The overall RRC design for Artificial Intelligence and Machine Learning in 3GPP Release 19 represents a fundamental shift from static, rule-based signaling to a dynamic, data-driven framework. This architecture is built on a non-critical extension of the UE capability reporting system, specifically nested within the Release 19 capability structures to ensure backward compatibility with legacy 5G NR deployments. The design philosophy centers on model observability and control, allowing the network to manage the entire lifecycle of an AI model—from initial capability discovery and activation to continuous performance monitoring and eventual fallback.
At the heart of this RRC design is the abstraction of AI models into specific functional blocks, such as channel state information prediction and beam management. Rather than standardizing the internal black-box logic of a neural network, 3GPP focuses on standardizing the interfaces: the specific input data (Set B) provided to the model, the expected output format (Set A), and the computational budget required for processing. This approach preserves vendor innovation in algorithm design while ensuring that the gNodeB and UE share a mutual understanding of the model's intent and reliability.
The signaling framework is further refined by a sophisticated resource management protocol that accounts for the unique computational demands of AI inference. By incorporating parameters for CPU pooling across carrier components and introducing relaxation timelines, the RRC design allows the UE to negotiate the necessary latency "buffer" required to complete complex mathematical operations. This ensures that the integration of AI does not compromise the strict timing requirements of the 5G air interface, but instead enhances it through proactive, multi-slot predictions of the radio environment
UE Capability
The introduction of AIML-Parameters-r19 shows the transition toward an AI/ML-based air interface. Instead of the network trying to infer how the UE sees the radio environment, the UE can now use trained models to support functions such as beam management, CSI feedback, and positioning.
The ASN.1 shown in this section follows the usual 3GPP non-critical extension structure. By placing these fields under UE-NR-Capability-v1900, 3GPP is making it clear that Release 19 is the formal starting point for standardized AI/ML capability signaling.
This matters because AI and ML in mobile networks cannot work as an uncontrolled black box. The network needs to know whether the UE is capable of AI-related processing, whether the UE’s model is still valid in the current environment, and whether the UE can contribute useful data for improving future models. This capability framework makes AI and ML a controlled and predictable part of radio resource management rather than just an experimental feature.
UE-NR-Capability-v1860 ::= SEQUENCE {
ntn-CHO-OnlyLocationTimeTrigger-r18 ENUMERATED { supported } OPTIONAL,
nonCriticalExtension UE-NR-Capability-v1900 OPTIONAL
}
UE-NR-Capability-v1900 ::= SEQUENCE {
aiml-Parameters-r19 AIML-Parameters-r19 OPTIONAL,
ue-RadioPagingInfo-r19 OCTET STRING (CONTAINING UE-RadioPagingInfo-r19) OPTIONAL,
ntn-VSAT-AntennaTypeKuBand-r19 ENUMERATED { electronic, mechanical } OPTIONAL,
ntn-VSAT-MobilityTypeKuBand-r19 ENUMERATED { fixed, mobile } OPTIONAL,
ntn-ERedCap-FR1-r19 ENUMERATED { supported } OPTIONAL,
onDemandSIB1-r19 ENUMERATED { supported } OPTIONAL,
onDemandPosSIB-ConnectedCtrlParam-r19 ENUMERATED { supported } OPTIONAL,
ntn-Redirection-r19 ENUMERATED { supported } OPTIONAL,
drx-PreferenceCellDTX-DRX-r19 ENUMERATED { supported } OPTIONAL,
lpwus-SupportedBandsIdleInactiveOFDM-r19 SEQUENCE (SIZE (1..maxBands)) OF FreqBandIndicatorNR OPTIONAL,
lpwus-SupportedBandsIdleInactiveOOK-r19 SEQUENCE (SIZE (1..maxBands)) OF FreqBandIndicatorNR OPTIONAL,
nonCriticalExtension UE-NR-Capability-v1920 OPTIONAL
}
AIML-Parameters-r19 ::= SEQUENCE {
applicabilityReportingCSI-r19 ENUMERATED { supported } OPTIONAL,
applicabilityReportingOther-r19 ENUMERATED { supported } OPTIONAL,
loggedDataCollection-r19 ENUMERATED { supported } OPTIONAL,
eventBasedLoggedDataCollection-r19 ENUMERATED { supported } OPTIONAL,
dataThresholdAvailabilityIndication-r19 ENUMERATED { supported } OPTIONAL
}
Followings are short descriptions on important parameters
-
applicabilityReportingCSI-r19 : Indicates whether the UE can report the applicability of its AI model specifically for Channel State Information. This helps the network determine whether the current AI model is reliable under the current radio conditions. -
applicabilityReportingOther-r19 : Indicates whether the UE can report model applicability for other AI/ML use cases, such as beam management or positioning. -
loggedDataCollection-r19 : Confirms that the UE can collect and store radio signal data for use as training data for machine learning models. -
eventBasedLoggedDataCollection-r19 : Indicates that the UE can trigger data logging only when specific radio events occur, such as a handover failure or a signal drop. -
dataThresholdAvailabilityIndication-r19 : Allows the UE to inform the network that it has collected enough data to be useful for model training or model updates.
CSI Prediction
The core idea here is that the UE does not just measure the channel as it exists at the current moment. It uses an AI model to predict how the channel will evolve in the near future, especially in high-mobility conditions where Doppler effects become significant.
Overall, the flow is straightforward. The UE first informs the network about its AI processing capability, including its CPU budget and the resource types it supports. The network then configures a prediction or monitoring session. The UE uses the configured CSI-RS resources, including parameters such as N4, as input to its AI model. It then generates predicted CSI and reports it back to the network. If additional inference time is needed, the UE can rely on the relaxation timeline so that the report is delayed just enough to allow the neural-network-based processing to complete properly.
predictionConfiguration-r19 CHOICE {
csi-InferencePrediction-r19 NULL,
configurationForBM-PredictionAndDataCollection-r19 SEQUENCE {
resourcesForChannelPrediction-r19 CSI-ResourceConfigId,
associatedIdForChannelPrediction-r19 AssociatedId-r19 OPTIONAL, -- Need R
associatedIdForChannelMeasurement-r19 AssociatedId-r19 OPTIONAL, -- Need R
nrofReportedPredictedRS-r19 ENUMERATED { n1, n2, n3, n4 } OPTIONAL, -- Need R
nrofTimeInstance-r19 ENUMERATED { n1, n2, n4, n8 } OPTIONAL, -- Need R
timeGap-r19 ENUMERATED { ms10, ms20, ms40, ms80, ms160,
spare3, spare2, spare1 } OPTIONAL, -- Need R
...
