UE battery level, and UE speed, as described hereinbefore, one may also envision embodiments of the invention which refrain from, i.e. , omit, signaling a performance metric 314 indicating a measured performance of the first AI / ML model 110 if one or more of the conditions on link quality, UE battery level, and UE speed, are met. In this case, the RAN 200 may be configured to re-configure execution of the first AI / ML model 110 by the UE, either to a less resource consuming configuration or to a non-AIL / ML legacy mechanism, based on an absent performance metric 312 indicating a low performance of the first AI / ML model 110, which the RAN 200 had expected (e.g., in the case ofP110678W001 2025-03-2807:36

[0084] 18

[0085] periodic reporting, based on ACKs / NACKs which the RAN 200 has transmitted to the UE 100, or CQI indices which the RAN 200 has received from the UE 100).

[0086] Fig. 2 schematically illustrates embodiments of the processing circuitry 130 comprised in the UE 100. The processing circuitry 130 may comprise one or more processors 131, such as Central Processing Units (CPUs), microprocessors, application processors, application-specific processors, or a combination thereof, and a memory 132 comprising a computer program 133, i.e., software, comprising instructions. When executed by the processor(s) 131 , the instructions cause the UE 100 to become operative in accordance with embodiments of the invention disclosed herein. The processing circuitry 130 may alternatively or additionally comprise one or more Application-Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or the like, which are operative to cause the UE 100 to become operative in accordance with embodiments of the invention disclosed herein. The processing circuitry 130 may further comprise circuitry for interfacing one or more of a separate processing circuitry on which the first AI / ML model 110 is provisioned (if provided separately from the processing circuitry 130), the communications interface 120, an IMU and / or a positioning sensor (for determining the speed of the UE).

[0087] The communications interface 120 comprised in the UE 100 may comprise circuitry which is operative to communicate, i.e., exchange signals, messages, and / or data, with corresponding communications interfaces comprised in one or more access nodes 210 of the RAN 200. In particular, the communications interface 120 is operative to communicate with the one or more access nodes 210 in accordance with one or more 3GPP standards, e.g., standards commonly known as 5G and / or 6G.

[0088] In the following, embodiments of a method 400 of reporting performance of a first AI / ML model 110 executed by a UE 100 are describedwith reference to Fig. 4. The UE 100 communicates with a RAN 200 via one or more radio links 150. The first AI / ML model 110 may be configured to perform joint inference with a second AI / ML model 211 executed by the RAN 200. The method 400 is performed by the UE 100 and comprises receiving 401 first configuration information for configuring execution of the first AI / ML model 110 by the UE 100, for inference of one or more properties related to the one or more radio links 150. The one or more properties related to the one or more radio links 150 may, e.g., comprise one or more of: CSI reporting, beam management, and positioning. The first configuration information is received from the RAN.

[0089] The method 400 further comprises receiving 402 second configuration information for configuring reporting, by the UE 100, of a performance metric indicative of the performance of the first AI / ML model 110 when the first AI / ML model 110 is executed by the UE 100. The second configuration information is received from the RAN.

[0090] The method 400 further comprises signaling 412, to the RAN 200, a performance metric indicating a low performance of the first AI / ML model 110. The performance metric indicating a low performance of the first AI / ML model 110 is signaled in response to determining 411 that a link quality of the one or more radio links 150 is below a lower link-quality threshold and / or above an upper link-quality threshold.

[0091] The performance metric indicating a low performance of the first AI / ML model 110 may, e.g., correspond to a lowest value of a configured set of values for the performance metric.

[0092] The link quality of the one or more radio links 150 may, e.g., be determined based on a set of consecutive CQI indices signaled by the UE 100 to the RAN 200. Alternatively, the link quality of the one or more radio links 150 may be determined based on a set of consecutive ACKs and / or NACKs received by the UE 100 from the RAN 200 in response to transmissions by the UE 100 to the RAN 200.The method 400 may further comprise signaling 414, to the RAN 200, a performance metric indicating a measured performance of the first AI / ML model 110. The measured performance of the first AI / ML model 110 is signaled in response to determining 413 that the link quality of the one or more radio links 150 is in a range between the lower link-quality threshold and the upper link-quality threshold.

[0093] The method 400 may further comprise signaling 412, to the RAN 200, a performance metric indicating a low performance of the first AI / ML model 110 in response to determining 421 that a battery level of a battery comprised in the UE 100 is below a battery-level threshold. Optionally, the method 400 may further comprise signaling, to the RAN 200, UE Assistance Information (UAI) indicative of the UE’s 100 preference for power saving.

[0094] The method 400 may further comprise signaling 414, to the RAN 200, a performance metric indicating a measured performance of the first AI / ML model 110. The performance metric indicating a measured performance of the first AI / ML model 110 is signaled in response to determining 423 that the battery level of the battery comprised in the UE 100 is above, or equal to, the battery-level threshold.

