vivo · 9.1.3

Vivo identifies TXRU virtualization mapping as a critical NW-side additional condition that causes significant generalization performance loss, particularly at 60km/h or with high proportions of outdoor users. They propose adopting the associated ID solution from the Beam Management use case to ensure consistency between training and inference data for UE-side models. For CSI feedback mechanisms, vivo prefers separate processing unit pools for AI-based CSI and legacy CSI to avoid interference and simplify UE design, rather than a joint pool. Regarding performance monitoring, vivo supports using SGCS as the metric, calculated specifically between the precoding vectors of the predicted and real channels to account for codebook quantization effects. They prioritize Type 3 and Type 1 monitoring configurations over Type 2, citing lower overhead and complexity, and propose reusing the dedicated CSI report configuration mechanism defined for AI beam management Case 2.

vivo identifies TXRU virtualization mapping as a critical NW-side additional condition that significantly impacts the generalization performance of UE-sided CSI prediction models. They present simulation evidence showing that mismatches in TXRU mapping between training and inference data cause substantial SGCS loss, particularly in high-speed (60km/h) and 100% outdoor user scenarios, with losses reaching up to -44.4%. To address this consistency issue, vivo proposes following the solution based on associated ID as established in the Beam Management (BM) use case. This approach ensures that data collected under specific NW-side additional conditions, such as varying TXRU mappings due to energy saving shutdowns, are linked to unique identifiers so UEs can select the correct corresponding models for inference.

Vivo advocates FOR addressing training-inference consistency issues in AI/ML CSI prediction by adopting associated ID solutions from beam management use cases, and FOR recognizing TXRU mapping as a critical network-side condition. They push AGAINST ignoring the significant performance degradation (up to 44.4%) caused by TXRU virtualization mapping mismatches between training and inference phases.