},
configurationForBM-Monitoring-r19 SEQUENCE {
refToPredictionConfig-r19 CSI-ReportConfigId,
nrofBestBeamForMonitoring-r19 ENUMERATED { n1, n2 } OPTIONAL, -- Need R
nrofTransmissionOccasion-r19 ENUMERATED { n1, n3, n7, n15 } OPTIONAL, -- Need R
timeInstanceForRS-PAI-r19 ENUMERATED { n1, n2, n3, n4, n5, n6, n7, n8 } OPTIONAL, -- Need R
mappingToResourcesForChannelPrediction-r19
BIT STRING (SIZE (1..maxNrofNZP-CSI-RS-ResourcesPerSet)) OPTIONAL, -- Need R
...
},
configurationForCSI-Monitoring-r19 SEQUENCE {
refToPredictionConfig-r19 CSI-ReportConfigId,
timeInstanceForCSI-PAI-r19 ENUMERATED { n1, n2, n3, n4, n5, n6, n7, n8 } OPTIONAL, -- Need R
...
}
} OPTIONAL, -- Need R
reportQuantity-r19 CHOICE {
noneBM-r19 NULL,
noneCSI-r19 NULL,
p-CRI-r19 NULL,
p-SSB-Index-r19 NULL,
p-CRI-RSRP-r19 NULL,
p-SSB-Index-RSRP-r19 NULL,
rs-PAI-r19 NULL,
csi-PAI-r19 NULL,
cjtc-Dd-r19 NULL,
cjtc-F-r19 NULL,
cjtc-P-r19 NULL,
cjtc-Dd-F-r19 NULL,
cli-RSSI-r19 NULL,
cli-SRS-RSRP-r19 NULL
} OPTIONAL -- Need R
CA-ParametersNR-v1900 ::= SEQUENCE {
aiml-CSI-PredictionDopplerPerBC-r19 CodebookParametersCSI-PredictionDoppler-r19 OPTIONAL,
-- R1 58-0-1: CSI report framework for UE-side inference
aiml-CSI-ReportPerBC-r19 SEQUENCE (SIZE (1..2)) OF CPU-PoolInfo-r19 OPTIONAL,
-- R1 58-3-1: CSI prediction for UE-sided inference when N4=1
aiml-CSI-PredictionPerBC-r19 SEQUENCE {
supportedCSI-RS-ResourceList-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16),
scalingFactor-r19 ENUMERATED { n1, n2, n4 },
numberOfOccupiedCPU-r19 INTEGER (0..8),
numberOfOccupiedCPUx-r19 INTEGER (0..8),
relaxationTimelineT-r19 SEQUENCE {
scs15kHz-r19 ENUMERATED { n14, n28, n56, n112 },
scs30kHz-r19 ENUMERATED { n28, n56, n112, n224 },
scs60kHz-r19 ENUMERATED { n56, n112, n224, n448 },
scs120kHz-r19 ENUMERATED { n112, n224, n448 },
scs480kHz-r19 ENUMERATED { n448, n896, n1792 },
scs960kHz-r19 ENUMERATED { n896, n1792 }
},
occupiedResourcePool-r19 INTEGER (1..2),
inferenceReportType-r19 ENUMERATED { aperiodic, semiPersistent }
} OPTIONAL,
-- R1 58-3-1a-1: DD unit size when A-CSI-RS is configured for CMR N4>1 for UE side inference of CSI prediction
aiml-CSI-PredictionUnitDurationDD-PerBC-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-2: CSI prediction for UE-sided inference when N4>1
aiml-CSI-PredictionN4PerBC-r19 SEQUENCE {
supportedCSI-RS-ReportSettingAcrossCC-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF
SupportedCSI-RS-ReportSetting-r18,
supportedCSI-RS-ReportSettingOneReport-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF
SupportedCSI-RS-ReportSetting-r18,
numOccupiedCPU-r19 INTEGER (0..8),
numOccupiedCPUx-r19 INTEGER (0..8),
occupiedPool-r19 ENUMERATED { p1, p2 }
} OPTIONAL,
-- R1 58-3-4: UE side data collection for CSI prediction
aiml-CSI-PredictionUE-DataCollectionPerBC-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-5: Performance monitoring for CSI prediction model
aiml-CSI-PredictionMonitoringPerBC-r19 SEQUENCE {
suppportedCSI-RS-ResourceList-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16),
numOccupiedCPU-r19 INTEGER (1..2)
} OPTIONAL
}
CodebookParametersCSI-PredictionDoppler-r19 ::= SEQUENCE {
-- R1 58-3-1b: Maximum number of aperiodic CSI-RS resources that can be configured in the same CSI report setting for Rel-16-based
-- doppler measurement for UE side inference of CSI prediction
maxNumberOfAperiodic-CSI-RS-Resource-r19 ENUMERATED { n4, n8, n12 } OPTIONAL,
-- R1 58-3-1-2: Support R=2 for Rel-16-based doppler codebook for UE side inference of CSI prediction
eType2DopplerR2-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL,
-- R1 58-3-1-3: Support X=1 based on first and last slot of WCSI, for Rel-16-based doppler codebook for UE side inference of CSI prediction
eType2DopplerX1-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-1-3a: Support X=2 CQI based on 2 slots for Rel-16-based doppler codebook for UE side inference of CSI prediction
eType2DopplerX2-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-1-4: support of l = (n – nCSI,ref ) for CSI reference slot for Rel-16 based doppler codebook for UE side inference of CSI prediction
eType2DopplerL-N4D1-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-1-5: Support of L=6 for Rel-16 based doppler codebook for UE side inference of CSI prediction
eType2DopplerL6-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-1-6: Support of rank equals 3 and 4 for Rel-16 based doppler codebook for UE side inference of CSI prediction
eType2DopplerR3R4-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-1-7: Active CSI-RS resources and ports for mixed R16 based doppler codebook for CSI prediction via UE side model with
-- other codebooks in any slot
codebookComboParameterMixedTypePrediction-r19 SEQUENCE {
type1SP-Type1SP-N4-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL,
type1SP-eType2SP-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL,
type1SP-eType2SP-N4-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL,
type1SP-N4-eType2SP-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL,
type1SP-N4-eType2SP-N4-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL,
eType2SP-eType2SP-N4-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL
} OPTIONAL
}
CPU-PoolInfo-r19 ::= SEQUENCE {
maxNumCPU-PerCC-r19 INTEGER (1..8),
maxNumCPU-AllCC-r19 INTEGER (1..32)
}
Followings are short descriptions on important parameters
-
