3GPP TSG RAN WG1 Meeting #120bis	R1-2501718
Wuhan, China, April 7th – 11th, 2025
Agenda Item:	9.1.4.1
Source:	Futurewei
Title:	Discussion of CSI compression on AI/ML for NR air interface
Document for:	Discussion

Conclusions
In this contribution, we discussed our evaluation results for scalable model structure and complexity-performance tradeoff by using different model structures. In addition, we shared our views regarding remaining issues and potential specification impact. Our observations and proposals are as follows.
For evaluation of scalable model structure:
Observation 1: In model structure scalability study for temporal domain Case 0 of CSI compression using two-sided model, when handling scalability over payload configurations using Alt2 (truncation/masking of the output linear layer output) and/or Alt3 (by varying quantization parameters), performance evaluation shows there is small SGCS performance loss compared to payload size-specific models.
Alt2 has 1.5% - 3.5% SGCS performance degradation compared to the payload size-specific models.
Observation 2: In model structure scalability study for temporal domain Case 0 of CSI compression using two-sided model, using scalable model structure can potentially save significant overhead in model training, the associated LCM burden and the storage concern on both UE-side and NW-side while achieving comparable SGCS performance compared to the benchmark (payload size-specific models), at least for Alt2 and Alt3.
For evaluation of tradeoff between complexity and performance:
Observation 3: For CSI compression using two-sided model, when comparing SGCS performance across models with different space and computational complexities, minor to medium performance degradation is observed on the less complex models (from 0% to ~5.5% from our study), compared to the benchmark model with much higher complexity.
Observation 4: For CSI compression using two-sided model, it is feasible for less complex models to achieve reasonable performance that is comparable with a more complex benchmark model.
Proposal 1: For CSI compression using 2-sided model, regarding the trade-off between performance and complexity/overhead, consider adopting model structure with less complexity to reduce both space and computational overhead.
For other topics in the study:
Observation 5: If Option 3b in Direction B is supported, it is preferrable for UE-side to start using the encoder or encoder parameters as soon as possible, especially for model (parameter) update case.
Observation 6: For Option 3a-1 and Option 4-1 in Direction A, the latency requirement for model parameter/dataset exchange/transfer can be more relaxed compared to Option 3b in Direction B if supported.
Proposal 2: If inter-vendor training collaboration option 3b in Direction B is supported, consider adopting over-the-air delivering method(s).
Proposal 3: If inter-vendor training collaboration option 3a-1 and/or 4-1 in Direction A are/is supported, consider adopting standardized signalling in upper layers or other offline method(s) as delivery options.
Proposal 4: For AI/ML-based CSI compression using two-sided model, further study the following potential specification impact related to quantization of CSI feedback at least for Option 3a-1/4-1 in Direction A and Option 3b in Direction B in alleviating/resolving the issues related to inter-vendor training collaboration:
Vector quantization:
Exchange of vector quantization codebook(s).
Segmentation information (if segmentation is used) of the CSI output.
Scalar quantization:
Configuration of quantization granularity and the corresponding range values.
Exchange of scalar quantization dictionary.
Proposal 5: In AI/ML-based CSI compression using two-sided model, for CQI determination, if the actual or reference CSI reconstruction model is available at UE, adopt Option 2a to determine CQI at UE. 
For CSI processing unit:
Observation 7: AI/ML-based CSI reporting needs both legacy CSI processing resources and AI/ML-specific processing resources.
Observation 8: For AI/ML-based CSI reporting, the needed legacy CSI processing part can share CPU with legacy CSI reporting.
Observation 9: The AI/ML engines / hardware on the UE are likely shared among different AI/ML-based features on the UE.
Observation 10: Separating the CPU counting for legacy and AI/ML-based CSI reporting provides a more accurate, flexible, and manageable approach to handling the computational demands of this new technology. It aligns with the potential for dedicated AI/ML processing hardware, facilitates better resource management and prioritization, enables more informative UE capability reporting, and supports efficient sharing of AI/ML resources within the UE.
Observation 11:  Supporting the sharing of AI/ML-based computing resources among different AI/ML features and functionalities is a more efficient, flexible, and scalable approach. It optimizes resource utilization, potentially reduces power consumption, simplifies UE capability reporting, and aligns with industry trends in AI hardware design. Counting the resources of this shared pool separately from legacy CPU usage provides a clearer understanding of the processing demands of AI/ML-based operations, facilitating better network management and configuration.
Proposal 6: Support to separate CPU Counting for Legacy and AI/ML-based CSI Reporting, i.e., legacy CPU and AI/ML-based CPU are from different resource pools.
Proposal 7: Support the sharing of AI/ML-based computing resources among different AI/ML features and functionalities.
Proposal 8: Adopt a new/updated timeline for model inference for the AI/ML-based counting approach of CPU and model inference processing resource. 

3GPP TSG RAN WG1 #120bis 		R1- 2501798
Wuhan, China, April 7th – 11th, 2025

Source:	vivo
Title:	Discussion on CSI compression and other aspects on AI/ML model/data
Agenda Item:	9.1.4.1
Document for:	Discussion and Decision

Conclusion
In this contribution, we have the following observations:
Given the same model dimension, trade-off between number of attention head and dimension of attention head does not affect performance.
Given the same complexity, very deep (e.g., 64 blocks) or very shallow (e.g., 1 block) models may show some performance degradation. So, it is suitable to consider a moderate value such as 4 blocks in the specified model structure.
For comparison between Structure 1 (zero padding based) and case 2 dedicated structure, Case 1 with different parameter sets and fixed hyperparameters across different (Tx, SB, PC) has no performance loss (up to -0.1%), compared to Case 2.
For comparison between Structure 1 (zero padding based) and case 2 dedicated structure, Case 1 with single parameter set (A: same transformer backbone parameters, different embedding layer parameter / output linear layer parameters for different branches) and fixed hyperparameters across different (Tx, SB, PC) has minor performance loss (average -0.8%, up to -2.3%), compared to Case 2.
For comparison between Structure 1 (zero padding based) and dedicated structure, Case 1 with single parameter set (B: same transformer backbone parameters, same embedding layer parameter) across different (Tx, SB, PC) has minor performance loss (average -1.0%, up to -2.3%), compared to Case 2.
For comparison between Structure 2 (atomic units based) and case 2 dedicated structure, Case 1 with different parameter sets and fixed hyperparameters across different (Tx, SB, PC) has no performance loss (average +0.4% gain), compared to Case 2.
For comparison between Structure 2 (atomic units based) and case 2 dedicated structure, Case 1 with single parameter set (A: same transformer backbone parameters, different embedding layer parameters / output linear layer parameters for different branches) and fixed hyperparameters across different (Tx, SB, PC) has minor performance loss (average -0.3%, up to -1.2%), compared to Case 2.
For comparison between Structure 2 (atomic units based) and case 2dedicated structure, Case 1 with single parameter set (B: same transformer backbone parameters, same embedding layer parameter) across different (Tx, SB, PC) and fixed hyperparameters has minor performance loss (average -0.7%, up to -2.1%), compared to Case 2.
For comparison between Structure 3 (token extraction based) and case 2 dedicated structure, Case 1 with different parameter sets and fixed hyperparameters has minor performance loss (average -0.9%, up to -2.6%), compared to Case 2.
For comparison between Structure 3 (token extraction based) and case 2 dedicated structure, Case 1 with single parameter set (A: same transformer backbone parameters, different embedding layer parameters / output linear layer parameters for different branches of fixed hyperparameters) and fixed hyperparameters has small performance loss (average -2.0%, up to -2.7%), compared to Case 2.
For comparison between Structure 3 (token extraction based) and case 2 dedicated structure, Case 1 with single parameter set (B: same transformer backbone parameters, same embedding layer parameters and same output linear layer parameters) and fixed hyperparameters across different (Tx, SB, PC) has medium performance loss (average -3.2%, up to -5.4%), compared to Case 2.
Scalable structure across different (Tx, SB, PC) settings is feasible.
Scalable structure and parameter across different (Tx, SB, PC) settings are feasible.
From initial results, it is seen that CNN based angular-delay domain input has small SGCS loss compared to spatial-frequency domain input. However, CNN in angular-delay domain input is expected to be simple.
To solve the scalability issue, similar methods for spatial-frequency domain input could be used in angular-delay domain input
Tx antenna ports: no scalability issue;
Number of subbands and higher-layer parameter (numberOfPMI-SubbandsPerCQI-Subband): input padding;
Payload sizes: adaption layers.
The proposed structure for case3 performs better than the R18 DD codebook compression under both ideal and realistic AR CSI prediction in terms of both inter-mediate KPI and throughput.
Model structures mentioned in the agreement require non-trivial additional efforts for specification on top of case0 Transformer backbone structure: 
Study and definition of new layers (e.g., Conv-LSTM/LSTM layers) in option 1
Study and definition of new operations (e.g., new quantization method on top of conventional SQ/VQ) in option 2
The proposed case2 model can reduce the model complexity by about one order of magnitude both in Flops and parameter scale while achieves even higher performance gain than case0.
It is not necessary for NW to further exchange target CSIs in option 3a-1 to achieve a good performance. 
For precoded RS-based monitoring, using KPIDiff =f ( KPIActual , KPIGenie ) for monitoring decision is unrealistic as KPIGenie is not available.
Two-sided model complexity could be reduced without performance loss by cell-specific model.
Field data simulation results reveal that the parameter scale/FLOPs of cell-specific model can be reduced to ~1/10 while achieving similar performance gain.
Two-sided model complexity could be reduced without performance loss by TSF model.
Our simulation results reveal that the parameter scale/FLOPs of TSF model can be reduced to ~1/10 while achieving even higher performance gain.

