R1-2407653
discussion
Discussion on AI/ML for beam management
From Huawei
Summary
This Huawei document presents a comprehensive analysis of AI/ML for beam management in NR, containing 38 detailed proposals and 9 observations covering data collection, inference procedures, performance monitoring, and UE capability reporting for both network-side and UE-side AI/ML models.
Position
Huawei advocates FOR pragmatic reuse of existing CSI framework and signaling mechanisms while pushing AGAINST overly complex new procedures. They strongly support expanding beam measurement capabilities (up to 256 beams vs legacy 64) and unified solutions that work across BM-Case 1/2. They oppose introducing dedicated AI/ML activation signaling and extending TCI state mechanisms for future time instances, arguing these add complexity without significant benefit. Their stance emphasizes cell-specific rather than cross-cell associated IDs to preserve proprietary information and reduce network management burden.
Key proposals
- Proposal 1 (Sec 2.1.1): Enable UE to perform CSI measurements on larger beam sets by either using multiple resource sets with legacy size (up to 64) or one resource set with increased size up to 256 resources
- Proposal 5 (Sec 2.2.2.1): Support adaptive number of beams based on threshold - report L1-RSRPs and beam info of up to M beams within X dB gap to largest measured L1-RSRP, with gNB-configured X and M>4
- Proposal 8 (Sec 2.3.1): For UE-side model measurements, indicate the purpose (training, inference, monitoring, non-AI/ML) to UE since corresponding behavior and report content differ for each purpose
- Proposal 11 (Sec 2.4.1): Study associated ID subject to cell-specific manner rather than globally unique or area-unique ID to avoid network burden and proprietary disclosure risks
- Proposal 16 (Sec 3.1.3): Do not support extending Rel-17 TCI state activation methods for N future time instances in BM-Case 2 due to insignificant overhead savings and substantial implementation complexity
- Proposal 19 (Sec 3.3.3.1): Support reporting predicted beam IDs/RSRPs of more than 4 beams for UE-side inference to improve accuracy, generalization performance, and maintain symmetry with NW-side model capabilities
- Proposal 23 (Sec 3.3.4.2): For BM-Case 2 observation window, investigate at least P/SP CSI-RS resources while studying aperiodic CSI-RS applicability due to long observation window constraints
- Proposal 25 (Sec 4.1): No need to specify monitoring procedure/metric for NW-side model except for data collection since monitoring is implementation-dependent
- Proposal 27 (Sec 4.2.2): Consider L1 signaling with higher priority for UE-side model monitoring Type 1 due to more stringent latency requirements compared to training
- Proposal 33 (Sec 5.1.2): Reuse legacy CPU mechanism for AI/ML-based CSI processing where low priority CSI is not updated if required CPU exceeds supported CPU capacity
- Proposal 35 (Sec 5.1.3): Achieve AI/ML functionality activation by reusing existing P-CSI/SP-CSI/A-CSI report signaling rather than introducing dedicated activation procedures
- Proposal 38 (Sec 5.2): For UE capability reporting, discuss supported configurations including data samples needed for training/monitoring, RS/CSI report configurations, Top-K values, and target performance/robustness information