predictionConfiguration-r19 defines how the UE and the network coordinate AI-based prediction activity. It is configured per CSI-ReportConfig and works together with reportQuantity-r19 : BM prediction or data collection requires reportQuantity-r19 set to p-CRI, p-SSB-Index, p-CRI-RSRP, p-SSB-Index-RSRP or noneBM, BM monitoring requires rs-PAI, and CSI prediction monitoring requires csi-PAI. -
csi-InferencePrediction-r19 represents an inference mode in which the UE runs a pre-trained AI model, predicts CSI, and reports the result back to the gNB. -
Monitoring configurations allow the network to compare predicted CSI with the actual channel condition, so it can evaluate prediction accuracy and monitor AI model performance over time. -
CPU-PoolInfo-r19 defines the amount of processing resource that the UE can dedicate to AI and ML tasks. -
maxNumCPU-AllCC-r19 indicates how much AI processing capacity the UE can share across multiple component carriers in carrier aggregation. -
relaxationTimelineT-r19 tells the gNB how much extra processing time, measured in slots, the UE may need to complete AI inference for different subcarrier spacings. -
CodebookParametersCSI-PredictionDoppler-r19 handles Doppler and high-mobility support for AI-based CSI prediction. -
eType2DopplerR3R4-r19 indicates support for higher-rank transmission cases, such as rank 3 and rank 4, even when AI-based prediction is used. -
X1 and X2 related parameters define the time window, in slots, that the AI model uses to analyze Doppler and frequency shift behavior. -
scalingFactor-r19 defines the complexity or size of the AI model. A larger value usually means a larger or more precise model. -
inferenceReportType-r19 defines how the UE reports AI-based results, such as by aperiodic or semi-persistent reporting. -
supportedCSI-RS-ResourceList indicates which CSI-RS resources the UE’s AI model is capable of processing. -
codebookComboParameter supports hybrid operation where some CSI reporting uses AI-based prediction while other parts still use legacy Type-I or Type-II codebook methods. -
resourcesForChannelPrediction-r19 indicates the CSI-ResourceConfig whose CSI-RS resources are used as the AI model input for channel prediction. -
associatedIdForChannelPrediction-r19 / associatedIdForChannelMeasurement-r19 link the prediction report to the associated resource configurations so that the network can correlate the predicted channel with the measured channel. -
nrofReportedPredictedRS-r19, nrofTimeInstance-r19 and timeGap-r19 define how many predicted reference signals are reported, how many future time instances are predicted and the time spacing between those instances. -
refToPredictionConfig-r19 indicates the linked CSI-ReportConfigId corresponding to a prediction report configuration, so that a monitoring report can be tied to the prediction it is evaluating. -
nrofBestBeamForMonitoring-r19, nrofTransmissionOccasion-r19 and timeInstanceForRS-PAI-r19 control how many best beams and transmission occasions are used for beam prediction accuracy monitoring, and which predicted time instance the RS prediction accuracy indicator (rs-PAI) refers to. -
timeInstanceForCSI-PAI-r19 indicates which predicted time instance the CSI prediction accuracy indicator (csi-PAI) refers to. -
reportQuantity-r19 introduces the AI/ML related report quantities : predicted CRI / SSB index and predicted RSRP (p-CRI, p-SSB-Index, p-CRI-RSRP, p-SSB-Index-RSRP), prediction accuracy indicators (rs-PAI, csi-PAI) and noneBM / noneCSI used for data collection without a regular report.
N4 Logic
N4 logic is the idea that the UE does not make an AI-based CSI prediction from only one instant of radio measurement. Instead, it can collect multiple CSI-RS observations over time and use them together as input to the AI model. In this sense, N4 represents the observation depth or the number of channel snapshots the UE buffers before running inference. When N4 is small, the model works with very limited history and the processing is simpler. When N4 is larger, the UE can observe channel evolution over multiple time points, which is especially useful for tracking Doppler behavior, fading trends, and other time-varying effects in high-mobility scenarios. At a high level, N4 defines how much past channel context the AI model uses in order to predict future channel behavior more accurately.
What is N4 for
- N4 means the number of CSI-RS observations or channel snapshots that the UE uses as input for one AI-based CSI prediction.
- It defines how much past channel history the UE buffers before running the AI inference.
- The main purpose of N4 is to give the AI model enough time-domain context to predict future channel behavior more accurately.
Why N4 is Needed
- If the UE uses only one observation, the model sees only the current channel condition.
- If the UE uses multiple observations, the model can track how the channel changes over time.
- This is especially important in high-mobility scenarios where Doppler and fading change quickly.
Small N4 vs Large N4
- If N4 is small, the processing is simpler and the memory requirement is lower.
- If N4 is large, the model has more historical context and can make a better time-series style prediction.
- A larger N4 usually increases processing load, memory usage, and inference complexity.
High-Level Meaning
- N4 is essentially the observation window size for AI-based CSI prediction.
- It tells the network that the UE can buffer and process multiple CSI-RS resources before generating a prediction.
- At a high level, N4 controls the tradeoff between prediction accuracy and implementation complexity.