And the following proposals:
For standardized model structure, the following hyper-parameter needs to be specified: 
Input dimension 
Number of tokens
Feature dimension of each token
Output dimension/Dimension of transformer block:  per token
Number of transformer blocks 
Number of self-attention heads 
Dimension of attention head 
Dimension of latent space inside feedforward module 
Activation function
Output dimension 
Quantization parameter
Scalar quantization: Number of bits per latent dimension: N bit
Vector quantization: segment size, bits per segment 
The candidate values are suggested for the hyper-parameters for standardized model structure
Hyper-parameters of Transformer are suggested as: 
Output dimension/Dimension of transformer block:  =256
Number of transformer blocks =4
Number of self-attention heads =8
Dimension of attention head =32
Dimension of latent space inside feedforward module =
Towards a better trade-off between performance and complexity, the following combination can be candidate hyper-parameter for standardized model structure: 
“4 Transformer blocks, model dimension 256, dimension of each head 32, 8 attention heads” for a relatively better performance, 
“4 Transformer blocks, model dimension 128, dimension of each head 16, 8 attention heads” for a relatively lower complexity while keeping most gain of the case.
For scalable structure and scalable structure and parameter, Structure 1 and Structure 2 are suggested. 
Structure 1: structure with zero padding based operation (i.e., Alt 1 for choice of token and feature, Alt 2 for feature dimension, Alt 2 for token dimension, Alt 2 for payload configuration)
Structure 2: structure with atomic unit based operation (i.e., Alt 3 for choice of token and feature, no padding or other techniques needed for scalability over feature dimension, Alt 2 for token dimension, Alt 2 for payload configuration)
Towards model structure standardization for temporal domain case3, reuse the model structure of Case 0 with input/output adaptation, i.e., directly extending the range of self-attention operation to multiple slots.
Seek a model structure for temporal domain case2 which only requires straightforward extension of Transformer for case0 to reduce the specification efforts and facilitate the SI progress.
Towards model structure standardization for temporal domain case2, consider to use Transformer backbone (same backbone as the model structure for Case 0) with the following adaptation: 
Extending the range of self-attention operation to multiple slots in auto regressive way.
Hyper parameter adjustment
For UE side monitoring, ID of an inference report configuration is configured in the configuration for monitoring to link the inference report configuration and monitoring report configuration and indicate the pairing relationship.
The pairing ID should be assigned by the NW.
Performance metric may include:
SGCS{target CSI, reconstructed CSI} for end to end performance target, 
NMSE{ PMI, PMI’}for encoder only performance target 
UE-side target CSI is input data for evaluating the performance.
For data collection in CSI compression
For NW to collect data for training
For Data format of NW to collect data for training, support using codebook-based Rel-16 eType2 as starting point.
For the FFS part
Report additional information regarding the samples, e.g., CQI, RI, timestamp
For CSI-RS and SRS configuration, reuse existing CSI-RS and SRS configuration as starting point
For case 2 and case 3 configuration, inform UE to collect multiple continuous data as a sample and number/time occasions of inputs and outputs.
For UE to collect data for training,
Different to NW side data collection, UE request RS configuration/transmission for data collection can be considered since the NW side didn’t know the UE data requirement (e.g., periodicity of CSI-RS, subbands of CSI-RS). 
For CSI-RS and SRS configuration, similar to NW side data collection, reuse existing CSI-RS and SRS configuration as starting point.
For configuration of temporal aspects for temporal case 2/3, P/SP CSI-RS can be supported
UE side monitoring can be supported for CSI compression with two-sided models for proxy model based solution.
For precoded RS-based monitoring  
Utilize Prob(KPIActual < thr) as a KPI for UE-side monitoring, when Prob(KPIActual < thr) exceeds thresthold can trigger a report, or trigger NW-side monitoring/monitoring decision;
Utilize KPIActual < thr as an event for UE-side monitoring, and the event report can trigger NW-side monitoring and monitoring decision
For data collection of temporal domain aspects Case 3, option 2 (The target CSI for training is derived based on the measured CSI of the future slot(s)) is supported
For monitoring of temporal domain aspects Case 3, Option 1(the monitoring label is derived based on the predicted CSI of the future slot(s))
Separate processing unit pools for AI based CSI and legacy CSI for quantifying AI-based CSI processing
The Processing Unit should be shared by CSI related report (i.e., CSI compression, CSI prediction, beam prediction) at least.
For AIML functionality switching/activation, consider the following steps for defining timeline:
Step 1: loading model to AI engine (e.g., APU/CPU/GPU)
Example 1: Loading model from flash storage to AI engine 
Example 2: Loading model from Global memory to AI engine 
Example 3: model has been loaded in AI engine memory 
Step 2: Initialization and scheduling of AI resources in AI engine
Step 3: running the model based on the input data
Step 4: return output results
Step 5: release model/AI resource in AI engine
For AIML functionality switching/activation, consider the following value range for defining timeline in step1/3:

3GPP TSG RAN WG1 #120bis		R1- 2501861
Wuhan, China, April 7th – 11th, 2025
Agenda Item:     9.1.4.1
Source:	Spreadtrum, UNISOC
Title:	             Discussion on AIML for CSI compression
Document for:	Discussion and decision

Conclusion 
In this contribution, we provide our opinions on standard impacts of CSI compression.
Observation 1: Further consideration is needed for how the metric should be used in Option 3a-1 Alt 2.
Observation 2: Whether to use common or multiple encoders depends on UE capabilities. 
If multiple UE capabilities are similar, a common encoder can be used.
 If multiple UE capabilities are different, multiple encoders should be used.
Observation 3: Performance degradation due to data mismatch for Direction B can not be addressed.
Observation 4: For option 3b-1, there is no risk of privacy disclosure without considering additional information.

Proposal 1: For additional information of Direction A (Options 3a-1 and 4-1), the performance target including performance metric:
At least SGCS should be considered as the performance metric.
Whether the performance metric are applied to the pair of encoder and decoder or only to encoder will depend on how the UE trained its encoder part.
Proposal 2: For the overhead concern of Direction A, it can be considered to alleviate it by transferring a small amount of data each time.
Proposal 3: If the NW-side data does not contain the UE-side data, performance may deteriorate. UE can report UE-side data to alleviate performance degradation caused by data mismatch
Proposal 4: Direction B does have overhead concern and can't be addressed.
Proposal 5: For direction C, 3GPP’s statistical channel model should be used for reference model(s) training.
Proposal 6: For network side data collection, support enhanced Rel-16 eTypeII codebook design to achieve high-resolution CSI for model training.
Proposal 7: For NW to collect data for training, both L1 signaling based reporting and RRC signaling based reporting should be supported for ground-truth reporting.
Proposal 8: For NW to collect data for training, enhancement on the design of CSI-RS is not needed.
Proposal 9: For the study of CQI determination in inference, support Option 1b (CQI is calculated based on target CSI with realistic channel measurement and potential adjustment).
Proposal 10: For performance monitoring, support the following monitoring options.
NW side monitoring based on the ground-truth CSI reported by UE.
UE side monitoring based on the recovery CSI indicated by NW.

3GPP TSG RAN WG1 Meeting #120bis	R1-2501919
Wuhan, China, April 7th – 11th, 2025

Source:	ZTE Corporation, Sanechips
Title:	Discussion on study for AI/ML CSI compression
Agenda Item:	9.1.4.1
Document for:	Discussion and Decision

Conclusion
In this contribution, we discuss the inter-vendor collaboration issues and some remaining issues in Rel-18. We have the following observations and proposals:
General views
To reduce the efforts for specification impact analysis and standardization, conduct comparison between Case 2 and Case 3 for potential down selection.
Conduct further evaluations and comparisons of different inter-vendor training collaboration options for feasibility study and potential down selection.
The specification impact analysis related to the inter-vendor training collaboration of AI/ML-based CSI compression should be deferred until the feasibility study and comparison of different options are concluded. 

Inter-vendor training collaboration
Direction A 
For option 3a-1 of Direction A, additional dataset for UE-side encoder retraining can be collected by UE implementations, while the sharing of target CSI from NW-side to UE-side is not essential.
For option 4-1 of Direction A, model backbone information should be shared from NW-side to UE-side to enable UE-side encoder training, validation, and testing.
For the determination of additional information sharing for options 3a-1 and 4-1, the feasibility of addressing UE-side/NW-side data distribution mismatch with respect to NW-side additional conditions should be concluded.
For inter-vendor-collaboration Options 3a-1 and 4-1 in Direction A, adopting SGCS statistic as the performance metric to facilitate UE-side offline engineering as a starting point.
For inter-vendor-collaboration Options 3a-1 and 4-1 in Direction A, the input data for evaluating the performance is the UE-side target CSI autonomously collected by the UE.
Observation 1: For options 3a-1 and 4-1 of Direction A, there is no proprietary information concern of disclosing performance requirements from NW side to UE side.
For option 4-1 of Direction A, the model backbone or structure information can only be shared to the opposite node if the proprietary information of the NW side can be maintained.
Observation 2: For Direction A, option 4-1 faces serious overhead concerns, while options 3a-1 incur much lower overhead due to the nature of their parameter exchange mechanisms.
For data distribution mismatch with respect to NW-side additional conditions for Direction A, it can be resolved by the indication of associated ID which implicitly abstracts the NW-side additional conditions.
For data distribution mismatch with respect to UE-side additional conditions for Direction A, it can be resolved by NW side timely data collection and subsequent update of delivered dataset/parameter to the UE side.

Direction B
For Direction B, the size of the encoder is much smaller than that of the training dataset and the transferring of parameters is not expected to be occurred frequently within a cell, making option 3b an attraction option in comparison to the dataset transfer related options from overhead consumption perspective.
For Direction B, it may be infeasible for the NW side to train multiple encoders tailored to different UEs due to the potential UE proprietary information disclosure risks and offline collaboration efforts.
For Direction B, the implementation of universal encoder across various UEs may or may not sacrifice the overall performance due to the lack of UE-specific encoder design and optimization.
Observation 3: For Direction B, there may be negative performance impact due to mismatch between NW side data distribution and UE side inference data distribution arising from the potential significant temporal gap between the model training phase and model deployment phase.
For Direction B, data categorization using associated IDs and continuous monitoring can be considered for addressing the data distribution mismatch issues.
Direction B may not carry risks of disclosing proprietary information if the model parameters to be transferred are widely recognized and devoid of device-specific design decisions.

Direction C
For Direction C, synthetic data generated under 3GPP’s statistical channel model can be a starting point for reference model training.
Observation 4: For Direction C, as synthetic data may not capture all characteristics of real world, mismatch between the distribution of the dataset used for reference model(s) training, UE-side data distribution, and NW-side data distribution happens. 
For Direction C, whether there is any performance impact arising from the data distribution mismatch depends on the generalization capability of the standardized reference model and thus further evaluations are needed.
For Direction C, model retraining based on data collected in real world can be performed to bridge the gap between synthetic data and field data.
Issues 8 and 9 are also applicable to Directions A and B due to the standardization requirement of reference model structure.

Model structure specification feasibility and scalability
Observation 5: If a proper model structure with a set of typical hyper-parameters is standardized, the impact on the inference performance is mainly due to the trained model parameters.
For model structure specification for Direction A Option 3a-1, Transformer should be adopted as the backbone structure for two-sided model of CSI compression as a starting point.
RAN1 and RAN4 can share the same standardized model structure for two-sided model of CSI compression to avoid the duplicated specification efforts, where the detailed specification and testing can be up to RAN4 study.
The so-called standardized reference model structure discussed in RAN1 for Direction A Option 3a-1 is assumed to be equivalent to the to-be-specified model structure, if specified, for performance requirements in RAN4.
Observation 6: With appropriate scalability solution performed to scale the dimension of the AI/ML model, the scalability over various CSI payload sizes, various bandwidths, and various Tx port numbers can be achieved by training the AI/ML model with mixed data samples.
The scalability solution for scaling the input/output dimension of the AI/ML model (e.g., pre/post-processing of truncation/padding, various quantization granularities, and adaptation layer in the AL/ML model) causes minor or no effects on the model structure specification.
For the choice of token dimension and feature dimension, Alt 1 is more appropriate considering the inherent channel correlation within the subbands in the frequency domain. 
For the choice of token dimension and feature dimension, Alt 3 may involves additional pre-processing steps like padding to handle non-whole divisions and its performance benefit needs further validation.
Observation 7: For scalability over the feature dimension, Alt 1 (specific embedding layer for each feature size) necessitates a specific embedding layer (e.g., a Fully-Connected layer) for each distinct feature size, leading to an increased number of model parameters and computational complexity. 
Observation 8: For scalability over the feature dimension, Alt 2 (a common embedding layer with padding) employs padding operations to fulfil the predefined feature sizes without introducing extra complexity to either the model architecture or inference process.
Observation 9: For scalability over the token dimension, Alt 2 (padding at the input) is a simple way to operate and incurs minor performance degradation compared with the dedicated model design.
Observation 10: For scalability over payload configurations, compared with Alt 1 (specific output linear layer for each payload configuration), Alt 2 (truncation/masking of the output linear layer output) can reduce number of model parameters by over 40% with only around 1% SGCS degradation for two layers.
For model structure scalability study for temporal domain Case 0, further consider
For the choice of token dimension and feature dimension,
Alt 1: Use subband as the token dimension and Tx port as a feature dimension
The number of tokens varies with the number of subbands.
For scalability over the feature dimension, 
Alt 2: A common embedding layer with padding
For scalability over the token dimension, 
Alt 2: Padding at the input
For scalability over payload configurations,
Alt 2: Truncation/masking of the output linear layer output

