AI/ML Life Cycle Management for Interoperable AI Native RAN

I. INTRODUCTION

• AI/ML enhances nonlinear RAN tasks including beam management, CSI feedback, and mobility prediction, while deployment raises model drift, version consistency, performance, and cross-vendor compatibility challenges.These challenges motivate standardized LCM to sustain model performance across heterogeneous and multi-vendor networks,.Source: §I · p. 1
• Rel-17 introduced generic AI/ML management for model training and retraining, and Rel-18 expanded it into a broader life-cycle framework.The management capabilities were integrated into existing OAM processes and later extended to additional life-cycle phases.Source: §I · p. 1
• Rel-19 addresses AI/ML-assisted CSI prediction, beam management, and positioning through data collection, applicability assessment, performance monitoring, and RRC configuration.The core Rel-19 air-interface work was completed in 2025, with performance evaluation and UE conformance activities continuing through 2026.Source: §I · p. 1
• Rel-20 extends RAN standardization toward two-sided AI/ML operation, including UE–gNB CSI compression, while ongoing work targets inter-vendor collaboration and model operation.These capabilities are positioned as a foundation for AI-native transceivers and broader AI-RAN functions in 6G.Source: §I · pp. 1–2
• Rel-18 established parallel SA5 OAM and RAN air-interface tracks: the former governs longer-timescale training and deployment, while the latter handles time-sensitive operation and interoperability.Rel-19 and Rel-20 progressively add normative air-interface support, two-sided operation, and interoperability mechanisms.Source: §I · pp. 1–2

A. Functional Framework

• The framework extends the five TR 38.843 logical functions with interactions among monitoring, KPI derivation, environment detection, adaptation, and control elements.Additional elements support deriving appropriate control actions across the functional blocks.Source: A. Functional Framework · pp. 2–3
• A fast RAN-local controller handles monitoring configuration, applicability assessment, model activation or deactivation, switching, and temporary fallback, while slower OAM manages training and repositories.The Management function is logical and does not prescribe one physical entity or common timescale.Source: A. Functional Framework · pp. 2–3
• Monitoring and KPI derivation convert observations into indicators of model accuracy, robustness, and trustworthiness, while environment detection compares current distributions with training data.Adaptation provides fast, lightweight RAN-local parameter adjustment alongside longer-timescale offline training or retraining.Source: A. Functional Framework · p. 3
• Decision logic combines inference, monitoring, and environment information to maintain a model, switch models, or fall back to a non-AI procedure; persistent degradation can trigger OAM replacement or retraining.The controller coordinates storage, training or adaptation, and inference through internal or external signaling.Source: A. Functional Framework · p. 3

B. Performance Monitoring in RAN

• RAN-based monitoring reacts to physical-environment changes on smaller timescales than OAM performance management, enabling evaluation of whether an active AI/ML function remains applicable.The monitoring requirement study underpins Rel-19 procedures and requirements.Source: B. Performance Monitoring in RAN · pp. 3–4
• Configured CSI-RS, DM-RS, or SRS monitoring resources provide reference measurements for ground-truth comparison with inference reports.In beam prediction, the UE measures signal strength or CSI on predicted beams, and the gNB associates measurements with the corresponding reports.Source: B. Performance Monitoring in RAN · p. 4
• Beam prediction accuracy counts monitoring intervals in which at least one of the M best measured beams appears among the top-K predicted beams.The metric is defined over the N most recent monitoring intervals [7].Source: B. Performance Monitoring in RAN · p. 4
• CSI prediction quality uses SGCS to quantify alignment between predicted and true precoders, while positioning accuracy uses the 90% CDF percentile of horizontal accuracy.Metrics may be computed at the UE, gNB, or jointly through UE-assisted reporting, supporting continued operation, switching, deactivation, or fallback.Source: B. Performance Monitoring in RAN · p. 4

C. Case Study: CSI Prediction and Compression

• CSI prediction and compression reduce the overhead or channel-aging effects associated with periodic CSI reporting, with prediction standardized in Rel-19 and two-sided compression addressed in Rel-20.Typical CSI reports occur every 40–160 ms, creating channel aging for mobile UEs.Source: C. Case Study: CSI Prediction and… · p. 4
• The UE collects CSI-RS measurements, performs local prediction or compression, and reports PMI or a latent message with optional CQI and RI extensions to the gNB.The RAN-local controller uses monitoring results to retain, deactivate, switch, or fall back from the active model.Source: C. Case Study: CSI Prediction and… · p. 4
• For CSI prediction, the gNB schedules reference signals for ground-truth CSI and Rel-19 supports UE-assisted SGCS reporting through the CSI reporting framework.CSI compression monitoring schemes remained under discussion in Rel-20.Source: C. Case Study: CSI Prediction and… · p. 4
• Approximately 10% SGCS improvement resulted when the LCM controller switched from the UMi-trained model to the UMa-trained model after an environment change.Without LCM, the system retained the UMi model; the case study also shows that SGCS can rise under a mismatched model when UMa channels are less complex.Source: C. Case Study: CSI Prediction and… · p. 4, Figure 4 · p. 5

• Two-sided functions require model management to preserve interoperability between independently developed UE- and gNB-side models.Different vendor interfaces make these models difficult to connect seamlessly, creating hurdles for management, training, and inference.Source: §III · p. 5
• CSI compression uses a neural UE encoder to send a latent message that a gNB decoder reconstructs into the precoder.Source: §III · p. 5
• Joint training provides high fidelity for encoder-decoder CSI compression but is impractical in multi-vendor deployments.Source: §III · p. 5
• 3GPP is studying training collaboration mechanisms that let vendors train corresponding models independently while preserving compatibility with the counterpart model.The decoder must correctly reconstruct the intended precoder from the encoder output.Source: §III · p. 5