Beam Management
While the previous section focused on predicting channel quality through CSI, this Beam Management section focuses on predicting the spatial direction of the signal. In Release 19, AI and ML are used to estimate which beam is likely to become the best beam in the near future, so the network can react before radio quality drops too much. This changes beam management from a reactive process into a more proactive one. Instead of waiting until beam quality degrades and then switching, the UE can use AI-based prediction to anticipate future beam behavior and help the gNB maintain more stable connectivity, especially in mobility scenarios where beam conditions can change very quickly.
MIMO-ParametersPerBand ::= SEQUENCE {
aiml-CSI-PredictionDoppler-r19 CodebookParametersCSI-PredictionDoppler-r19 OPTIONAL,
-- R1 58-0-1: CSI report framework for UE-side inference
aiml-CSI-Report-r19 SEQUENCE (SIZE (1..2)) OF CPU-PoolInfo-r19 OPTIONAL,
-- R1 58-1-2: UE-side beam prediction for BM Case1 for inference
aiml-BM-Case1-r19 SEQUENCE {
maxNumberOfInferenceReportPerBWP-r19 SEQUENCE {
periodicReporting-r19 INTEGER (1..4),
aperiodicReporting-r19 INTEGER (1..4),
semiPersistentReporting-r19 INTEGER (1..4)
},
maxNumberOfInferenceReportAcrossAllCC-r19
ENUMERATED { n1, n2, n3, n4, n8, n10, n12, n16 },
maxNumberOfResourceSetB-r19 ENUMERATED { n4, n8, n16 },
maxNumberOfResourceSetA-r19 ENUMERATED { n8, n16, n32, n64 },
resourceTypeSetB-CSI-RS-r19 SEQUENCE {
periodic-r19 ENUMERATED { supported } OPTIONAL,
aperiodic-r19 ENUMERATED { supported } OPTIONAL,
semiPersistent-r19 ENUMERATED { supported } OPTIONAL
},
inferenceReportType-r19 SEQUENCE {
periodic-r19 ENUMERATED { supported } OPTIONAL,
aperiodic-r19 ENUMERATED { supported } OPTIONAL,
semiPersistent-r19 ENUMERATED { supported } OPTIONAL
},
subUseCases-r19 ENUMERATED { subset, diffSet, both },
maxNumberOfPredictedBeamPerReportingInstance-r19
INTEGER (1..4),
numberOfOccupiedCPU-r19 INTEGER (0..8),
numberOfOccupiedCPUx-r19 INTEGER (0..8),
relaxationTimelineD-r19 SEQUENCE {
scs15kHz-r19 ENUMERATED { n7, n14, n21, n28, n35, n42, n56 },
scs30kHz-r19 ENUMERATED { n14, n28, n42, n56, n70, n84, n112 },
scs60kHz-r19 ENUMERATED { n28, n56, n84, n112, n140, n168, n224 },
scs120kHz-r19 ENUMERATED { n56, n112, n168, n224, n280, n336, n448 },
scs480kHz-r19 ENUMERATED { n224, n448, n672, n896, n1120, n1344, n1792 },
scs960kHz-r19 ENUMERATED { n448, n896, n1344, n1792 }
},
relaxationTimelineD1-r19 SEQUENCE {
scs15kHz-r19 ENUMERATED { n7, n14, n21, n28, n35, n42, n56 },
scs30kHz-r19 ENUMERATED { n14, n28, n42, n56, n70, n84, n112 },
scs60kHz-r19 ENUMERATED { n28, n56, n84, n112, n140, n168, n224 },
scs120kHz-r19 ENUMERATED { n56, n112, n168, n224, n280, n336, n448 },
scs480kHz-r19 ENUMERATED { n224, n448, n672, n896, n1120, n1344, n1792 },
scs960kHz-r19 ENUMERATED { n448, n896, n1344, n1792 }
},
occupiedResourcePool-r19 INTEGER (1..2)
} OPTIONAL,
-- R1 58-1-3: UE-side beam prediction for BM Case1 with predicted RSRP for inference
aiml-BM-Case1-PredictedRSRP-r19 INTEGER (1..4) OPTIONAL,
-- R1 58-1-4: UE-side beam prediction for BM Case2 for inference
aiml-BM-Case2-r19 SEQUENCE {
maxNumberOfInferenceReportPerBWP-r19 SEQUENCE {
periodicReporting-r19 INTEGER (1..4),
aperiodicReporting-r19 INTEGER (1..4),
semiPersistentReporting-r19 INTEGER (1..4)
},
maxNumberOfInferenceReportAcrossAllCC-r19
ENUMERATED { n1, n2, n3, n4, n8, n10, n12, n16 },
maxNumberOfResourceSetB-r19 ENUMERATED { n4, n8, n16, n32, n64 },
maxNumberOfResourceSetA-r19 ENUMERATED { n4, n8, n16, n32, n64 },
minNumberOfKBM-SetB-r19 ENUMERATED { n2, n4, n8 },
resourceTypeSetB-CSI-RS-r19 SEQUENCE {
periodic-r19 ENUMERATED { supported } OPTIONAL,
semiPersistent-r19 ENUMERATED { supported } OPTIONAL
},
inferenceReportType-r19 SEQUENCE {
periodic-r19 ENUMERATED { supported } OPTIONAL,
aperiodic-r19 ENUMERATED { supported } OPTIONAL,
semiPersistent-r19 ENUMERATED { supported } OPTIONAL
},
maxNumberOfPredictedBeamPerPerTimeInstance-r19
INTEGER (1..4),
maxNumberOfPredictedTimeInstance-r19 ENUMERATED { n1, n2, n4, n8 },
maxTotalNumberOfPredictedBeamPerReport-r19
ENUMERATED { n1, n2, n4, n6, n8, n12, n16, n32 },
dummy-TimeGap-r19 ENUMERATED { ms10, ms20, ms40, ms80, ms160 },
numberOfOccupiedCPU-r19 INTEGER (0..8),
numberOfOccupiedCPUx-r19 INTEGER (0..8),
relaxationTimelineD-r19 SEQUENCE {
scs15kHz-r19 ENUMERATED { n14, n28, n42, n56, n70, n84, n98 },
scs30kHz-r19 ENUMERATED { n28, n56, n84, n112, n140, n168, n196 },
scs60kHz-r19 ENUMERATED { n56, n112, n168, n224, n280, n336, n392 },
scs120kHz-r19 ENUMERATED { n112, n224, n336, n448 },
scs480kHz-r19 ENUMERATED { n448, n896, n1344, n1792 },
scs960kHz-r19 ENUMERATED { n896, n1792 }
},
relaxationTimelineD1-r19 SEQUENCE {
scs15kHz-r19 ENUMERATED { n14, n28, n42, n56, n70, n84, n98 },
scs30kHz-r19 ENUMERATED { n28, n56, n84, n112, n140, n168, n196 },
scs60kHz-r19 ENUMERATED { n56, n112, n168, n224, n280, n336, n392 },
scs120kHz-r19 ENUMERATED { n112, n224, n336, n448 },
scs480kHz-r19 ENUMERATED { n448, n896, n1344, n1792 },
scs960kHz-r19 ENUMERATED { n896, n1792 }
},
occupiedResourcePool-r19 INTEGER (1..2)
} OPTIONAL,
-- R1 58-1-5: UE-side beam prediction for BM-Case2 with predicted RSRP for inference
aiml-BM-Case2-PredictedRSRP-r19 SEQUENCE {
maxNumPredictedBeamPerInstance-r19 INTEGER (1..4),
maxNumPredictedTime-r19 ENUMERATED { n1, n2, n4, n8 },
maxTotalNumPredictedBeamInOneReport-r19