Evaluation on model scalability
For studying the scalability of standardized model structure, temporal domain Case 0 with spatial-frequency domain input can be adopted as a primary use case and Transformer is suggested as the basic backbone structure.
Observation 11: For model scalability over 32 Tx ports and 16 Tx ports, compared with generalization Case 1 that AI model is trained and tested on the same configuration of 16 Tx ports, generalization Case 2 (i.e., AI model is trained on 32 Tx ports and tested on a different configuration of 16 Tx ports) shows large performance degradation of -4.85% for layer 1 SGCS and -7.90% for layer 2 SGCS.
Observation 12: For model scalability over 32 Tx ports and 16 Tx ports, compared with generalization Case 1 that AI model is trained and tested on the same configuration of 16 Tx ports, generalization Case 3 shows marginal performance degradation of -0.44% for layer 1 SGCS and -0.52% for layer 2 SGCS for testing on 16 Tx ports. In addition, less performance loss is observed for testing on 32 Tx ports, i.e., -0.26% for layer 1 SGCS and -0.44%  for layer 2 SGCS.
For the scalability verification of AI/ML based CSI compression over various numbers of Tx ports, the model scalability can be achieved by training the model with mixed dataset from different numbers of Tx ports.
Observation 13: For model scalability over low, medium, and high CSI payload sizes, compared with generalization 1 case that AI model is trained and tested on the same configuration, AI model trained with mixed dataset of multiple CSI feedback payload sizes can achieve minor performance loss of -0.76% for layer 1 SGCS and -0.93% for layer 2 SGCS under high payload size.
Observation 14: For model scalability over low, medium, and high CSI payload sizes, compared with generalization 1 case that AI model is trained and tested on the same configuration, AI model trained with mixed dataset of multiple CSI feedback payload sizes can achieve over 1.58% SGCS gains for low payload sizes and over 2.46% SGCS gains for medium payload sizes, respectively
For the scalability verification of AI/ML based CSI compression over various CSI feedback payload sizes, the model scalability can be achieved by training the model with mixed dataset from different CSI feedback payload sizes.
Observation 15: For model scalability over different configurations of {Tx ports, subbands, payload sizes}, compared with Case 2 (i.e., dedicated model structure), Case 1-1 with a single parameter set shows minor performance degradation of -1.45% for layer 1 SGCS and -1.82% for layer 2 SGCS in average. Case 1-1 can achieve -3.79%~+1.27% SGCS gain for two layers over Case 2.
Observation 16: For model scalability over different configurations of {Tx ports, subbands, payload sizes}, compared with Case 2 (i.e., dedicated model structure), Case 1-2 with different parameter sets shows minor performance degradation of -0.24% for layer 1 SGCS and -0.61% for layer 2 SGCS in average. Case 1-2 can achieve -2.34%~+0.92% SGCS gain for two layers over Case 2.
For the scalability verification of AI/ML based CSI compression over various configurations of {Tx ports, subbands, payload sizes}, the model scalability can be achieved by training the model with mixed dataset from different configurations.

Remaining issues for AI/ML CSI compression
For ground-truth reporting for NW-side data collection, support to further study enhanced Rel-16 eTypeII codebook design to achieve high-resolution CSI.
To enable high-quality data collection from UE to network, at least support
UE reports data quality related information to NW, e.g., SINR, CQI, positioning information
NW configures a threshold of data quality to UE and UE only reports the qualified data to NW
For CQI determination, at least prioritize the specification impact discussions on Options 1a and 1b.
Prioritize to study the specification impacts on at least the following case for model performance monitoring, 
NW-side monitoring based on the target CSI with realistic channel estimation associated to the CSI report, reported by the UE.
In CSI compression using two-sided model use case, deprioritize the study on UE-side monitoring in Rel-19 study phase.
Observation 17: For the definition of KPIDiff defined in TR 38.843, KPIDiff = (KPIActual > KPIth 2, KPIGenie < KPIth 3) and KPIDiff = (KPIActual < KPIth 2, KPIGenie > KPIth 3) correspond to the missed detection probability and false alarm probability, respectively.
In CSI compression using two-sided model use case, further study a set of appropriate monitoring methodology and reasonable monitoring metrics for the given KPI under rank>1.
Regarding training data collection for temporal domain aspects Case 3, prioritize Option 2 (i.e., the target CSI for training is derived based on the measured CSI of the future slot(s)) due to its reduced data collection overhead and potentially higher CSI compression performance.
For temporal domain aspects Case 3, prioritize separate prediction and compression, where Option 1 (i.e., the monitoring label is derived based on the predicted CSI of the future slot(s)) is used for monitoring CSI compression only.

3GPP TSG-RAN WG1 Meeting #120bis	Tdoc R1-2501924
Wuhan, China, April 7th – April 11st, 2025

Agenda Item:	9.1.4.1
Source:	Ericsson
Title:	AI/ML for CSI compression and other aspects on AI/ML model/data
Document for:	Discussion, Decision
1 

Conclusion
In the previous sections we made the following observations: 
Observation 1	Without access to the actual NW decoder, for inter-vendor training option 3, 4 and 5, RAN4 pre-deployment testing cannot provide any assurance that the two-sided CSI compression functionality performs well in the field.
Observation 2	RAN1 should consider robust performance testing and monitoring that is not dependent on offline collaboration between UE and NW vendors.
Observation 3	Eventual KPI based monitoring has low complexity, low overhead, and can capture network MU-MIMO performance. The NW can perform frequent monitoring of eventual KPIs via NW implementation-based solutions and use it as a first step for detecting potential AI/ML feature/functionality failure.
Observation 4	It is necessary to specify UE reporting high resolution target CSI to enable NW-side intermediate KPI based monitoring of the two-sided CSI-compression model performance.
-	Alternative 1: Evaluating end-to-end intermediate KPI(s).
-	Alternative 2: Evaluation the KPIs related to the encoder output.
Observation 5	NW-side monitoring of the two-sided CSI-compression model based on target CSI reporting is expected to be executed infrequently (e.g., event triggered or periodically with a large periodicity), hence, the monitoring data collection overhead for this model monitoring method is in general not an issue.
Observation 6	UE-side based monitoring of eventual KPI is not feasible as the UE does not have CSI-RS precoding information and cell shaping information nor can it capture the model’s performance in MU-MIMO which is the main motivation for AI/ML based CSI reporting.
Observation 7	Input/output data distribution-based monitoring method put requirements on computation power and memory at the UE side. Data drifts detected at the UE-part of a two-sided model does not necessarily mean that the two-sided model is not functioning.
Observation 8	It is non-trivial to define conditions and measurable statistic KPI metrics to represent input/output data distribution for a data drift detection at UE. Hence, the feasibility of supporting UE reporting its real-time data drift detection related metrics and the testability of the report are questionable.
Observation 9	For CSI compression using two-sided model use case, the method of UE-side monitoring based on a proxy CSI construction model at the UE (Case 2-1) may not provide accurately monitoring results, since the proxy intermediate KPI statistics derived/obtained from the proxy model may not reflect the actual intermediate KPI statistics of the two-sided CSI-compression model, and it may impose an additional set of model LCM overhead and signaling overhead for training/deploying/monitoring/testing the proxy model.
Observation 10	For CSI compression using two-sided model use case, the method of UE-side monitoring via direct estimation of intermediate KPI (Case 2-2) may not provide accurately monitoring results, since the proxy intermediate KPI statistics derived/obtained from the intermediate KPI estimator may not reflect the actual intermediate KPI statistics of the two-sided CSI-compression model, and it may imposes an additional set of model LCM overhead and signaling overhead for training/deploying/monitoring/testing the intermediate KPI estimator.
Observation 11	For CSI compression using two-sided model use case, with all considered inter-vendor training collaboration options, the method of UE-side monitoring based on the approximation of the CSI reconstruction model output indicated/provided by NW does not seem to be feasible in practice, since it can significantly increase the signalling overhead and UE complexity, and the monitoring accuracy for this method has not been evaluated.
Observation 12	For CSI compression using two-sided model use case, comparing to the method of UE-side monitoring via direct estimation of intermediate KPI (e.g., SGCS), the method of UE-side monitoring based on estimation of monitoring output (e.g., SGCS range indicatin), may provide better monitoring accuracy and generalization performance. Further study is needed.
Observation 13	For CSI compression using two-sided model use case, similar to case 2-1 and case 2-2, the method of UE-side monitoring based on estimation of monitoring output may imposes an additional set of model LCM overhead for training/deploying/monitoring/testing the monitoring output estimator.
Observation 14	For CSI compression using two-sided model use case, the method of UE-side monitoring based on precoded RS transmitted from NW based on the output of the CSI construction model introduced RS signaling overhead, latency and processing complexity without clar benefit.
Observation 15	For CSI compression using two-sided model use case, it is not feasible for UE-side alone to detect the error cause of performance degradation of the two-sided model.
Observation 16	In the AI-TSF compression Case 3, where UE performs prediction in a separate step before compression (i.e., joint UE-sided CSI prediction followed by two-sided TSF CSI compression), under ideal channel estimation assumption, a small to moderate SGCS gain compared to the UE-sided AI-based CSI prediction with Rel-18 MIMO eType II codebook for CSI feedback, is observed.
Observation 17	For AI CSI compression Case 3, the computational complexity ratio in terms of FLOPs between the AI model and legacy Rel-18 eType II is around 300 for payload Category X and around 200 for payload Categories Y and Z.
Observation 18	Compared to Rel-16 eType II benchmark, AI CSI compression Case 2 provides 8.5% and 20.4% SGCS gain at CSI payload X, for layer 1 and layer 2, respectively. The gain decreases as the CSI payload size increases.
Observation 19	Compared to a non-AI based benchmark (Rel. 16 eType II with W1, past Wf, past ), AI CSI compression Case 2 provides 5.9% and 17.1% SGCS gain at CSI payload X, for layer 1 and layer 2, respectively. The gain decreases as the CSI payload size increases.
Observation 20	Compared to the CSI compression Case 0, AI CSI compression Case 2 provides 1.7% and 2.4% performance gain at CSI payload X, for layer 1 and layer 2, respectively. The gain decreases as the CSI payload size increases.

