3GPP TSG-RAN WG2 Meeting#130		R2-2503372
St Julian’s, Malta, May 19th – 23rd, 2025

Source:	vivo
Title:	Configuration and report of inference input for NW-sided model
Agenda Item:	8.3.3
Document for:	Discussion and Decision

Conclusion
In the contribution, we have the following observations and proposals:
General principle
For NW-sided model of RRM measurement prediction, the existing RRM measurement framework can be reused as a baseline for configuration and reporting of the inference input.
Both periodic and event-triggered reporting schemes can be configured for UE to report inference input for NW-sided model of RRM measurement prediction.
Use cases
Consider L1-filtered RSRP as a candidate inference input for the NW-sided model, while the L1 sample (without filtering) is not considered.
UE can be configured to report the following objects as the contents of inference inputs for NW-sided model:
- For temporal domain prediction: a list of L1-filtered beam-level RSRP, L3 beam-level RSRP, and L3 cell-level RSRP of one cell at multiple time instances;
- For spatial domain prediction: a list of L1-filtered beam-level RSRP of one cell;
- For frequency domain prediction: a list of L1-filtered beam-level RSRP, L3 beam-level RSRP and L3 cell-level RSRP of one or more cells on specific frequency(s).
To support sub-use cases 2 and 5 for NW-sided model, UE can be configured to report L3 beam-level and L3 cell-level measurement results per cell.
To support sub-use cases 1, 3, 4 or 6 for NW-sided model, UE can be configured to report L1-filtered beam-level measurement results.
Prediction domains
In the current RRM measurement report, one measurement result corresponds to the results measured at one time instance.
NW configures the specific inference inputs as measurement object, where UE is agnostic about the NW-sided model.
At least for temporal domain prediction, support reporting RRM measurement results of one cell at multiple time instances in one measurement report as the inference input of NW-sided model.
For the inference input of NW-sided model, RAN2 to discuss how to configure the UE to report measurement results at multiple time instances in one report in WI Phase.

3GPP TSG-RAN WG2 Meeting #130	 R2-2503466
St Julian’s, Malta, May 19-23, 2025

Agenda Item:8.3.3
Source: NEC
Title: Discussion on AIML mobility Configuration and Report of Inference input to network sided model
Document for: Discussion and Decision
1 

Conclusions
We have the following proposals for the configuration and reporting of inference inputs for the network sided model.
Proposal-1: Introduce periodic/event triggered inference reporting for network sided AI/ML models for mobility management based on delay thresholds in order to ensure timely mobility decisions.
Proposal-2: UE may indicate the confidence level of the inference inputs in terms of some reliability indicators to help network take some decision based on the quality of data received. 

3GPP TSG-RAN WG2 Meeting #130	R2-2503515
St. Julians, Malta, May 19th – May 23rd, 2025 	

Agenda item:	8.3.3
Source:	Qualcomm Incorporated
Title:	Configuration and reporting for inference for network-side model
WID/SID:	FS_NR_AIML_Mob – Release 19
Document for:	Discussion and Decision

Conclusions
Based on the discussion above, we recommend RAN2 to discuss the proposals and observations below.
Observation 1. If UE provides RRM measurement reports to support network-side RRM measurement prediction, then no additional information or reports are needed by the source gNB from the UE to support network-side Measurement Event prediction.
Proposal 1: For network-side RRM measurement prediction, the legacy RRM measurement configuration and RRM measurement reporting framework can be used. Any additional impact to the specification is left up to the WI phase discussion. 