ENUMERATED { n1, n2, n3, n4, n6, n8, n12, n16, n24, n32 }
} OPTIONAL,
-- R1 58-1-6: Performance monitoring for UE-sided model
aiml-BM-Monitoring-r19 SEQUENCE {
maxNumTotalResource-r19 ENUMERATED { n4, n8, n16, n32, n64 },
maxNumReportPerBWP-Periodic-r19 INTEGER (1..4),
maxNumReportPerBWP-Aperiodic-r19 INTEGER (1..4),
maxNumReportPerBWP-SP-r19 INTEGER (1..4),
maxNumReportAcrossAllCC-r19 ENUMERATED { n1, n2, n4, n8 },
maxNumOccasion-r19 ENUMERATED { n1, n3, n7, n15 },
monitoringResourceType-r19 ENUMERATED { periodic, semipersistent },
monitoringReportType-r19 ENUMERATED { periodic, aperiodic, semipersistent }
} OPTIONAL,
-- R1 58-1-7: Data collection for UE-side beam prediction
aiml-BM-UE-DataCollection-r19 SEQUENCE {
subCase-r19 ENUMERATED { equal, subset, notSubset },
maxNumResourceSetB-r19 ENUMERATED { n4, n8, n16, n32, n64 },
maxNumResourceSetA-r19 ENUMERATED { n8, n16, n32, n64 }
} OPTIONAL,
-- R1 58-3-1: CSI prediction for UE-sided inference when N4=1
aiml-CSI-Prediction-r19 SEQUENCE {
supportedCSI-RS-ResourceList-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16),
scalingFactor-r19 ENUMERATED { n1, n2, n4 },
numberOfOccupiedCPU-r19 INTEGER (0..8),
numberOfOccupiedCPUx-r19 INTEGER (0..8),
relaxationTimelineT-r19 SEQUENCE {
scs15kHz-r19 ENUMERATED { n14, n28, n56, n112 },
scs30kHz-r19 ENUMERATED { n28, n56, n112, n224 },
scs60kHz-r19 ENUMERATED { n56, n112, n224, n448 },
scs120kHz-r19 ENUMERATED { n112, n224, n448 },
scs480kHz-r19 ENUMERATED { n448, n896, n1792 },
scs960kHz-r19 ENUMERATED { n896, n1792 }
},
occupiedResourcePool-r19 INTEGER (1..2),
inferenceReportType-r19 SEQUENCE {
aperiodic-r19 ENUMERATED { supported },
semiPersistent-r19 ENUMERATED { supported }
}
} OPTIONAL,
-- R1 58-3-1a-1: DD unit size when A-CSI-RS is configured for CMR N4>1 for UE side inference of CSI prediction
aiml-CSI-PredictionUnitDurationDD-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-2: CSI prediction for UE-sided inference when N4>1
aiml-CSI-PredictionN4-r19 SEQUENCE {
supportedCSI-RS-ReportSettingAcrossAllCC-r19
SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF
SupportedCSI-RS-ReportSetting-r18,
supportedCSI-RS-ReportSettingAcrossOneReport-r19
SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF
SupportedCSI-RS-ReportSetting-r18,
numOccupiedCPU-r19 INTEGER (0..8),
numOccupiedCPUx-r19 INTEGER (0..8),
occupiedPool-r19 INTEGER (1..2)
} OPTIONAL,
-- R1 58-3-4: UE side data collection for CSI prediction
aiml-CSI-PredictionUE-DataCollection-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-5: Performance monitoring for CSI prediction model
aiml-CSI-PredictionMonitoring-r19 SEQUENCE {
supportedCSI-RS-ResourceList-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16),
numOccupiedCPU-r19 INTEGER (1..2)
} OPTIONAL,
...,
[[
-- Time gap for BM Case2: bitmap of supported time gaps (ms10, ms20, ms40, ms80, ms160)
aiml-BM-Case2-v1920 SEQUENCE {
timeGap-r19 BIT STRING (SIZE (5))
} OPTIONAL
]]
}
CodebookParametersCSI-PredictionDoppler-r19 ::= SEQUENCE {
-- R1 58-3-1b: Maximum number of aperiodic CSI-RS resources that can be configured in the same CSI report setting for Rel-16-based
-- doppler measurement for UE side inference of CSI prediction
maxNumberOfAperiodic-CSI-RS-Resource-r19 ENUMERATED { n4, n8, n12 } OPTIONAL,
-- R1 58-3-1-2: Support R=2 for Rel-16-based doppler codebook for UE side inference of CSI prediction
eType2DopplerR2-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL,
-- R1 58-3-1-3: Support X=1 based on first and last slot of WCSI, for Rel-16-based doppler codebook for UE side inference of CSI prediction
eType2DopplerX1-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-1-3a: Support X=2 CQI based on 2 slots for Rel-16-based doppler codebook for UE side inference of CSI prediction
eType2DopplerX2-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-1-4: Support of l = (n – nCSI,ref) for CSI reference slot for Rel-16 based doppler codebook for UE side inference of CSI prediction
eType2DopplerL-N4D1-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-1-5: Support of L=6 for Rel-16 based doppler codebook for UE side inference of CSI prediction
eType2DopplerL6-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-1-6: Support of rank equals 3 and 4 for Rel-16 based doppler codebook for UE side inference of CSI prediction
eType2DopplerR3R4-r19 ENUMERATED { supported } OPTIONAL,
-- R1 58-3-1-7: Active CSI-RS resources and ports for mixed R16 based doppler codebook for CSI prediction via UE side model with
-- other codebooks in any slot
codebookComboParameterMixedTypePrediction-r19 SEQUENCE {
type1SP-Type1SP-N4-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL,
type1SP-eType2SP-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL,
type1SP-eType2SP-N4-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL,
type1SP-N4-eType2SP-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL,
type1SP-N4-eType2SP-N4-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL,
eType2SP-eType2SP-N4-r19 SEQUENCE (SIZE (1..maxNrofCSI-RS-ResourcesExt-r16)) OF INTEGER
(0..maxNrofCSI-RS-ResourcesAlt-1-r16) OPTIONAL
} OPTIONAL
}
CPU-PoolInfo-r19 ::= SEQUENCE {
maxNumCPU-PerCC-r19 INTEGER (1..8),
maxNumCPU-AllCC-r19 INTEGER (1..32)
}
In addition to MIMO-ParametersPerBand, Release 19 adds per-band RF capabilities in BandNR indicating support of beam prediction under the known Rx beam condition :
BandNR ::= SEQUENCE {
...