[0095] The method 400 may further comprise signaling 412, to the RAN 200, a performance metric indicating a low performance of the first AI / ML model 110 in response to determining 431 that a speed of the UE 100 is below a speed threshold.

[0096] The method 400 may further comprise signaling 414, to the RAN 200, a performance metric indicating a measured performance of the first AI / ML model 110. The performance metric indicating a measured performance of the first AI / ML model 110 is signaled in response to determining 433 that the speed of the UE 100 is above, or equal to, the speed threshold.

[0097] The method 400 may further comprise signaling 414, to the RAN 200, a performance metric indicating a measured performance of the first AI / ML model 110. The performance metric indicating a measured performance ofP110678W001 2025-03-2807:36

[0098] 21

[0099] the first AI / ML model 110 may be signaled in response to determining 441 that a reporting condition according to the second configuration information is met.

[0100] It will be appreciated that the method 400 may comprise additional, alternative, or modified, steps in accordance with what is described throughout this disclosure. An embodiment of the method 400 may be implemented as the computer program 133 comprising instructions which, when the computer program 133 is executed by one or more processor(s) 131 comprised in the UE 100, cause the UE 100 to carry out the method 400 and become operative in accordance with embodiments of the invention described herein. The computer program 133 may be stored in a computer-readable data carrier, such as the memory 132. Alternatively, the computer program 133 may be carried by a data carrier signal, e.g., downloaded to the memory 132 via the communications interface 120.

[0101] The person skilled in the art realizes that the invention by no means is limited to the embodiments described above. On the contrary, many modifications and variations are possible within the scope of the appended claims.

Claims

P110678W001 2025-03-2807:3622CLAIMS1. A User Equipment, UE, (100) for reporting performance of a first Artificial Intelligence / Machine-Learning, AI / ML, model (110) executed by the UE (100), the UE (100) comprising a communications interface (120) operative to communicate with a Radio Access Network, RAN, (200, 210) via one or more radio links (150), and processing circuitry (130) causing the UE (100) to be operative to:receive, from the RAN (200, 210), first configuration information (301) for configuring execution of the first AI / ML model (110) by the UE (100), for inference of one or more properties related to the one or more radio links (150),receive, from the RAN (200, 210), second configurationinformation (302) for configuring reporting, by the UE (100), of a performance metric indicative of the performance of the first AI / ML model (110) when the first AI / ML model (110) is executed by the UE (100), andin response to determining (311 ) that a link quality of the one or more radio links (150) is below a lower link-quality threshold and / or above an upper link-quality threshold, signal, to the RAN, a performance metric (312) indicating a low performance of the first AI / ML model (110).

2. The UE (100) according to claim 1 , operative to determine the link quality of the one or more radio links (150) based on a set of consecutive Channel Quality Indicator, CQI, indices signaled by the UE (100) to the RAN (200, 210).

3. The UE (100) according to claim 1 , operative to determine the link quality of the one or more radio links (150) based on a set of consecutive ACKs and / or NACKs received by the UE (100) from the RAN (200, 210) in response to transmissions by the UE (100) to the RAN (200, 210), or basedP110678W001 2025-03-2807:3623on a set of consecutive ACKs and / or NACKs transmitted by the UE 100 to the RAN 200 in response to transmissions by the RAN 200 to the UE 100.

4. The UE (100) according to any one of claims 1 to 3, further operative to signal, to the RAN (200, 210), in response to determining (313) that the link quality of the one or more radio links (150) is in a range between the lower link-quality threshold and the upper link-quality threshold, a performance metric (314) indicating a measured performance of the first AI / ML model (110).

5. The UE (100) according to any one of claims 1 to 4, further operative to signal, to the RAN (200, 210), in response to determining (321) that a battery level of a battery comprised in the UE (100) is below a battery-level threshold, a performance metric (312) indicating a low performance of the first AI / ML model (110).

6. The UE (100) according to claim 5, further operative to signal, to the RAN (200, 210), UE Assistance Information, UAI, indicative of the UE’s preference for power saving.

7. The UE (100) according to claim 5 or 6, further operative to signal, to the RAN (200, 210), in response to determining (323) that the battery level of the battery comprised in the UE (100) is above, or equal to, the battery-level threshold, a performance metric (314) indicating a measured performance of the first AI / ML model (110).

8. The UE (100) according to any one of claims 1 to 7, further operative to signal, to the RAN (200, 210), in response to determining (331) that a speed of the UE (100) is below a speed threshold, a performance metric (312) indicating a low performance of the first AI / ML model (110).P110678W001 2025-03-2807:36249. The UE (100) according to claim 8, further operative to signal, to the RAN (200, 210), in response to determining (333) that the speed of the UE (100) is above, or equal to, the speed threshold, a performance metric (314) indicating a measured performance of the first AI / ML model (110).

10. The UE (100) according to any one of claims 1 to 9, wherein the performance metric (312) indicating a low performance of the first AI / ML model (110) corresponds to a lowest value of a configured set of values for the performance metric.