Based on the discussion in the previous sections we propose the following:
Proposal 1	For inter vendor training collaboration option 3a-1 in Direction A and Direction C, the following issue shall be studied before concluding their feasibility:
	How to represent a standardized reference model (structure + parameters) and/or a standardized reference model structure in 3GPP specifications?
Proposal 2	Conclude that it is necessary to specify UE reporting high resolution target CSI to enable NW-side intermediate KPIs based performance monitoring and performance degradation error cause detection for two-sided CSI-compression use case.
Proposal 3	In CSI compression using two-sided model use case, capture in TR that ground-truth CSI report based on enhancements of the eType-II format with new parameters shall be defined to ensure high-accuracy model performance monitoring and error cause detection at the NW-side. Potential specification impact include:
	Define the target-CSI format (e.g., Rel16 eType II CB with new parameters) for NW-side data collection (can reuse the ground truth defined for model training data collection)
	Mechanisms (e.g., RRC-message based methods) to support UE reporting the target CSI together with the encoder output for NW-side data collection for performance monitoring.
	Signaling and configuration for event triggered and periodical data collection at the NW-side.
Proposal 4	For CSI compression using two-sided model use case, for any UE-sided performance monitoring method (if its feasibility and performance are justified), to enable the testability of the quality of the UE reported monitoring metrics, at least the following spec impact are identified:
	The format of the monitoring metrics
	Singaling and mechanisms for UE reporting monitoring metrics
	RAN4 testing of the quality of UE reported monitoring metrics
	Data collection at the NW-side based on UE reporting monitoring data samples including target CSI, encoder output and monitoring metrics (to enable NW-side test the quality of the reported monitoring metric in the field).
Proposal 5	Further study methods to reduce the AI computational complexity for AI CSI compression Case 3.
Proposal 6	Deprioritize AI CSI compression use case 2.

3GPP TSG-RAN WG1 Meeting #120-bis	R1- 2501934
Wuhan, China, April 7th – April 11th, 2025

Agenda Item:	9.1.4.1
Source:	NVIDIA
Title:	Additional study on AI-enabled CSI compression and other aspects of AI model       and data
Document for:	Discussion
1	

Conclusion
From RAN1 perspective, the study progress of model identification and model transfer/delivery in Agenda item 9.1.4.2 are sufficient. 
New work/discussion in AI 9.1.4.2 can be triggered by LS from other working group(s) or other TSG(s), if any
The work on pCR (if any) to capture the output of AI 9.1.4.2 to be done in future meeting(s)
Other on-demand work (if any)
From RAN1 perspective, there is no consensus in Rel-19 on the need of standardized solution for model delivery/transfer of one-sided model
In RAN1, if needed, whether model identification is needed for CSI compression or not is to be decided in Agenda item 9.1.4.1
In RAN1, if needed, whether standardized solution(s) for model transfer/delivery are needed for CSI compression or not is to be decided in Agenda item 9.1.4.1

3GPP TSG RAN WG1 #120bis		R1-2501949
Wuhan, China, April 7th – 11th, 2025

Agenda Item:	9.1.4.1
Source:	Google
Title:	ML based CSI compression
Document for:	Discussion/Decision

Conclusion
In this contribution, we provided discussion on AI/ML based CSI compression. Based on the discussion, the following proposals are provided.
Proposal 1: For case 2 and case 3, support the UE to report W1 and compressed W2 for a configured rank
The compressed W2 is calculated based on the AI/ML
NW can configure the UE to report RI and CQI based on precoded CSI-RS resources after receiving the W1 and compressed W2
Proposal 2: The priority for non-ML based CSI report should be higher than the priority of ML based CSI report.
Proposal 3: Support the CPU occupancy rule for ML based CSI based on two types processing unit
Type1 CPU: a measurement processing unit (MPU) used for channel estimation and pre-processing
Type2 CPU: an inference processing unit (IPU) used for inference for ML based CSI
Proposal 4: For CSI report based performance monitoring, the following spec impact should be considered:
New report quantity PMI only should be introduced, where UE reports the PMI based on a configured rank
Support a further enhancement to report subband L1-SINR in addition to the PMI to facilitate the precoder and MCS selection for PDSCH
With regard to measurement accuracy in low SINR case, support the UE to report a state of CSI indicating the CSI is invalid for performance monitoring
With regard to joint ML based CSI compression and prediction, support to configure whether the UE should perform the CSI predication based on ML or non-ML and the CSI quantization based on ML or non-ML for separate performance monitoring for ML based CSI prediction and ML based CSI compression.
Proposal 5: For SRS based performance monitoring, the following spec impact should be considered:
Support to configure the SRS linked with a CSI-RS report configuration for ML based CSI, where the UE uses the same ports including antenna virtualization scheme to transmit the SRS and to receive the CSI-RS for ML based CSI
Support burst based SRS with frequency hopping to facilitate the performance monitoring for wideband channel for coverage-limited UE
Proposal 6: For UE-side monitoring, the following spec impact should be considered: 
Introduce the hypothetical BLER as the metric for performance calculation
Configuration of precoded CSI-RS for hypothetical BLER calculation
Proposal 7: For NW-side data collection, support the following enhancement for Rel-16 eType2 codebook and Rel-18 eType2 codebook for PMI prediction
Additional O1 and O2 value
Additional codebook parameter combinations at least for rank 3 and 4
UE reports the measured CSIs instead of predicted CSI for Rel-18 eType2 codebook for PMI prediction
Additional number of DD basis for Rel-18 eType2 codebook for PMI prediction
Proposal 8: For NW-side data collection, the potential spec impact for the configured number of layers can be based on introducing a constraint for the RI restriction to configure one RI for the CSI report.
Proposal 9: For NW-side data collection, support the UE to report the following additional information 
L1-SINR measured based on the CSI-RS to indicate the measurement quality
Timing for the transmission occasion(s) of the CSI-RS for the measured CSI
Proposal 10: Support to configure multiple uplink power control parameter sets for SRS
The SRS can be used for data collection and other functionalities, where different uplink power control parameter sets can be used for different functionalities
The NW can indicate which one is applied for the SRS transmission dynamically
Proposal 11: Support both NW-configured and UE-requested based CSI-RS for UE-side data collection
Corresponding number of CPUs should be occupied for the CSI-RS 
Associated ID should be provided
Proposal 12: Support hybrid AI/ML based and non-AI/ML based CSI measurement and report
UE reports the CSI based on AI/ML if it reports a small RI and the UE can report the CSI based on Type1 codebook if it reports a large RI
Proposal 13: Conclude that from RAN1’s perspective, both option 4-1 and 3a-1 for Direction A are feasible.
Proposal 14: Conclude that Direction B is deprioritized.
Proposal 15: For direction C, Option 2 (Latent adaptation) is preferred compared to Option 1 (Input/output adaptation with additional layers)
Proposal 16: Correct the agreed transformer structure with the following changes:
Change the order of the normalization layer and multi-head attention layer
Change the order of the normalization layer and MLP layer
Proposal 17: Support to define the standardized model structure for case 0 for single-CRI based report as follows:
The dimension of a token is the CSI for a subband for 2 ports
The number of tokens is Np/2 * Nsubband
Np indicates the number of ports for CSI
Nsubband indicates the number of subbands
Proposal 18: Support to define the standardized model structure for case 2 and case 3 for single-CRI based report as follows:
The dimension of a token is the CSI for a subband for 2 ports for a transmission occasion
The number of tokens is Np/2 * Nsubband * NTO
Np indicates the number of ports for CSI
Nsubband indicates the number of subbands
NTO indicates the number of measured/predicted CSI occasions
Proposal 19: For multi-CRI based CSI report, study the following options:
Option 1: The CSI compression is performed per CRI separately
Option 2: The CSI compression is performed across all CRIs jointly

3GPP TSG RAN WG1 Meeting #120bis                                                                             R1-2501958
Wuhan, China, April 7th – 11st, 2025
Source:	TCL
Title:	Discussion on AI/ML beam management	
Agenda Item:	9.1.4.1
Document for:	Discussion and Decision

Conclusions
This contribution has led to the following observations:
Observation 1: The temporal domain compression Case 2 will improve the CSI compression efficiency.
Observation 2: Certain additional information may impact the priority of AI/ML CSI reporting for CSI compression:
AI/ML LCM procedure
AI/ML characteristics for CSI compression
AI/ML reporting contents
Other factors that may be pertinent to the value of AI/ML-specific priority should not be precluded.
Observation 3: The memory and computational resources as well as the power consumption may become bottlenecks for UE when implementing AI/ML models for CSI compression and restoration.
Observation 4: The computational time for CSI restoration at NW side should be estimated and reconfigured to the UE. This may include the time duration for both model activation and model inference.
Observation 5: The CSI transferring for model monitoring in AI/ML based CSI compression introduces considerable overhead, RAN 1 should strive for an efficient signaling and overhead reduction mechanism for the CSI transferring for model monitoring purpose.
Observation 6: The reliability of the overhead reduction scheme for CSI transferring for model monitoring in AI/ML based CSI compression should be guaranteed.

Furthermore, the following proposals have been put forward:
Proposal 1: RAN1 should study the benefits and potential issues if temporal domain compression Case 2 is adopted, especially in the following aspects:
The overall performance gain considering both the accuracy loss and the overhead reduction of the temporal domain compression Case 2.
The performance monitoring metrics of the temporal domain compression Case 2.
The potential problem of error accumulation and its solutions.
Proposal 2: RAN1 should further study the following specification impact on the monitoring of temporal domain compression Case 3:
the monitoring data report configuration, e.g., new report quantity, indication of data type.
the format of CSI quality
Proposal 3: Investigate the following aspects for mitigating the impact of UCI loss.
How to define the UCI event triggering the CSI synchronization.
NW-signaling to reset the CSI information at UE to predefined values.
Proposal 4: In the case of CSI compression using a two-sided model, the design of an AI/ML-specific CSI-RS resource and CSI reporting configuration that may be compatible with the traditional CSI reporting scenario should be considered in the following aspects:
AI/ML-specific CSI-RS resource configuration for CSI compression
AI/ML-specific fields in CSI-ReportConfig IE
Dedicated report quantities and report configurations for AI
Proposal 5: The definition of AI/ML-specific priority for CSI reporting in relation to CSI compression should be considered in comparison with the traditional CSI priority rules
Proposal 6: When the UE supports both AI/ML and non-AI/ML CSI reporting, it is necessary to redefine the priority rule considering different types of CSI reporting.
Proposal 7: The study of how to describe the capabilities of a UE to implement AI/ML models for inference on CSI compression and calculations should be undertaken.
Proposal 8: Option 1 and 2 should not be standardized, at least they are out of the scope of R19.
Proposal 9: RAN 1 should down select among option 3, 4 and 5 considering if unified model format or structure is shared between the NW and UE side model respectively.
For option 4, there may be no need for offline-engineering.
Option 3 and 5 are preferred if model structure or format is exchanged from NW to UE.
Proposal 10: Ran1 should consider whether/how to transmit the raw CSI for monitoring purpose.
FFS: The overhead reduction scheme based on spectral or temporal processing.