-- R4 59-1: UE-side beam prediction for BM Case1
aiml-BM-Case1-KnownRxBeam-r19 ENUMERATED { supported } OPTIONAL,
-- R4 59-2: UE-side beam prediction for BM Case2
aiml-BM-Case2-KnownRxBeam-r19 ENUMERATED { supported } OPTIONAL,
...
}
Followings are short descriptions on important parameters
maxNumPredictedBeam-r19 : Indicates how many future candidate beams the UE can predict and report in a single AI-based Beam Management report.maxNumberOfPredictedTimeInstance-r19 : Indicates how many future time points the UE can predict, especially for Case 2 temporal beam prediction.dummy-TimeGap-r19 / aiml-BM-Case2-v1920 : Defines the spacing between future prediction points. In the released ASN.1 the original enumerated timeGap field is frozen as dummy-TimeGap-r19 (not used by the network), and the time gap capability is instead reported in aiml-BM-Case2-v1920 as a 5-bit bitmap in which each bit indicates support of ms10, ms20, ms40, ms80 or ms160. This tells the network how far apart the predicted beam states are in time.maxNumberOfResourceSetA-r19 : Indicates how many measured beam resource sets the UE can use as observation input to the AI model.maxNumberOfResourceSetB-r19 : Indicates how many target beam resource sets the UE can predict as output of the AI model.aiml-BM-Case1-PredictedRSRP-r19 : Indicates support for reporting predicted RSRP rather than only measured RSRP for beam management.maxNumOccasion-r19 : Defines how many monitoring occasions can be used when the network evaluates prediction accuracy over time.relaxationTimelineD-r19 : Indicates the extra processing time the UE may need to complete beam prediction inference and prepare the report.relaxationTimelineD1-r19 : Indicates additional processing time needed for more complex operations such as concurrent data collection or advanced multi-beam handling.occupiedResourcePool-r19 : Identifies which internal UE processing or hardware resource pool is assigned to the beam management task.subset / diffSet related parameters : Indicate the relationship between the measured beam set and the predicted beam set, such as whether prediction is made within the same set or across different sets.aiml-BM-Case1-r19 : Represents support for AI-based beam prediction for the current or immediate next beam decision.aiml-BM-Case2-r19 : Represents support for more advanced temporal prediction where the UE predicts a sequence of future best beams over multiple time instances.aiml-BM-Case1-KnownRxBeam-r19 / aiml-BM-Case2-KnownRxBeam-r19 : Per-band RF capabilities signalled in BandNR, indicating that the UE supports beam prediction for Case 1 / Case 2 under the known Rx beam condition, i.e. the RAN4 measurement requirements assuming the UE already knows the suitable Rx beam.
AI-BM Case 1 vs. Case 2
Case 1 and Case 2 are the two main AI-based Beam Management approaches in Release 19, but they emphasize different dimensions of prediction. Case 1 is more of a spatial prediction problem. It focuses on identifying the best beam among candidate beams for the current moment or the immediate next moment based on the current observation. In other words, it is mainly about selecting the best direction in space. Case 2 extends this idea into the temporal domain. It predicts how the best beam will change over multiple future time instances, so the network can anticipate beam evolution ahead of time. In simple terms, Case 1 is mainly about spatial beam selection, while Case 2 is about time-evolving beam prediction.
Case 1 refers to intra-unit prediction(Spatial prediction). The UE predicts the best beam for the current time instance or the immediate next time instance based on current measurements.Case 2 refers to temporal prediction. The UE predicts a sequence of future best beams across multiple future time instances.maxNumberOfPredictedTimeInstance-r19 indicates how many future time points the UE can predict.timeGap-r19 (signalled via aiml-BM-Case2-v1920) defines the spacing between those future prediction points.- Case 1 is mainly about immediate refinement and accuracy, while Case 2 is about proactive beam tracking over time.
Resource Set A and Resource Set B
Resource Set A and Resource Set B define the basic input-output structure of AI-based Beam Management. One set represents what the UE actually measures from the radio environment, and the other set represents what the UE tries to predict with AI. In high-level terms, this allows the UE to observe a limited number of real reference signals and then use those observations to estimate the quality of a larger or different set of candidate beams. This structure is important because the UE cannot always measure every possible beam directly in real time. Instead, it measures a practical observation set and uses AI to infer the expected quality or best beam choice for the prediction set.
maxNumberOfResourceSetA-r19 indicates the number of measured resource sets that can be used as AI model input.maxNumberOfResourceSetB-r19 indicates the number of target resource sets whose beam quality the AI model can predict.Resource Set A is the observation set. These are the SSB or CSI-RS resources that the UE actually measures.Resource Set B is the prediction set. These are the beams or resources whose future quality the UE tries to predict.subset means the predicted beams in Set B are a smaller group within the measured beams in Set A.diffSet means the predicted beams are different from the measured beams, such as using measurements on one frequency to predict beams on another frequency.