Collaboration Modes

• 3GPP identified joint training at one entity, joint training split across network and UE sides, and separate training of the encoder and decoder.Joint training uses the same forward-backward loop, while separate training assigns the encoder and decoder to different vendors or sides.Source: A. Collaboration Modes · pp. 5–6
• Rel-19 focused mainly on Type 3 separate training because Type 1 needs less interface design and Type 2 air-interface study was deprioritized.Source: A. Collaboration Modes · p. 6
• Direction I provides standardized reference models and parameters, while Direction II provides datasets or encoder parameters for UE-side offline engineering.The dataset may contain {target CSI, feedback CSI}; target CSI is the UE-estimated precoder and feedback CSI is the latent message sent to the gNB.Source: A. Collaboration Modes · p. 6
• Under Direction I, vendors train their model against a fixed counterpart reference model; under Direction II, vendors train from labeled target-CSI and feedback-CSI data.A UE vendor can train an encoder against a fixed reference decoder, while a gNB vendor can train a decoder against a fixed reference encoder.Source: A. Collaboration Modes · p. 6
• Rel-20 considers standardized test-procedure CSI models as references in Direction I and dataset exchange between BS and UE vendors in Direction II.An associated ID pairs encoder and decoder models by confirming use of the same reference model, dataset, or parameter set.Source: A. Collaboration Modes · p. 6

B. Proposed Reference Model Derivation and Selection Procedure

• Careful reference-model derivation is foundational because the selected models shape commercial encoder and decoder development for Directions I and II.Source: B. Proposed Reference Model Derivation… · p. 6
• Reference-model derivation begins by defining the simulation setup and encoder-input dataset generation procedure for candidate-model evaluation.Source: B. Proposed Reference Model Derivation… · p. 7
• The procedure then selects a model backbone, training hyperparameters, and test decoder reference models.Candidate backbones include CNNs, transformers, or variants chosen using initial performance, robustness, and complexity assessment.Source: B. Proposed Reference Model Derivation… · p. 7
• Reference models are evaluated using achievable performance, complexity, and robustness criteria.Performance may use SGCS, throughput, or block error rate; complexity includes FLOPs and model storage size.Source: B. Proposed Reference Model Derivation… · p. 7
• Robustness requires satisfactory performance when a reference decoder or encoder connects with a collaboratively trained counterpart using a different backbone.The collaboration procedure targets strong system performance with reasonable implementation cost and a flexible vendor design space.Source: B. Proposed Reference Model Derivation… · p. 7

• The section frames key technical challenges and future directions for AI-native RAN design.Source: §IV · p. 7

A. Design Challenges in Resource-Efficient Performance Monitoring

• More monitoring resources improve AI/ML performance tracking but reduce the overhead savings delivered by AI/ML models.In beam prediction, transmitting more beams in the predicted-beam set improves tracking but diminishes reference-signal overhead reduction.Source: A. Design Challenges in… · p. 7
• Monitoring-resource design must intelligently trade off monitoring performance against resource consumption.Source: A. Design Challenges in… · p. 7
• Two-sided CSI compression complicates monitoring because the lightweight UE encoder cannot assess downlink quality without access to the gNB-side reconstructed precoder.Replicating the decoder at the UE would close this assessment gap but significantly increase device complexity.Source: A. Design Challenges in… · p. 7

B. Input Distribution Shifts and Their Impact on Model Reliability

• Input-statistics monitoring can detect environmental drift when operating-data characteristics deviate from training data and model performance may degrade.The LCM controller is relied upon to recover performance, while existing 3GPP work has focused mostly on output-based monitoring.Source: B. Input Distribution Shifts and Their… · p. 7
• Input-distribution descriptors can both detect drift and retrieve stored models whose KPI performance is expected to remain similar in adjacent statistical regions.The stored-model population should sufficiently cover the statistical input space.Source: B. Input Distribution Shifts and Their… · p. 7
• The LCM controller must fuse input analysis, performance monitoring, and assistance signaling collected across heterogeneous functions and time slots to make timely decisions.Source: C. Key Challenges in Intelligent LCM… · pp. 7–8
• Performance degradation does not necessarily indicate model misalignment, because conditions such as an SNR drop can reduce CSI prediction accuracy even when channel dynamics remain properly captured.After identifying environmental mismatch, the controller must choose among switching models, adapting parameters, or falling back to non-AI/ML algorithms.Source: C. Key Challenges in Intelligent LCM… · p. 8

D. Rethinking Interoperability vs. Flexibility in AI-native RAN

• Standardizing reference models or their generated datasets can guarantee two-sided CSI-feedback interoperability from both training and inference perspectives.Source: D. Rethinking Interoperability vs.… · p. 8
• Reference-model standardization limits AI/ML flexibility to adapt across environments and channel types, despite that flexibility being a major advantage of AI/ML models.Source: D. Rethinking Interoperability vs.… · p. 8
• Standardized models may delay adoption of newer architectures because the standards must be updated before vendors can use them, motivating reconsideration of interoperability mechanisms.3GPP has developed related SA5 and RAN life-cycle tracks, while inter-vendor collaboration and two-sided operation remain under development.Source: D. Rethinking Interoperability vs.… · p. 8, §V · p. 8