3GPP TSG-RAN WG2 Meeting #130	                                          R2-2503639
Malta, 19 – 23, May, 2025  							    

Source: 	Xiaomi
Title:  	Discussion on configuration and report for NW sided model
Agenda Item:	8.3.3
Document for:	Discussion and Decision

Conclusions
According to the analysis given above, we have the following observations and proposals:
Observation 1: Huge spec impact is foreseen to support L1 beam as model input for NW sided model, i.e. sub case 1, 3, 4 and 6, including 
new signalling to report L1 beam measurement result of serving and neighbour cell for RRM, as current L1 beam for RRM is configured by CSI-RS-ResourceConfigMobility, rather than using CSI framework
new requirement on UE measurement with potential RAN4 impact
new signalling to report how L1 filtering is done, which is now completely up to UE implementation
Proposal 1: L1 beam measurement result as model input, i.e. sub case 1 and 3, is not supported for NW sided model in RRM prediction and L3 beam prediction.
Proposal 2: Legacy measurement report procedure can be reused for UE to report L3 cell level or L3 beam level measurement result as model input for NW sided model.
Proposal 3: There is no specification impact associated to NW-side model inference. It’s up to NW how to use the inference results.
Proposal 4: RAN2 confirms following agreements from AI air is also applicable to AI mobility
RAN2 confirms that UE will not be informed about any gNB -sided model/functionality management decision (e.g., selection, (de)activation, switching, fallback, etc.)
RAN2 confirms that UE will not be involved in any gNB -sided model/functionality management decision making (e.g., selection, (de)activation, switching, fallback, etc.), except being configured to provide the required measurement/data. 
For the network-side model, required network side additional condition is left to the network implementation

3GPP TSG-RAN WG2 #130	R2-2503687
St Julian, Malta, May 19th – 23rd, 2025

Agenda item:	   8.3.3
Source:		Lenovo
Title:                            Configuration and Report of Inference Input to NW-sided model
Document for:		Discussion and Decision

Conclusion

Based on the discussion above, we request RAN2 to discuss and agree on the following proposals:

Observation: The following agreements can be reviewed and extended to RRM measurement prediction:
LCM for NW-sided model
RAN2 focuses on the data collection procedure from UE to NW (e.g., gNB, LMF, or OAM) for the sake of NW-sided model LCM (including training, inference, management).
RAN2 to consider an RRC configuration to configure radio measurements and the related reporting to enable data collection for NW-side training
For AI/ML based beam management, RAN2 assumes the L1 measurement framework shall be used for configuring the input data of the NW side AI/ML model inference.  FFS if further enhancements are needed

Proposal 1: For the network-sided model for AI/ML Mobility, RAN2 confirms that L1/L3 signaling can be used for UE inputs for inference.
Proposal 2: The input from UE for inference at the network side could be either legacy L1/L3 beam measurement, L3 cell measurement, or L1 beam prediction results (if UE is capable of L1 beam prediction in the AI for air interface). 
Proposal 3: NW can configure UE to report in a periodic, one-shot (aperiodic), or event-triggered manner for RRM measurement prediction for NW-sided model. FFS details.

3GPP TSG-RAN WG2 Meeting #130													     R2-2503706
St. Julians, Malta, May 19th – 23rd, 2025									  
Agenda item:	8.3.3
Source:	Apple
Title:	On NW-sided AI/ML mobility
Document for:	Discussion and Decision

1	

Proposals
Observation 1: the topic of the present contribution is NW-sided model inference and the air interface enhancements it requires.
Observation 2: the latency, signaling overhead, and the extra burden on UE of network-sided models have never been studied.
Proposal 1: for NW-sided inference, a UE should never be configured with extra measurements (in terms of periodicity, cells and frequencies to measure, etc ) compared to what it would have been configured for legacy operation.
Proposal 2: UE should be aware whether the measurements configured are for network-sided inference.
Proposal 3: it should be demonstrated how NW-sided inference can possibly reduce measurement overhead in order to graduate this work to a normative phase.
Proposal 4: at least some advantages of NW-sided prediction compared to UE-sided should be demonstrated in order to graduate this work to a normative phase.
5	

3GPP TSG-RAN WG2 Meeting #130                                  R2-2503984
St.Julians, Malta, May 19th – 23rd, 2025                                                                                                              
Agenda item:		8.3.2 
Source:				Samsung
Title:					Discussion on configuration and report of inference input for NW-side model
Document for:		Discussion & Decision
1	