3GPP TSG RAN WG1 #120bis		                             R1-2501971
Wuhan, China, April 7th – 11th, 2025

Source:	CATT
Title:	Discussion on AI/ML-based CSI compression
Agenda Item:	9.1.4.1
Document for:	Discussion and Decision

Conclusions
In this contribution, we provided our views on specification of model structure and remaining issues on the three directions for inter-vendor training collaboration. We also provide our evaluation results on model scalability and our views on specification impacts for AI/ML based CSI compression. The following observations and proposals are provided:
Observation 1: For Case 0 encoder model structure standardization, hyper-parameter combination provided in Table 1 can provide good SGCS performance with relatively low complexity.
Observation 2: Case 1 scalable structure demonstrates substantial performance improvement (+13.33%~+17.83% in small payload, +3.56%~+7.44% in moderate payload, +1.2%~+4.04% in large payload) against R16 eType II baseline.
Observation 3: Case 1 scalable structure incurs minor performance degradation (-3.35%~+1.19%) against Case 2 dedicated structure in all configurations.
Observation 4: Case 1 scalable structure incurs small complexity increase (+0.6978M FLOPs, +6.8%) compared with the most complex Case 2 dedicated structure. Compared with the total FLOPs of all 12 configurations, 87.98% FLOPs are reduced by the scalable structure.
Observation 5: For CSI compression Option 4-1, quantization misalignment can lead to -54.95%~ -64.51% SGCS performance degradation.
Observation 6: Regarding inter-vendor-collaboration Direction A Option 3a-1 for CSI compression, additional information of target CSI dataset exchanged from NW-side to UE-side is not needed.
Observation 7: Regarding dataset and parameter sharing from NW-side to UE or UE-side, if OTA signaling is considered, Option 4-1 has more overhead concern than Option 3a-1.
Observation 8: The scalable structure for SF domain CSI compression has obvious advantage in performance-complexity trade-off over the dedicated structure.
Observation 9: Regarding the scalable structure for SF domain CSI compression,
when increasing the complexity from 3.097004 M to 10.88606 M FLOPs, there is 4.5~5.47% (1.4~3.35%) SGCS performance gain for the configuration of 32 (16) ports, respectively. 
when increasing the complexity from 10.88606 M to 20.58668 M FLOPs, there is 1.02-1.68% (-0.13~0.44%) SGCS performance gain for the configuration of 32 (16) ports, respectively. 
Observation 10: Regarding the scalable structure for SF domain CSI compression, hyper-parameter config 2 has the best performance-complexity trade-off.
Observation 11: In CSI compression using two-sided model use case, UE-side performance monitoring Case 2-1 should not be used for the inter-vendor training collaboration options other than the ones (Option 3/Option 5) with UE-side transmit model/mode parameters to the NW-side.
Observation 12: In CSI compression using two-sided model use case, for data collection for model training, enhancement on CSI-RS is not needed.
Observation 13: Regarding training data collection for temporal domain aspects Case 3, both Option 1 and Option 2 can use Rel-18 Doppler eType II codebook for CSI report with enhancement on the time domain position of the  consecutive slot intervals. Option 1 and Option 2 might have different numbers of CSI-RS resources and different time domain configurations.

Proposal 1: For studying the standardized model structure, prioritize the standardization on Case 0 model structure with model input in spatial-frequency.
Proposal 2: For AI/ML based CSI comparison, to support different bandwidths for inference, in additional to support scalable models, the following methods are also considered:
Obtaining CSI feedback for larger bandwidths through multiple inferences;
Adjusting the input bandwidth;
Cropping the model output.
Proposal 3: Scalability over various number of Tx ports, CSI feedback payload size and bandwidth can be achieved by
Using subband as token dimension and Tx port as feature dimension
Specific embedding layer for each number of Tx ports
Padding at the input for different number of subbands
Specific output linear layer for each payload configuration
Proposal 4: Regarding inter-vendor-collaboration Direction A Option 4-1 for CSI compression, backbone/structure-related information is adopted as additional information.
Proposal 5: Regarding inter-vendor-collaboration Direction A Option 3a-1/4-1 for CSI compression, support at least one of the following options:
Option 1: Standardize quantization method for CSI feedback, e.g., scalar quantization.
Option 2: Quantization related information is adopted as additional information, at least for Option 4-1.
Proposal 6: In CSI compression using two-sided model use case, support NW-side monitoring based on the target CSI reported by the UE via legacy eT2 codebook or eT2-like high-resolution codebook.
Proposal 7: In CSI compression using two-sided model use case, for NW-side monitoring Case 1, further study the signaling and procedures for reporting target CSI, with the following two options considered:
Option 1: The target CSI is reported separately from its associated CSI feedback;
Option 2: The target CSI is reported together with its associated CSI feedback.
Proposal 8: In CSI compression using two-sided model use case, performance monitoring at UE-side based on reference model or proxy model can be deprioritized.
Proposal 9: In CSI compression using two-sided model use case, if eventual KPI is adopted as monitoring metric, how to exclude the impacts of other factors other than AI/ML model performance should be studied.
Proposal 10: In CSI compression using two-sided model use case, support UE-side monitoring based on precoded RS (e.g., CSI-RS, DMRS) transmitted from NW based on the output of the CSI reconstruction model.
Proposal 11: In CSI compression using two-sided model use case, regarding separately monitoring of CSI prediction and CSI compression for temporal domain aspects Case 3, for NW-side monitoring for CSI compression, the specification impact of monitoring of CSI compression includes:
For Option 1, configuration and report for target CSI of CSI compression corresponding to the output of CSI prediction;
For Option 2b, configuration and report of target CSI of CSI compression corresponding to the measured CSI of the future slot(s) and CSI feedback based on the measured CSI of the future slot(s).
Proposal 12: In CSI compression using two-sided model use case, regarding separately monitoring of CSI prediction and CSI compression for temporal domain aspects Case 3, for UE-side monitoring for CSI compression, the specification impact of monitoring of CSI compression includes configuration of CSI-RS for CSI measurement of the future slot(s) for Option 2b.
Proposal 13: In CSI compression using two-sided model use case, regarding separately monitoring of CSI prediction and CSI compression for temporal domain aspects Case 3, study whether the selection of monitoring methods for CSI prediction needs to be comprehensively considered in conjunction with the monitoring methods for CSI compression.
Proposal 14: In CSI compression using two-sided model use case, L1 signaling based reporting of target CSI is supported for NW-side data collection for performance monitoring.
Proposal 15: In CSI compression using two-sided model use case, both L1 signaling based reporting and RRC signaling based reporting are supported for ground-truth CSI for NW-side data collection for model training.
Proposal 16: In CSI compression using two-sided model use case, for L1 signaling based reporting of ground-truth CSI/target CSI for NW-side data collection, legacy CSI feedback framework can be reused for Case 0.
Proposal 17: In CSI compression using two-sided model use case, for NW-side data collection for model training, collecting ground-truth data in type of precoding matrix is supported.
Proposal 18: In CSI compression using two-sided model use case, for NW-side data collection for model training, NW determines the number of layers for ground-truth CSI data collection.
Proposal 19: In CSI compression using two-sided model use case, at least for measured CSI of the future slot(s) collected as target CSI for training, if Rel-18 Doppler eType II codebook is used, enhancement on the time domain position of the  consecutive slot intervals should be introduced.
Proposal 20: In CSI compression using two-sided model use case, for L1 signaling based reporting of ground-truth CSI/target CSI for NW-side data collection, multiple time instants CSI can be reported in the same report for Case 3.
Proposal 21: In CSI compression using two-sided model use case, quantization alignment between UE-side and NW-side in a 3GPP non-transparent manner is supported.
Proposal 22: In CSI compression using two-sided model use case, legacy CSI reporting principles are reused as much as possible.
Proposal 23: In CSI compression using two-sided model use case, if CQI in CSI report is configured, for CQI determination in CSI report, one of the sub options of Option 1 is adopted:
Option 1: CQI is NOT calculated based on the output of CSI reconstruction part from the realistic channel estimation, including
Option 1a: CQI is calculated based on target CSI with realistic channel measurement
Option 1b: CQI is calculated based on target CSI with realistic channel measurement and potential adjustment
Option 1c: CQI is calculated based on legacy codebook
Proposal 24: For CQI reporting in CSI compression using two-sided model use case, reuse the same quantization scheme as in legacy CQI reporting.
Proposal 25 Introduce a new type of processing unit shared among different AI functionalities, which is independent of legacy CPU.
Proposal 26：AI/ML-based CSI reporting occupies both legacy CPU and new types of processing unit. The CPU counting for legacy CSI reporting and AI/ML-based CSI reporting shares the same CPU pool.

3GPP TSG RAN WG1 #120-bis		                                    R1-2502037
Wuhan, China, April 7th – 11th, 2025

Source:	Ofinno
Title:	Views on UCI loss mitigation
Agenda Item:	9.1.4.1
Document for:	Discussion and Decision

Conclusion
In this contribution, we make the following proposals.
Support NW-signaling to reset historical CSI information at UE to initial values, e.g., UE flush all CSI historical information on a CSI report once a reset is triggered for the CSI report. 
Support gNB signaling to reset historical CSI information to the initial state at UE (e.g., all zero or known random numbers):
For periodic CSI reports, support RRC-configured periodical resets, e.g., a reset periodicity
For semi-permanent CSI report, support SPS grant based resets
For aperiodic CSI report, support a DCI field indicating the resets
Proposal 3: Support gNB signaling to reset (or retransmit as 2nd priority but not both) historical CSI information to the most recent one at UE (e.g., the one before the UCI loss):
Support P/SP/AP-CSI
For PDSCH with an associated DCI, reuse the “DMRS sequence initialization” field to indicate resetting
For PDSCH without an associated DCI, gNB initializes the DMRS with “DMRS sequence initialization” as either 0 or 1. Use one of the values to indicate resetting
UE blindly decodes the DMRS sequence initialization.

3GPP TSG RAN WG1 #120bis	R1-2502073
Wuhan, China, April 7th – 11th, 2025

Agenda item:	9.1.4.1
Source:	NEC
Title:              	Discussion on CSI compression
Document for:	Discussion and Decision

Conclusion
In this contribution, we provide our views mainly on inter-vendor training collaboration and necessary specification impacts. Specifically, we have the following observations and proposals:
Proposal 1: For Direction A, option 3a-1 w/ target CSI, support NW-side to adopt the training dataset mix to ensure the generalization ability of the trained encoder.
Proposal 2: RAN1 to prioritize studying Direction A and C over Direction B.
Proposal 3: For Direction A, the following options should be prioritized.
Option 4-1.
Option 3a-1 w/ target CSI sharing form NW to UE.
Proposal 4: Support NW-side monitoring based on the target CSI reported by UE (Case 1).
Proposal 5: For NW-side monitoring based on the target CSI reported by UE (Case 1), for reporting the compressed CSI and corresponding target CSI (i.e., ground-truth CSI), the following can be considered:
The compressed CSI and corresponding target CSI are reported in the same report.
The compressed CSI and corresponding target CSI are reported in separate reports.
Observation 1: Compared to NW-side monitoring, the most significant advantage of UE-side monitoring based on the output of the decoder at UE is significant reduction in reporting overhead.
Observation 2: If the decoder at UE side is the same as the actual decoder used at NW-side, or a reference model provided by NW, the monitoring result obtained by UE-side monitoring is accurate in principle. And this would only bring a small amount of additional burdens to UE.
Observation 3: If the decoder at UE side is a proxy decoder developed by UE side, only a moderate amount of additional effort is required to use the monitoring results of NW-side monitoring as a baseline for assessing the performance the proxy decoder. And the update of the proxy decoder will be together with the update of the encoder at UE side.
Proposal 6: Support UE-side monitoring based on the output of the CSI reconstruction model at the UE (Case 2-1).
Proposal 7: Reset of historical CSI information should be supported to address UCI loss. And the definition, determination or indication of the reset value (e.g., previous historical CSI information, initial state) need to be further studied.
Proposal 8: For the case of UCI loss is known to both NW and UE, e.g., CSI report is dropped, NW-signaling is not needed for reset of historical CSI information at UE side.
Proposal 9: At least for Case 2, further study effective availability of historical CSI information over time.
Proposal 10: For data collection for NW-first training and NW-side monitoring, L1 signaling (e.g., the existing CSI report framework) can be used to report ground-truth CSI.
Proposal 11: For NW-side data collection, RAN1 to discuss and evaluate whether ground truth reporting from UE can be transmitted over L1 or higher layer signaling (e.g. RRC).
Proposal 12: RAN1 to select the model identification procedure based on the outcome of the study to address inter-vendor collaboration issues for CSI compression use case, for example, MI-Option 1/2 to be adopted if Option 4 is agreed, and a combination of MI-Option 1/2 and MI-Option 3 to be adopted if Option 3a-1 is agreed.
Proposal 13: Support Alt. B for model transfer methodology z4. 
Proposal 14: Support RRC signaling for transfer of AI/ML model parameters from gNB to UE. And discuss further if UE can store the model parameters when it goes into idle/inactive state so that gNB can avoid providing the full model parameters to the UE when UE switches to connected state.
Proposal 15: For model management purposes, support model structure within the first indication to be identified/associated with a model structure id and the model parameters in the second indication to be identified/associated with a model id value, i.e., Option 2 and a transferred AI/ML model is uniquely identified at least within the scope of the cell, where AI/ML model transfer occurs, using both model structure id and model id value.