11. The UE (100) according to any one of claims 1 to 10, further operative to signal, to the RAN (200, 210), in response to determining (341) that a reporting condition according to the second configuration information is met, a performance metric (314) indicating a measured performance of the first AI / ML model (110).

12. The UE (100) according to any one of claims 1 to 11 , wherein the one or more properties related to the one or more radio links (150) comprise one or more of: Channel State Information, CSI, reporting, beam management, and positioning.

13. The UE (100) according to any one of claims 1 to 12, wherein the first AI / ML model (110) is configured to perform joint inference with a second AI / ML model (211 ) executed by the RAN (200, 210).

14. A method (400) of reporting performance of a first Artificial Intelligence / Machine-Learning, AI / ML, model (110) executed by a User Equipment, UE, (100) communicating with a Radio Access Network, RAN,P110678W001 2025-03-2807:3625(200, 210) via one or more radio links (150), the method performed by the UE (100) and comprising:receiving (401), from the RAN, first configuration information for configuring execution of the first AI / ML model (110) by the UE (100), for inference of one or more properties related to the one or more radio links (150),receiving (402), from the RAN, second configuration information for configuring reporting, by the UE (100), of a performance metric indicative of the performance of the first AI / ML model (110) when the first AI / ML model (110) is executed by the UE (100), andin response to determining (411 ) that a link quality of the one or more radio links (150) is below a lower link-quality threshold and / or above an upper link-quality threshold, signaling (412), to the RAN (200, 210), a performance metric indicating a low performance of the first AI / ML model (110).

15. The method (400) according to claim 14, wherein the link quality of the one or more radio links (150) is determined based on a set of consecutive Channel Quality Indicator, CQI, indices signaled by the UE (100) to the RAN (200, 210).

16. The method (400) according to claim 14, wherein the link quality of the one or more radio links (150) is determined based on a set of consecutive ACKs and / or NACKs received by the UE (100) from the RAN (200, 210) in response to transmissions by the UE (100) to the RAN (200, 210) or based on a set of consecutive ACKs and / or NACKs transmitted by the UE 100 to the RAN 200 in response to transmissions by the RAN 200 to the UE 100.

17. The method (400) according to any one of claims 14 to 16, further comprising signaling (414), to the RAN (200, 210), in response to determining (413) that the link quality of the one or more radio links (150) isP110678W001 2025-03-2807:3626in a range between the lower link-quality threshold and the upper link-quality threshold, a performance metric indicating a measured performance of the first AI / ML model (110).

18. The method (400) according to any one of claims 14 to 17, further comprising signaling (412), to the RAN (200, 210), in response to determining (421) that a battery level of a battery comprised in the UE (100) is below a battery-level threshold, a performance metric indicating a low performance of the first AI / ML model (110).

19. The method (400) according to claim 18, further comprising signaling, to the RAN (200, 210), UE Assistance Information, UAI, indicative of the UE’s (100) preference for power saving.

20. The method (400) according to claim 18 or 19, further comprising signaling (414), to the RAN (200, 210), in response to determining (423) that the battery level of the battery comprised in the UE (100) is above, or equal to, the battery-level threshold, a performance metric indicating a measured performance of the first AI / ML model (110).

21. The method (400) according to any one of claims 14 to 20, further comprising signaling (412), to the RAN (200, 210), in response to determining (431) that a speed of the UE (100) is below a speed threshold, a performance metric indicating a low performance of the first AI / ML model (110).

22. The method (400) according to claim 21 , further comprising signaling (414), to the RAN (200, 210), in response to determining (433) that the speed of the UE (100) is above, or equal to, the speed threshold, aP110678W001 2025-03-2807:3627performance metric indicating a measured performance of the first AI / ML model (110).

23. The method (400) according to any one of claims 14 to 22, wherein the performance metric indicating a low performance of the first AI / ML model (110) corresponds to a lowest value of a configured set of values for the performance metric.

24. The method (400) according to any one of claims 14 to 23, further comprising signaling (414), to the RAN (200, 210), in response to determining (441) that a reporting condition according to the second configuration information is met, a performance metric indicating a measured performance of the first AI / ML model (110).

25. The method (400) according to any one of claims 14 to 24, wherein the one or more properties related to the one or more radio links (150) comprise one or more of: Channel State Information, CSI, reporting, beam management, and positioning.

26. The method (400) according to any one of claims 14 to 25, wherein the first AI / ML model (110) is configured to perform joint inference with a second AI / ML model (211) executed by the RAN (200, 210).

27. A computer program (133) comprising instructions which, when the computer program (133) is executed by one or more processors (131) comprised in a User Equipment, UE, (100), cause the UE (100) to carry out the method (400) according to any one of claims 14 to 26.

28. A computer-readable data carrier (132) having stored thereon the computer program (133) according to claim 27.2829. A data carrier signal carrying the computer program (133) according to claim 27.

Images