Conclusion
Based on the above, RAN2 is requested to discuss and agree on the following proposals:
Common part:
Proposal. 1: For NW-side model inference, RAN2 should consider the ways to reduce the signalling overhead for the report of inference input (e.g., event-triggered reporting, adaptive reporting based on mobility status, …).
Temporal domain Case A:
Proposal. 2: For temporal domain Case A, UE can log the L3 Cell-level measurement results of N recent time instances and report them to NW for inference input report of NW-side model. The N (i.e., OW length) can be configured by NW.
Temporal domain Case B:
Observation. 1: For temporal domain Case B, the motivation is to reduce UE’s measurement effort for energy saving.
Observation. 2: For temporal domain Case B, for UE-side model case, the event-triggered measurement reporting can be used. I.e., UE can report the measurement report only when it detects some event.
Observation. 3: For temporal domain Case B, for NW-side model case, only the periodical measurement reporting can be used. I.e., UE should periodically report the measurement report with very short interval (e.g., 80 ms), which can increase the energy consumption of UE.
Proposal. 3: For temporal domain Case B, RAN2 exclude the NW-side model case.
Frequency domain:
Proposal. 4: For frequency domain prediction, UE can report one latest measurement result for each cell/beam to NW as in the legacy for inference input report of NW-side model. No further enhancement is needed.
L3 beam-level prediction:
Proposal. 5: For L3 beam-level prediction, UE can log the L1/L3 Beam-level measurement results of N recent time instances and report them to NW for inference input report of NW-side model. The N (i.e., OW length) can be configured by NW.
Observation. 4: L1/L3 Beam-level logging and reporting will increase the signalling overhead.

3GPP TSG-RAN WG2 Meeting #130					   			 R2-2504109
St. Julian’s, Malta, 19 - 23 May, 2025

Title: 	Discussion on NW-sided model support
Source: 	Huawei, HiSilicon
Agenda item:	8.3.3
Document for:	Discussion/Decision

Conclusion
In this contribution, we have the following observations and proposals:
Spatial Domain Prediction
Observation 1: Spatial domain prediction is proved to be feasible and can help to realize both UE and NW-side energy saving.
Proposal 1: RAN2 to support spatial domain beam-level prediction for both NW-sided model and UE-sided model.

Potential Impacts on Measurement Framework
Observation 2: There are no measurement framework impacts to enable NW-sided RRM measurement inference using sub-use case 2 and sub-use case 5.
Observation 3: Changes to the current L1 or L3 measurement framework are needed to support NW-sided inference using sub-use case 1, 3, 4 and 6.
Proposal 2: All analyzed RRM sub-use cases should be supported for NW-sided model inference.
Proposal 3: The following options should be considered to support L1 beam level measurements availability at the gNB-CU:
Extending L3 measurement framework to support L1 beam level results reporting
Introduce L1 beam levels forwarding from gNB-DU to gNB-CU

Assistance information for the NW-side model
Proposal 4: No assistance information from the UE is needed to support NW-side model inference other than L1/L3 RSRP beam/cell level measurement results.

3GPP TSG-RAN2 Meeting #130	R2-2504460
St.Julians, Malta,  May 19th – 23rd , 2025
Source: 			ZTE Corporation
Title: 	Discussion on configuration and report of inference input to the network side model
Agenda item:	8.3.3
Document for: 	Discussion and Decision

Conclusion and proposals
Observation 1: In the existing measurement report, the measurement report always includes RSRP, RSRQ and available SINR results for the serving cell configured with servingCellMO, which may be not needed for model inference.

Proposal 1: For inference of network side model, the network can configure the frequency that the UE shall perform inference on and the type of the reported measurement results (e.g. L1 filtered beam results, L3 filtered cell results) to the UE. 
Proposal 2: For temporal domain prediction case A and case B, the network can configure an expected measurement interval to the UE.
Proposal 3: For temporal domain prediction case B, the network can configure the number of continuous measurements to the UE. And the existing measurement framework is enhanced to support the UE to report the measurement results of multiple historical time instances in one measurement report.
Proposal 4: For inference in network side model, enhancement is needed to allow the UE to not report RSRP/RSRQ results of all serving cells that configured with servingCellMO.