Auto-Generated Summary

> Provide an extensive summary of the agreements by topic. No need to mention CRs.

Summary of Agreements by Topic

Beam Management

Inference Report Content:

For BM-Case1, Top K beam(s) report is configured in inference report configuration.
Report format for BM-Case1 includes CRI/SSBRI mapped to RSRP values, with differential RSRP relative to the largest predicted RSRP.
For BM-Case2, report format includes time instance indicators, CRI/SSBRI, and RSRP values for multiple time instances.

Resource Configuration:

Only always-on SSB and P/SP CSI-RS are supported for data collection.
CSI-ReportConfig can be used for configuring resources for data collection without CSI report.

Monitoring:

Type 1 Option 2 monitoring supports periodic CSI-RS, semi-persistent CSI-RS, and SSB.
Performance metric for Top 1 or Top K beam prediction accuracy is defined based on linked inference reports and monitoring resources.
Monitoring report types are supported for combinations of periodic, semi-persistent, and aperiodic inference reports.

Processing Unit:

Options for processing include dedicated AI/ML PU, legacy CPU, or both, reported by UE capability.

Time Instance Configuration:

For BM-Case2, time gap between consecutive future time instances is configured by RRC ([10ms, 20ms, 40ms, 80ms, 160ms]), and N = [1, 2, 4, 8].

Positioning Accuracy Enhancement

Measurement Types:

For Case 3b, enhanced measurement includes Nt' values selected from Nt consecutive values with timing granularity T.
No offset is transmitted from gNB to LMF for Case 3b.

Training Data Collection:

Part B label includes timing information provided by LMF to gNB using "Timing Measurement Quality" in TS 38.455.
LOS/NLOS indicator provided using existing IE LOS-NLOS-Indicator-r17.

Monitoring:

For Case 1, monitoring outcome includes an indication that the target UE cannot perform the positioning method.

CSI Prediction

Inference:

Rel-18 CSI framework is reused for inference.
CSI-RS resource types for CMR include periodic, semi-persistent, and aperiodic CSI-RS.
New RRC parameter introduced to distinguish AI-CSI prediction and non-AI CSI prediction.

Training Data Collection:

CSI-ReportConfig can be used for configuring resources for data collection without CSI report.
Aperiodic CSI-RS resource for CMR is not supported.

Performance Monitoring:

Type 3 monitoring is supported, with SGCS confirmed as the performance metric.
Reporting contents for monitoring include SGCS values calculated based on predicted CSI and ground truth CSI.

Processing Unit:

Dedicated AI/ML PU introduced for simultaneous processing of multiple CSI reports across AI/ML use cases.

CSI Compression

Inter-Vendor Collaboration:

Options 3a-1 and 4-1 are feasible, including with or without NW-side target CSI sharing.
Fully standardized reference model associated with an ID for pairing-related discussions.

Quantization:

Standardized quantization configurations include scalar/vector quantization, segment size, and codebook size.
Exchanged quantization codebook may include parameterized formulas.

Scalability:

Scalable model structure specification is feasible for temporal domain Case 0.
Evaluations confirm small loss in SGCS for scalable structures compared to dedicated structures.

Monitoring:

NW-side monitoring with target CSI reporting is feasible, with potential benefits from enhanced CSI codebooks.

Other Aspects of AI/ML Model and Data

Model Identification:

ID associated with exchanged dataset or model parameters for pairing-related discussions.
Same ID used for training, applicability inquiry, inference configuration, and NW-side data collection.

Model Transfer/Delivery:

Study progress on model identification and transfer/delivery is sufficient.
No consensus in Rel-19 on the need for standardized solutions for one-sided model delivery.

Processing Unit:

Dedicated AI/ML PU introduced for quantifying simultaneous processing across AI/ML use cases.

This summary captures the key agreements across topics discussed in the document.