3GPP TSG RAN WG1 #120-bis			R1-2502096
Wuhan, China, April 7th – 11th, 2025

Agenda item:	     9.1.4.1
Source:	Tejas Networks Ltd.
Title:	Discussion on AI/ML for CSI Compression
Document for:	Discussion 

Conclusion
The following observation proposals are made in this contribution:
Proposal 1: For the EVM of temporal domain CSI compression, consider the following assumptions for the CSI generation part and CSI reconstruction part, respectively:
CSI generation part at t=T: 
Model input: Pre-processed channel matrix is given as a input to the CsiNet encoder.
First layer of encoder is Convolutional layer with real and imaginary parts of channel as its input.  This layer uses 33 kernel size to generate two feature maps. Following the convolutional layer, we reshape the feature maps into a vector and use a fully connected layer to generate the codeword s, which is a real-valued vector of size M.

Model output: Encoder transforms channel matrix into a vector of size M.
CSI reconstruction part at t=T+∂ (where ∂ is an uplink latency)
Model input: Codeword of size Mis given as a input to the CsiNet decoder.
Model output: First layer of decoder is dense fully connected layer which gives the initial estimate of the channel followed by several RefineNet units gives the CSI.
       FFS: Study the effect of quantization
Observation 1: For discussion on performance degradation due to UE-side / NW-side data distribution mismatch with respect to UE side additional condition, CSI-Net is robust to the antenna Tilt angles.
Proposal 2: For discussion on performance degradation due to UE-side / NW-side data distribution mismatch with respect to UE side additional condition (issue 4 and 6), consider Model ID to represent the models which are robust to data distribution mismatch.
Observation 2: For the evaluation of temporal domain aspects of AI/ML-based CSI compression using two-sided model, for case2, Consider Scenario A (No UCI Loss) as the baseline for comparing all other scenarios.

Observation 3: For the evaluation of temporal domain aspects of AI/ML-based CSI compression using two-sided model, for case2, scenario B (UCI Loss Known at NW, Unknown at UE, with Mitigation at NW) should measure the effectiveness of NW-side mitigation techniques and assess how much they improve overall system performance despite the lack of UE-side correction.
Observation 4: For the evaluation of temporal domain aspects of AI/ML-based CSI compression using two-sided model, for case2, Scenario C (UCI Loss known at NW and UE, with mitigation at Both) should be evaluated for its potential to significantly improve performance through coordinated mitigation, it is preferred approach if signalling overhead is manageable.
Proposal 3: For the evaluation of temporal domain aspects of AI/ML-based CSI compression using two-sided model, for case2:
Scenario A provides the baseline for comparison, offering the best-case performance under ideal conditions.
Scenario B highlights the need for NW-side mitigation strategies, but its effectiveness is limited due to UE-side mismatches.
Scenario C offers the most comprehensive mitigation, may enable significant performance recovery through coordinated actions between NW and UE.

Proposal 4:
In the results template for capturing the evaluation of temporal domain aspects Case 3/4 of AI/ML based CSI compression, it is clarified that the upper bound is calculated based on ideal CSI prediction and without CSI compression.
Proposal 5:
For temporal domain aspects Case 3 and 4, study the impact on LCM aspects (e.g., data collection, training, monitoring, and model control) of separate prediction and compression vs. joint prediction and compression.
Proposal 6: For inter-vendor collaboration, for direction A, prioritise Option 4-1: Dataset exchanged from the NW-side to UE-side consists of (target CSI, CSI feedback).
Proposal 7: Consider the following to overcome data distribution mismatch: 
Option1: NW should timely update the datasets with samples collected for diverse UE conditions 
Option2: Establish a data set distribution drift detection mechanism at the NW level, to identify when updates are necessary. 
Option 3: Develop a standardized phase normalization method for CSI input, to reduce the dataset distribution mismatch due to UE side additional conditions. 

Proposal 8: For inter-vendor collaboration, for Direction A, consider the NW providing target performance SGCS as additional information to the UE. 
Proposal 9: For inter-vendor collaboration, in direction B, use a common encoder across all UEs and standardize model parameters and pre-processing. 
Proposal 10: For inter-vendor collaboration, consider standardizing a model structure with configurable hyperparameters to adapt effectively to different deployment scenarios.
Proposal 11: For inter-vendor collaboration, for Direction C, consider standardizing both the encoder and decoder (1-3).
Proposal 12: For Standardizing Scalable Model Structure, Consider Training and Specifying Parameters for One or a Few Combinations.
Proposal 13: For a scalable model structure, consider using transfer learning by training a base model on one or more parameter combinations. Add small adaptation layers to the pre-trained model to efficiently handle specific configurations.
Proposal 14: Consider enabling the network to configure the UE to report ground truth CSI via L1 signalling or RRC.
Proposal 15: For UE-side data collection, the network configures CSI-RS resources for the UE. An associated ID can be used to indicate network-side additional conditions.
Proposal 16: For NW side monitoring consider the target CSI reported by the UE via legacy eT2 codebook or eT2-like high-resolution codebook (Case 1) for better monitoring accuracy.
Proposal 17: For UE side monitoring consider the Case 2-1 for better monitoring accuracy
 Based on the output of the CSI reconstruction model at the UE (Case 2-1)
Note: CSI reconstruction model at the UE-side can be the same as the actual CSI reconstruction model used at the NW-side, a reference model provided by NW, or a proxy model developed by the UE side.
Proposal 18: For Option 3a-1 and 4-1, the exchanged dataset or the model parameters can be associated with an ID for pairing related discussion, then 
An ID required for Model performance monitoring

Proposal 19: For Option 3a-1 and 4-1, the exchanged dataset or the model parameters can be associated with an ID for pairing related discussion, then
The same ID used for UE to collect UE-side target CSI for UE-side training, applicability inquiry and reporting and inference configuration can also be used for model monitoring.
Proposal 20: For Model structure scalability, for temporal domain Case 0, we support:
Alt 1: Use subband as the token dimension and Tx port as the feature dimension.
The number of tokens varies with the number of subbands.
Proposal 21: For Model structure scalability, for temporal domain Case 0, for scalability over the feature dimension, we support:
Alt 1: Specific embedding layer for each feature size.

3GPP TSG RAN WG1 #120-bis   	                              R1-2502101
Wuhan, China, April 7th – 11st, 2025
Agenda item:		9.1.4.1
Source:		LG Electronics
Title:			Study on CSI compression
Document for:	Discussion and Decision

Conclusion
In this contribution, we discussed further potential aspects on improving CSI compression/feedback performance and training collaboration with two-sided AI/ML model for CSI compression. The following observation and proposals are provided.

Observation #1: For performance monitoring of CSI compression, signaling overhead mainly originates from ground-truth CSI report, and/or reconstructed CSI delivery. 

Proposal #1: Consider NW-side monitoring approaches via direct estimation of intermediate KPI (e.g., SGCS) without ground-truth CSI reporting. 
Proposal #2: Deprioritize the RS-based monitoring options, e.g., NW-side SRS-based monitoring, UE-side monitoring based on precoded CSI-RS, DMRS.
Proposal #3: Regarding TSF-domain CSI compression, discuss the format of past CSI information and how to report it at least for performance monitoring perspective. 
Proposal #4: First converge on whether to support the joint CSI compression and prediction (JPC) or the separate CSI compression and prediction (SPC), differing the detailed LCM aspects to a subsequent discussion.
Proposal #5: Regarding TSF-domain CSI compression Case 3, consider performance monitoring method on joint CSI compression and prediction by adapting the operation on the AI/ML model between CSI compression and prediction.
Proposal #6: To improve LCM aspects on TSF-domain CSI compression Case 3, consider the following:
Data quality-based sample selection for model updates
Example of the quality metrics: Prediction confidence, temporal consistency, CSI reconstruction error etc.
Proposal #7: On data collection for training at least for Case 3, consider CSI-RS enhancement from the following aspects:
Separate CMR for ground truth CSI: Adjusting power level and RS density
Temporal alignment: Linkage between inference and ground truth CSI measurements
Proposal #8: Regarding non-ideal UCI feedback on TSF-domain CSI compression, 
Consider two-step performance monitoring to check that the performance degradation of the AI/ML model is originated from whether the accumulated past CSI has a problem or the AI/ML model is not suitable for the deployed environment.
Consider to report past CSI information via NW-triggered signaling when UCI missing or UCI dropping.
Proposal #9: Consider the method on the rank adaptation based on the availability check of layer(s) for a given RI.
Proposal #10: For CQI determination in CSI compression using two-sided model, consider to prioritize Option 1. If Option 2 is supported, further consider 
Option 2a: Utilizing AI/ML model complexity reduction method to reduce the signaling overhead to deliver the CSI reconstruction part at NW-side.
Proposal #11: On inter-vendor training collaboration via dataset exchange (i.e., Option 4-1), consider assistant information regarding relationship between target CSI and reconstructed target CSI to reduce/alleviate signaling overhead and model alignment issue.
Proposal #12: Study on model complexity reduction method, e.g., knowledge distillation, to further reduce the CSI training/signaling complexity and issue on model alignment for Type 3 training collaboration.
Proposal #13: Study on methods for reducing the number of tokens, such as omitting or merging tokens, to enable more efficient CSI feedback with scalability across various token dimensions.

3GPP TSG RAN WG1 Meeting #120bis	R1-2502118
Wuhan, China, April 7th – 11th, 2025

Source:	Fujitsu
Title:	Discussion on CSI compression with AI/ML
Agenda item:	9.1.4.1
Document for:	Discussion and Decision

Conclusion
In conclusion, we have the following observations and proposals on CSI compression with AI/ML for Rel-19 further study.
Inter-vendor training collaboration
Observation 1:
For direction C, it could be observed that:
The performance degradation due to data distribution mismatch between synthetic data and field data could be mitigated by retraining/redeveloping the encoder or the decoder with field data. 
Significant performance degradation could be observed if UE-sided encoder and NW-sided decoder are independently retrained/redeveloped without inter-vendor collaboration.

