R1-2500274 discussion

Discussion on specification support for beam management

From CMCC
Status: not treated
WI: NR_AIML_air
Agenda: 9.1.1
Release: Rel-19
Source: 3gpp.org ↗
CMCC's prior position on 9.1.1 at RAN1#119 · AI-synthesized, paraphrased
verify sources →
Proposes that L1 signaling be supported for NW-sided training data collection and that Top-K beam sweeping be supported for NW-side inference to enhance prediction accuracy. Requires that for UE-side models, the overall CPU be separately counted between legacy CSI reporting and AI/ML-based CSI reporting, while being shared among AI/ML features. Prefers Option 1 for BM-Case 2 differential RSRP reporting, which includes CRI of top-K predicted beams per instance. Supports dedicated resource sets and report configurations for UE-side model monitoring (Type 1 Option 2), linking the monitoring configuration to an inference report configuration via CSI-ReportConfig ID. Proposes that the granularity of UE capability reporting for AI/ML be at the sub-use case level, including details on Set A/B size, RS type, and model outputs.

Summary

CMCC presents 55 proposals and 4 observations regarding the specification impact of AI/ML-based beam management in NR, covering data collection, inference, and monitoring for both NW-side and UE-side models. The document addresses critical aspects such as Set A/B configuration, Top-K beam sweeping necessity, reporting content for inference results, and performance monitoring mechanisms.

Position

CMCC proposes supporting Type 1 and Type 3 data collection contents for NW-side training, emphasizing flexibility for classification and regression models. They require that RS associated with TCI state indication must be measured at least once before application, ensuring QCL parameter validity. For UE-side inference, they prefer Option 1 for reference time determination in BM-Case 2 and support both predicted and measured RSRP reporting options for BM-Case 1. CMCC argues that applicable functionality determination should be up to UE implementation rather than strictly dependent on associated ID matching, allowing for model generalization. They propose separate CPU counting for legacy vs. AI/ML CSI reporting to prevent performance degradation of legacy features. For monitoring, they prefer statistical beam prediction accuracy over N instances and support monitoring sets as down-sampled subsets of Set A to reduce overhead.

Key proposals

Your notes

Private to your account