In summary :
|
Feature |
Set B (Input) |
Set A (Output) |
|---|---|---|
|
Role |
Observation Set: What the UE actually measures in the physical world. |
Prediction Set: The full list of potential beams the network wants to evaluate. |
|
Quantity |
Small (e.g., 8 beams) |
Large (e.g., 64 beams) |
|
UE Task |
Performs L1-RSRP measurements on these resources. |
Runs AI inference to predict the RSRP or "Best Beam ID" for this set. |
Predicted RSRP
Predicted RSRP is one of the key ideas that makes AI-based Beam Management different from legacy beam reporting. In the legacy approach, the UE can only report the signal power that it has already measured. In the AI-based approach, the UE can go one step further and report the signal power it expects to see in the near future for a candidate beam. Through aiml-BM-Case1-PredictedRSRP-r19, the UE provides the gNB with a forward-looking estimate rather than only a backward-looking measurement. This allows the network to make proactive beam decisions before signal quality actually degrades, which can improve beam stability and reduce the risk of sudden beam failure.
- In legacy Beam Management, the UE reports the RSRP that it has already measured.
- In AI-based Beam Management, the UE can report
Predicted RSRP throughaiml-BM-Case1-PredictedRSRP-r19 . - This means the UE reports the expected beam power before actually reaching that beam condition.
- This allows the gNB to make proactive beam decisions instead of waiting for signal quality to drop.
Relaxation Timeline D and D1
Relaxation Timeline D and D1 define the extra processing time that the UE may need when performing AI-based Beam Management. Unlike conventional beam reporting, AI-based prediction requires additional computation for inference, report generation, and sometimes simultaneous data collection or more complex multi-beam processing. relaxationTimelineD represents the extra time needed for the basic inference and reporting procedure, while relaxationTimelineD1 represents additional time that may be required for more advanced or parallel operations. At a high level, these parameters allow the UE to inform the network that AI processing cannot always fit into the same timing budget as legacy beam management, especially when NPU or GPU based computation is involved.
relaxationTimelineD indicates the extra time the UE needs to run AI inference and prepare the Beam Management report.relaxationTimelineD1 indicates additional time that may be needed for more complex operations, such as concurrent data collection or multi-beam processing.- These parameters are important because AI-based Beam Management requires more processing time than conventional beam reporting.
- They allow the UE to signal that extra time is needed for NPU or GPU based processing.
Model Specification
One of the most distinctive aspects of the Release 19 AI/ML framework is what 3GPP decided NOT to specify : the AI model itself. There is no IE in the RRC ASN.1 that describes, identifies, downloads or configures a specific neural network. The model remains a vendor proprietary black box inside the UE, and the specification only defines the interfaces around it : what the model takes as input (e.g. the CSI-RS observations of Set B), what it produces as output (e.g. predicted CSI or predicted beams of Set A), how much processing budget it consumes (CPU pools, relaxation timelines) and how its behavior is managed over time (applicability, monitoring, fallback). In fact, the word "model" appears in the AI/ML related ASN.1 of TS 38.331 only in comments and in a single field description.
The table below summarizes what goes into and what comes out of the model black box from the RRC point of view, i.e. which RRC IEs configure the model input and which IEs carry the model output :
|
Functionality |
Input to the Model (Black Box) |
Output from the Model (Black Box) |
|---|---|---|
|
BM Case 1 (spatial beam prediction) |
L1-RSRP measurements on the Set B resources (SSB or CSI-RS), taken from the CSI-ResourceConfig referenced by the CSI-ReportConfig that contains predictionConfiguration-r19 (configurationForBM-PredictionAndDataCollection-r19). associatedIdForChannelMeasurement-r19 identifies the deployment the measurements belong to. |
Predicted best beam(s) among Set A, optionally with predicted RSRP, reported as reportQuantity-r19 = p-CRI, p-SSB-Index, p-CRI-RSRP or p-SSB-Index-RSRP. The number of reported predicted beams is set by nrofReportedPredictedRS-r19, and associatedIdForChannelPrediction-r19 identifies the predicted beam set. |
|
BM Case 2 (temporal beam prediction) |
The same Set B L1-RSRP measurements, but observed over multiple past time instances (observation history buffered inside the UE). |
Predicted best beams / predicted RSRP for up to nrofTimeInstance-r19 future time instances spaced by timeGap-r19, reported with the same p-CRI / p-SSB-Index (-RSRP) report quantities. |
|
CSI Prediction |
N4 CSI-RS channel snapshots (CMR) from the resources referenced by the CSI-ReportConfig whose predictionConfiguration-r19 is set to csi-InferencePrediction-r19. The observation depth corresponds to the N4 / scalingFactor-r19 capability of the UE. |
Predicted CSI (RI / PMI / CQI) for a future reference slot, encoded with the Rel-16 eType-II Doppler based codebook configured by codebookConfig-r19 and carried in a legacy CSI report. |
|
Performance Monitoring |
Actual measurements of the previously predicted beams / channel, collected over nrofTransmissionOccasion-r19 occasions (BM) or at the configured time instance (CSI), linked to the prediction report via refToPredictionConfig-r19. |
Prediction accuracy indicators reported as reportQuantity-r19 = rs-PAI (beam prediction, for the instance set by timeInstanceForRS-PAI-r19) or csi-PAI (CSI prediction, timeInstanceForCSI-PAI-r19). |
|
UE side Data Collection (training) |
Set B (and Set A) measurements collected with reportQuantity-r19 = noneBM / noneCSI, or via logged data collection. |
Nothing over the air - the data is stored inside the UE as training data, and the model training itself takes place outside 3GPP signalling. |
Why the Model itself is not Standardized
Vendor innovation : chipset and device vendors compete on model architecture, model size and training quality. Standardizing the model itself would freeze the technology at the level of the first release.Testability : conformance testing can only be defined on observable input/output behavior (prediction accuracy, reporting timeline), not on internal weights of a neural network.IP protection : trained model weights are valuable trade secrets. Exchanging them over standardized interfaces raised strong inter-vendor concerns during the study phase.Upgradability : a UE vendor can retrain and replace its model at any time without any signalling impact, as long as the declared capability and the applicability status still hold.