Proposal 1:
For Option 3a-1 in Direction A, RAN1 to consider target CSI exchange from NW side to UE side.
Proposal 2:
Considering that there is no performance advantage without decoder related information disclose and more specification effort is required for option 3a-1, RAN1 to consider option 4-1 with high priority.
Proposal 3:
For Direction A, the validation/test dataset and the corresponding performance target should be shared from NW side to UE side as additional information.
Proposal 4:
For Direction A, the minimum performance requirement could be shared to UE side and the following metrics could be considered:
For the end-to-end performance requirement, NMSE and SGCS could be considered as the performance metric.
For the encoder performance requirement, NMSE and cross entropy could be considered as the performance metric. 
Proposal 5:
For Direction A, to enable UE-sided offline engineering and alleviate inter-vendor training collaboration, the following information could be considered to share from NW side to UE side:
Model input/output related information
Input related information, such as pre-processing (normalization/phase correction/domain-transform/scalability)
Output related information, such as post-processing (domain-transform/scalability)
Quantization related information
Quantization type
Quantization granularity (e.g., vector length, quantization bit)
Quantization table for trainable vector quantization if applied
Proposal 6: 
For Direction C, RAN1 to consider training reference models with synthetic data.
Proposal 7:
For Direction C, test/validation dataset and the corresponding performance target should be shared from NW side to UE side.
Proposal 8:
For Direction C, the minimum performance requirement could be shared to UE side and the following metrics could be considered:
For the end-to-end performance requirement, NMSE and SGCS could be considered as the performance metric.
For the encoder performance requirement, NMSE and cross entropy loss could be considered as the performance metric. 
Proposal 9:
For the choice of token and feature dimensions, considering that three alternatives have the similar performance and Alt.3 is more general than Alt.1/Alt.2, RAN1 to consider Alt.3 with high priority.

Spec impact for two-sided model
Observation 2:
For UE-side AI/ML model performance monitoring using a proxy model, the expectation of a simple structure and a small size contradicts to the needs of a strong generalization capability for a proxy model to work well in various scenarios.
Observation 3:
Using multiple proxy models for UE-side AI/ML model performance monitoring results in additional burden for model management, as well as potential additional overhead because of the assistance information required for choosing a right proxy model among multiple ones.
Observation 4:
The proxy model used in the UE-side AI/ML model performance monitoring is also data-driven. And the performance of the proxy model should also be monitored regularly.
Observation 5:
Only one SGCS is given by a proxy model, which may not be able to represent the performance of multiple CSI reconstruction models from multiple vendors.

Proposal 10:
For CSI compression using two-sided AI/ML models, support both the following alternatives of precoding matrix for output-CSI-UE and input-CSI-NW:
Alt 1: The precoding matrix in spatial-frequency/spatial-frequency-time domain for Case 2/Case 3
Alt 2: The precoding matrix in angular-delay/angular-delay-doppler domain projection for Case 2/Case 3.
Proposal 11:
For CSI compression using two-sided AI/ML models, RAN1 to further study the configurations and CSI reporting formats required for various AI/ML model settings. To reduce the normative workload, the following could be down selected:
AI/ML-model-setting-specific CSI configurations and CSI reporting formats.
A unified configuration and CSI reporting format adapting to various possibilities, including at least
layer specific and rank common.
layer specific and rank specific.
layer common and rank common.
layer common and rank specific.
Proposal 12:
For CSI compression using two-sided AI/ML models, deprioritize Option 2 proposed in RAN1 #112 for CQI determination.
Option 2: CQI is calculated based on the output of CSI reconstruction part from the realistic channel estimation.
Proposal 13:
In the CSI report generated by AI/ML, the information on the order of the spatial layers of the reported precoding matrix should be included, if the reported rank is larger than 1.  
Furthermore, the layer index/order related information should be included in the Part I CSI
Proposal 14:
In the CSI reports generated by AI/ML, the criteria of the Part II CSI priority can be set by the order of the spatial layers of the precoding matrix related information, if the reported rank is larger than 1. 
Furthermore, the order of the spatial layers in the first bullet can be
Alt 1: the same as the order of the reported layers of the precoding matrix related information.
Alt 2: the descending order of the singular values corresponding to the precoding vectors.
Proposal 15:
For the performance monitoring of AI/ML-based CSI compression, RAN1 to prioritize the study of NW-sided monitoring.
Proposal 16:
For the NW-side AI/ML model performance monitoring for CSI compression, RAN1 to prioritize the study of using the codebook-based quantization method to obtain the ground-truth CSI. Besides, adding new parameter values to legacy codebook for higher resolution ground-truth CSI should be studied.
Proposal 17: 
For signaling support to mitigate the impact of UCI loss for Case-2, NW-triggered CSI retransmission should be deprioritized.
Proposal 18: 
For Case 2, RAN1 to study how to address layer ordering issue for training data collection, model inference and performance monitoring.

3GPP TSG RAN WG1 #120bis			R1-2502152
Wuhan, China, April 7th – 11th, 2025

Source: 	CMCC
Title:	Discussion on AI/ML for CSI compression
Agenda item:	9.1.4.1
Document for:	Discussion & Decision

Conclusions
In this contribution, we discussed AI/ML based CSI compression, and the following observations and proposals are made.
Proposal 1: If the target CSI sharing for Option 3a-1 is supported, it should be performed via over the air signaling to alleviate the inter-vendor collaboration complexity.
Proposal 2: Some necessary model related information, such as model backbone, can be aligned between NW side and UE side to achieve better performance for Option 4-1.
Proposal 3: For inter-vendor-collaboration Options 3a-1 and 4-1 in Direction A, for the type of performance metric, take the intermediate KPI, like SGCS, as starting point.
Proposal 4: For performance monitoring, the following two options can be prioritized:
NW-side monitoring based on the ground-truth CSI report.
UE-side monitoring based on the recovery CSI indication.
Proposal 5: In CSI compression using two-sided model use case, regarding the ground truth CSI format for NW side data collection, the basic codebook structure could be reused, along with the basic concept of spatial domain, frequency domain and Doppler domain basis.
Proposal 6: In CSI compression using two-sided model use case, regarding the ground truth CSI format for NW side data collection, the exact supported values of codebook parameters can be studied to make sure high resolution data report.


Observation 1: The required amount of target CSI in Option 3a-1 is ~10% or similar to the amount of target CSI in Option 4-1, depending on the usage of target CSI in Option 3a-1.
Observation 2: In Option 4-1, if some model related information cannot be shared between the UE and the model training side via standardized signaling, inter-vendor training collaboration is still needed.
Observation 3: Based on the results in TR 38.843, the scalable model structure specification and model parameters specification over numbers of Tx ports, CSI feedback payload sizes, and bandwidths, number of slots is feasible for Direction C; the feasibility of the scalable model structure specification can be further improved for Direction A Option 3a-1.

3GPP TSG RAN WG1 #120b	R1-2502196
Wuhan, CN, April 07 – 11, 2025
	
Agenda item:	9.1.4.1
Source: 	Lenovo
Title: 	On AI/ML for CSI compression 
Document for:	Discussion and Decision
	

Conclusions
This contribution addressed AI/ML-based CSI feedback enhancements. We have the following observations and proposals:
Observation 1:	Other than data transfer for model training, transmission of quantized ground-truth CSI might be required for data transfer during the model monitoring or model update phase. The acceptable data transfer overhead/latency for model monitoring/model update phases is usually lower than that of initial model training phase.
Observation 2:	For fixed feedback overhead cases, there exist a trade-off between the number of samples that can be feedback and the resolution of each sample in the feedback data. Therefore, when analysing the gain of transmitting a dataset with higher sample resolutin, it should be compared with the case of not increasing the sample resolution but instead send more samples.
Proposal 1:	For transmission of ground-truth CSI samples, prioritize the performance of transmission of more samples, instead of fewer samples with higher resolution per sample (e.g., more samples with current parameter configurations for Rel-16 Type II, instead of less samples with a new parameter configuration for Rel-16 Type II), especially for cases that the overhead is more important, e.g., ground-truth data transfer for model monitoring or model update.
Observation 3:	Although it is desirable to have larger CSI dataset for training, due to the overhead associated with data collection, it is important to determine the proper size of the CSI dataset which provides enough information for training without imposing large overhead on the network.
Observation 4:	To have a reduced training CSI dataset size, it is desirable to only include CSI samples which are more informative for model training.
Observation 5:	Based on the statistics of the input data and the cost of data transfer, it might be beneficial not to transmit/transfer all CSI samples measured during the data collection phase. This might be due to lower information content of some CSI samples or high correlation between one CSI sample (a group of samples) and another CSI sample (group of samples).
Proposal 2:	Support procedures/signaling enabling transmission of subset of CSI samples among the set of measured/collected CSI samples from the environment.
Observation 6:	Transmission of additional information such as the level distortion when measuring a sample, quality indictor, or how frequent a particular sample (representative sample) occurs, can help the NW side to better train the model.
Observation 7:	As the information content of CSI samples depends on the experienced distortion level, the UE may use distortion level or quality indictor to determine which samples should be transmitted to the NW, or at least can send these information along the sample itself. The NW may be able to use these information further during the training of the model.
Proposal 3:	Support procedures/signaling enabling transmission of subset of CSI samples based on the experienced distortion level or quality indictor.
Proposal 4:	Support procedures/signaling for transmission of additional information such as sample-group size, quality indicator, distortion level along the transmission of the sample itself.
Observation 8:	Knowledge of the conditions/additional conditions under which the samples of the training data have been collected is needed can be used to train Specialized model .
Proposal 5:	Support procedures/signaling enabling UE to report UW-side additional conditions and also the NW conditions/additional conditions under which the data/samples have been collected.
Observation 9:	The model trained using synthetic data experiences performance loss when the inference data is sampled from the field data.
Observation 10:	In direction C, the UE-side and/or the NW-side may use their current field data to fine-tune the respective encoder and decoder models.
Observation 11:	In direction C, if the NW (UE) directly implement the fully specified decoder (encoder) model as its decoder (encoder) model, the fine-tuning of the other side (i.e., encoder (decoder)) can potentially improve the performance of the two-sided model in field data.
Observation 12:	In direction C, if both sides train/fine-tune their encoder/decoder model based on their own dataset (instead of direct implementation of the fully specified encoder/decoder), the final two-sided model could potentially experience performance loss. This is due to the mismatch between the datasets that both sides are using for training/fine-tuning.
Proposal 6:	If a NW (UE) does not intend to implement directly the fully specified decoder (encoder) model, further study how it can train/fine-tune its own version of decoder (encoder) model based on the fully specified decoder (encoder) model without causing mismatch between the encoder/decoder sides.
Proposal 7:	In direction C, further evaluate the feasibility and required procedures including the workload to agree and specify multiple fully specified models.
Proposal 8:	Due to performance limitation and also the required high specification effort, we suggest deprioritizing Direction C for inter-vendor training collaboration.
Observation 13:	In option 4-1 and 3a-1 of Direction A, the performance of the two-sided model (especially if there exists UE data mismatch) is not clear if the UE directly train the encoder model (without the decoder model) based on the exchanged information.
Observation 14:	In option 4-1 and 3a-1 with {target CSI} samples of Direction A, the UE can use the exchanged information to first train a local decoder and then uses the local decoder to train the encoder using the samples available at the UE-side. The trained encoder model can be used for new type UEs which their input data statistics have not been observed during the data collection/training step.
Proposal 9:	For option 4-1 and 3a-1 of Direction A, prioritize schemes based on first construction of the “local” decoder and then training of the encode model.
Proposal 10:	Confirm that the {target CSI} in option 3a-1 should be the {target CSI} from the samples available at the NW-side (used during the training phase).
Proposal 11:	Confirm that the trained “local” decoder model can be considered as a good representation (the proxy) of the actual NW-side decoder model.
Observation 15:	The applicability of a model can be determined based on the conditions/additional conditions under which the samples of the dataset used for training of the model has been collected.
Proposal 12:	Support definition of pairing information based on the conditions/additional conditions assigned to the samples of the datasets used for training of the model.
Observation 16:	Having different model between different UE-NW vendors has some advantages (e.g., higher performance, less cross-vendor efforts during training); however, it is not clear how the UE/gNB can determine the identity (e.g., vendor) of the connected gNB/UE duing the inference time (due to the transparency requirements)
Proposal 13:	Further study model identification/selection procedures during inference time when different models have been developed for different UE-NW vendor pairs.
Proposal 14:	In options 3a-1 (with no {target-CSI} samples), the encoder model cannot be adapted with the statistics of the new UE during the training phase. It is desirable, thus, to study mechanisms to ensure the applicability of the trained model during the inference phase.
Observation 17:	The lower performance of a two-sided model could be due to an issue in the “deployed” encoder and/or decoder model, an issue in the communication link, or data-drift of the input data.
Proposal 15:	Study mechanism to determine the main contributor(s) of the lower performance of the model at least for the issues related to a) the “deployed” encoder and/or decoder model, b) the communication link, or c) data-drift of the input data. We note that, in general, the “deployed” encoder/decoder model could have different performance that the “trained” (reference) encoder or decoder model.
Proposal 16:	Study mechanism for root-cause determination based on exchange of some test data-set between the NW and the UE.
Proposal 17:	Confirm that the local decoder model that is trained using data shared in options 4 and 3a-1 with {target-CSI} has a good match with the NW-side decoder model and it can be use for UE-side monitoring and UE-side root-cause determination.
Proposal 18:	Study mechanism for root-cause determination based on exchange of information regarding the NW-side trained encoder and/or decoder model.
Observation 18:	The performance of AI/ML-based CSI feedback models based on VQ outperforms the model based on SQ.
Observation 19:	The performance of AI/ML-based CSI feedback models using both SQ and VQ to construct the feedback bits outperforms the model based on only VQ and only SQ.
Proposal 19:	Support procedures/signalling enabling CSI-compression models having both Scalar and vector Quantizers for generation of the CSI-feedback bits.

