AI-based channel feedback locks its data formats but not its quality metric

Phones that use AI to compress channel feedback will report signal quality (CQI) using whatever method their chipset maker prefers — an outcome Apple, NTT DOCOMO, Huawei, Ericsson, CATT, ZTE and Samsung all objected to. They wanted one fixed calculation so operators could compare phones fairly; Qualcomm, vivo, OPPO, Nokia and eight others argued no single method works because the phone never sees the exact precoding matrix the network reconstructs from its report. The dispute repeated across three separate meeting sessions without resolution, leaving RAN4 — the group that writes conformance tests — unable to define what counts as an acceptable CQI implementation.

The scale of the new feature is settled: AI-compressed feedback will cover up to 128 antenna ports, rank 4 and 19 subbands, matching the massive-MIMO configurations already used in 5G-Advanced. Reports will travel in the same two-part uplink control structure as legacy high-resolution feedback, with CQI and rank in the better-protected first part and the AI-compressed detail in the second — reusing a design pattern from Enhanced Type II codebooks rather than inventing a new one.

For training data collected by the network from phones, companies picked amplitude-phase quantization (3 bits for amplitude, 4 bits for phase) over the alternative real-imaginary format, reusing the quantization tables already defined for 5G's Enhanced Type II codebook. When no custom codebook is exchanged, phones fall back to a standard set of values — 0.125, 0.375, 0.625 and 0.875 for 2-bit quantization — guaranteeing that phones and networks from different vendors interpret the same compressed numbers the same way.

Model monitoring will run on fast physical-layer (L1) signaling rather than the slower RRC-based (L3) layer: Nokia, Apple, Samsung, MediaTek, Xiaomi and LG argued L3 signaling lacks the latency control needed to catch a degrading AI model before it hurts link performance. Left unresolved is whether phones need their own beam-restriction logic (Codebook Subset Restriction) when preparing feedback for AI compression — Huawei, CATT, ZTE, KT, OPPO and vivo say the network can filter unwanted beams on its own, while Qualcomm, Ericsson, Nokia and Apple want it studied first, since it existed in earlier 5G releases.

RAN1 #126 will run the beam-restriction study companies agreed to at this meeting, comparing signal leakage and throughput when phones restrict beams themselves against when the network filters alone, using four restricted beams as the test case. The result decides whether AI-based CSI feedback needs new phone-side complexity for beam handling or whether that job stays entirely on the network side — a question that also bears on how RAN4 eventually tests these devices.