Functionality based LCM vs Model-ID based LCM
The Release 18 study (TR 38.843) defined two candidate approaches for the Life Cycle Management (LCM) of AI/ML models, and the choice between them decides whether a model ever becomes visible on the air interface.
Functionality based LCM : the network sees and controls only functionalities (e.g. beam prediction Case 1, CSI prediction). Activation, deactivation, switching and fallback are applied to a configured functionality, and which physical model the UE runs underneath is invisible to the network. This is what Release 19 standardized for the UE-side (one-sided) models.Model-ID based LCM : each model (or model pairing) is identified by a model ID known to both the network and the UE, and LCM actions are applied per model ID. This approach was studied but NOT standardized in Release 19.Model identification Type A : the model is identified offline, through co-engineering between the UE vendor, the network vendor and the operator, without any over-the-air procedure. Only the resulting identifiers would ever appear in signalling.Model identification Type B : the model is identified via over-the-air signalling, either UE initiated (Type B1) or network initiated (Type B2). Even in the UE initiated case the network assigns or confirms the ID in order to guarantee uniqueness within its domain. Type B was deferred and produced no Release 19 signalling.
How the Network manages Models without a Model ID
Even though no model is identified explicitly, Release 19 gives the network a complete set of indirect handles, and together they emulate most of what a model ID would provide :
Configuration level control : "activating a model" is realized by configuring the corresponding report configuration (predictionConfiguration-r19 inside CSI-ReportConfig), and "deactivating" it by releasing the configuration. The functionality, not the model, is the unit of control.Applicability reporting : through ApplicabilityReportConfig-r19 / ApplicabilityReport-r19 the UE declares each configuration applicable or inapplicable, optionally with releaseConfigurationPreference-r19. The field description explicitly mentions model unavailability as a trigger - this is how a UE says "I do not have a working model for this" without ever naming a model.AssociatedId-r19 : a 24 bit, PLMN-unique identifier indicating that the UE may assume similar properties of a DL Tx beam or beam set associated with the same value. It acts as a deployment fingerprint, letting the UE recognize whether the current site matches what its model was trained on. It identifies the scenario, not the model.Performance monitoring : the prediction accuracy indicators (rs-PAI for beam prediction, csi-PAI for CSI prediction) let the network detect a degraded model and fall back to legacy operation by releasing the configuration.Data collection : logged and event based data collection provide training data, but the training itself and the delivery of the resulting model to the UE happen entirely outside 3GPP signalling.
Who would configure a Model ID
If and when a model ID (or model pairing ID) is introduced, the roles are expected to follow the pattern that is already visible in Release 19. In every variant that 3GPP has seriously discussed, the network side owns the ID space and the UE never invents its own model ID.
ID assignment : the ID value is created offline (model identification Type A), during co-engineering between the UE vendor, the network vendor and the operator. The expected scope is operator administered, PLMN-unique IDs - exactly the convention already frozen for AssociatedId-r19, which is defined as "unique within a PLMN".ID configuration : the gNB configures the applicable ID to the UE via dedicated RRC signalling, after the UE has declared the supported IDs or pairings in its UE capability - the same flow used today for associatedIdForChannelPrediction-r19 inside predictionConfiguration-r19.Runtime selection : fast selection among RRC configured candidates (activation, switching, fallback) is expected via MAC CE, following the usual NR pattern such as TCI state activation or semi-persistent CSI reporting.
Where Model related Parameters would live
|
Aspect |
Layer / Mechanism |
|---|---|
|
Capability (supported functionalities, future model pairings) |
RRC UE capability (AIML-Parameters-r19, aiml-* fields in MIMO-ParametersPerBand / CA-ParametersNR / BandNR) |
|
Functionality configuration |
RRC dedicated signalling (predictionConfiguration-r19 in CSI-ReportConfig, ApplicabilityReportConfig-r19, DataCollectionPreferenceConfig-r19) |
|
Activation / switching / fallback |
RRC configuration and release in Release 19; MAC CE expected for fast LCM actions in later releases |
|
Model identification |
Offline co-engineering (Type A); over-the-air identification (Type B) deferred |
|
Model delivery (weights / binary) |
User plane / application layer and OAM, supported by the SA2 AI/ML cross-domain work - deliberately out of RRC scope |
|
Positioning use cases |
LPP (TS 37.355) rather than RRC carries the AI/ML positioning configuration |
Outlook
The first genuinely model related parameter on the air interface is expected to come from the two-sided model work on CSI compression, where the UE side encoder and the network side decoder must be trained as a pair : such an operation structurally requires some pairing identity in capability signalling and configuration, even if it identifies a pairing rather than a downloadable model. AI/ML for mobility, the other major normative candidate, is expected to reuse the functionality based framework without model IDs. Note that this section reflects the status as of the Release 19 ASN.1 freeze and the early Release 20 scoping discussions, and the details may still evolve.
Reference
- 3GPP TS 38.331 v19.2.0 : NR; Radio Resource Control (RRC); Protocol specification
- 3GPP TS 38.306 : NR; User Equipment (UE) radio access capabilities
- 3GPP TS 38.214 : NR; Physical layer procedures for data
- 3GPP TS 37.355 : LTE Positioning Protocol (LPP)
- 3GPP TR 38.843 : Study on Artificial Intelligence (AI)/Machine Learning (ML) for NR air interface
- 3GPP TR 38.744 : Study on Artificial Intelligence (AI)/Machine Learning (ML) for mobility in NR
- RP-234039 : New WID on Artificial Intelligence (AI)/Machine Learning (ML) for NR air interface (Release 19 Work Item Description)