References

3GPP TSG-RAN WG1 Meeting #120bis	  R1-2502213
Wuhan, China, April 7 – 11, 2025

Agenda Item:	9.1.4.1
Source:	Huawei, HiSilicon
Title:	Discussion on AI/ML for CSI compression
Document for:	Discussion and Decision

Conclusions
In this contribution, we discussed issues related with alleviating inter-vendor training collaboration, temporal domain cases, the leftover specification impact for CSI compression, and fusion of SRS measurement and AI/ML compressed CSI for TDD scenarios. Based on the discussions, we have the following observations and proposals.
Observation 1: For the Rel-19 study, there is strong need to explore the necessary extension of CSI compression to achieve high gains in widely deployed TDD scenarios.
Observation 2: For the existing approaches to generate DL precoding:
DL precoder based on SRS measurement may suffer severe channel aging issue, considering long latency for full bandwidth DL CSI acquisition in real network due to sparse SRS resources for a UE.
Configuring dense SRS resources to shorten the latency can alleviate channel aging issue, but it will lead to low efficient UL resource utilization.
DL precoder based on PMI feedback may have limited precision due to coarse frequency domain PMI granularity under large TDD bandwidth.
Introducing smaller frequency granularity of PMI measurement can improve precision of CSI obtained at NW, but with the cost of higher UE complexity, CSI overhead and power consumption.
Observation 3: For TDD scenario, fusion of SRS measurement and AI/ML compressed CSI feedback can be performed to improve the accuracy of the DL precoder.
For Decoder, fusion is performed to the SRS measurement and the CSI feedback (i.e., output of Encoder) before performing multi-head attention.
For Encoder, channel matrix is assumed as the model input to enable the fusion with SRS measurement, while other impacts are similar to the AI/ML CSI compression for FDD.
Observation 4: From the evaluation results over DL CSI acquisition methods, AI/ML PMI compression only is not justified for TDD scenario, as it does not outperform the legacy SRS measurement only.
There is still large margin (48% to the performance with ideal CSI) for further performance improvement for both methods.
Observation 5: From the evaluation results under TDD scenario, DL CSI acquisition based on fusion of SRS measurement and AI/ML compressed CSI feedback can significantly improve performance with 32% gain over DL CSI acquisition methods of both legacy SRS measurement only and AI/ML PMI compression only.
Observation 6: For inter-vendor collaboration Option 4-1, even though the model backbones are different between NW part model and UE part model, model pairing can still be achieved as similar performance is achieved compared with the case where the UE part model has the same backbone to the NW part model.
Observation 7: For the dataset sharing for Option 3a-1/4-1 over air-interface, UEs can upload the received subset of the dataset shared by NW to its own OTT server via offline manner without additionally consuming UL air-interface resources, which is similar to other use cases of UE-sided model.
Observation 8: For the overhead analysis of Direction A, the overhead between Option 3a-1 with Target CSI and Option 4-1 may be comparable.
Observation 9: For the overhead analysis of Direction A/B, the overall consumed air-interface overhead subject to Direction A (Option 3a-1/4-1) is much less than Direction B, since parameter delivery of Direction B needs to be repeatedly performed to each UE for per trained model.
Observation 10: For the proprietary disclosure analysis of Direction A/B, sharing the model information (Option 3a-1/3b) incurs more risk of NW side proprietary disclosure than sharing the dataset information (Option 4-1).
Observation 11: Regarding the complexity reduction methods for AI/ML processing:
It can be achieved with specific model design.
The complexity can be alleviated with additional temporal domain CSI compression (i.e., TSF Case 2), since the Transformer part of the model for achieving SF domain CSI compression may not need to be so sophisticated.
It can be achieved with implementation based approach, e.g., knowledge distillation. 
The student model achieves negligible performance loss compared with the teacher model.
Observation 12: Regarding the complexity reduction analysis for non-AI/ML processing, it can be achieved by adopting appropriate model input type.
Adopting channel matrix as model input has least non-AI/ML complexity as it requires no additional non-AI/ML processing to generate precoder representation.
Adopting spatial-frequency domain precoding matrix has more non-AI/ML complexity as it requires additional non-AI/ML processing to generate eigenvectors based on SVD/EVD decomposition.
Adopting angular-delay domain precoding matrix has heaviest non-AI/ML complexity as it requires additional SD/FD basis selection and matrix projection processing on top of the spatial-frequency domain precoding matrix.
Observation 13: The imbalanced generalization performances between the proxy model (Case 2-1/2-2) at UE and the actual CSI reconstruction part at NW will lead to a degraded monitoring accuracy at the UE side when the channel environment changes.
Observation 14: UE side proxy model (Case 2-1/2-2) is likely to operate under collaboration level x, since its additional LCM will impose huge burden on NW, including model/functionality identification, monitoring, activation/deactivation/switching/fallback, etc., of the UE side proxy model. Without such additional LCM, the performance and robustness of the proxy model are not trustable at NW.
In particular, how to monitor the performance of the UE side proxy model is not clear.

Proposal 1: To enable the fusion of SRS measurement and AI/ML compressed CSI for performance improvement of TDD scenario, support channel matrix as the type of Target CSI for AI/ML CSI compression.
Proposal 2: For the extended temporal domain compression cases, support to keep both Case 2 and Case 3.
Proposal 3: For the performance target of Direction A (Option 3a-1/4-1),
End-to-end performance metric (e.g., Alt.1 of Option 3a-1/4-1), encoder only performance metric (e.g., Alt.2 of Option 3a-1/4-1), and/or decoder only performance metric (e.g., Alt.1 of Option 3a-1/4-1).
Regarding the type of performance metric, both NMSE and SGCS should be adopted.
Regarding the input data for calculating the performance target, the performance target is only valid for using the testing dataset shared from NW.
Proposal 4: For additional information of Direction A Option 4-1, model backbone information is not needed since the UE side can autonomously select appropriate model structure based on the guidance of performance target information indicated by NW side.
Proposal 5: For additional information of Direction A Option 3a-1, the dataset of at least Target CSI should be also shared from NW side to UE side to ensure:
Aligned dataset for training NW part Decoder and UE part Encoder to avoid generalization problem.
Aligned dataset for metric calculation to test the performance target at UE side.
Note: it would be more complex from both spec perspective and UE side implementation perspective to align the dataset construction information if assuming UE side to autonomously collect training and/or testing dataset.
Proposal 6: For the dataset sharing for Option 3a-1/4-1 over air-interface, study the solution to relieve the overhead. E.g.,
NW splits the overall dataset into many subsets each with a limited number of data samples (e.g., with an overhead affordable by the RRC signaling). The subsets can be separately sent to different UEs, and all subsets are associated with a common dataset ID for the UE side re-combination.
Proposal 7: For studying the standardized dataset format of inter-vendor collaboration Option 4-1, consider at least the following aspects:
Data sample format related aspects, 
Type of ground-truth CSI, including precoding matrix and channel matrix.
Dimension of input/output.
CSI feedback related information.
Dataset construction related aspects, e.g., number of data samples, dataset ID, dataset split/segmentation, association between ground-truth CSI and CSI feedback subject to a data sample, observation window/prediction window information for temporal domain cases.
Scalability information.
Target performance information.
Proposal 8: For the NW side data collection, confirm the necessity and feasibility of UE report of the ground-truth CSI.
For the data type, consider the following: 
Precoding matrix
Channel matrix
For the data format, prioritize Rel-16 eType II CB based quantization with new parameters for both above data types, and take the following new parameters as candidates for discussion:
L= 8, 10, 12; pv = 0.8, 0.9, 0.95; reference amplitude = 6 bits, 8 bits; differential amplitude = 4bits; phase = 5 bits, 6 bits.
Proposal 9: For NW-side monitoring, support ground-truth CSI based monitoring.
eT2-like high-resolution codebook for reporting format with higher priority.
Further discuss the reporting mode, e.g., per sample reporting and reporting of a number of monitored samples.
Note: potential solution can be considered to alleviate UE complexity on generating the eT2-like high-resolution codebook, considering the overhead/latency of monitoring CSI report can be relaxed compared with legacy eT2 CSI.
Proposal 10: There is no strong motivation for specifying the UE side proxy model (Case 2-1/2-2) for monitoring.

3GPP TSG-RAN WG1 Meeting #120bis                                               	 R1-2502291
Wuhan, China, April 7th-11th, 2025
Source:	OPPO
Title:	Additional study on AI/ML-based CSI compression
Agenda Item:	9.1.4.1
Document for:	Discussion and Decision

Conclusion
In this contribution, we provide our views on two-sided AI/ML based CSI compression. Observations and proposals are summarized as follows:
Observation 1: When UE-side trains its actual encoder and NW-side trains its actual decoder, the compatibility of Direction C is questionable.
Observation 2: Regarding reference model in Direction C,
1-1: specify encoder only reduces the complexity of UE-side model training and deployment, and provides the degree of freedom for NW-side model implementation.
1-2: specify decoder only reduces the complexity of NW-side model training and deployment, and provides the degree of freedom for UE-side model implementation.
1-3: specify both encoder and decoder can address the inter-vendor collaboration issue with huge standardization efforts.
Observation 3: Regarding the real deployment for Direction C, sequential UE-side and/or NW-side separate training based on the fully specified reference model is necessary. 
Observation 4: One model structure with multiple sets of parameters cannot fully solve the scalability issue, which still requires additional inter-vendor collaborations when toke/feature dimensions and/or payload changes. 
Observation 5: Regarding Option 3a-1 and Option 3b, if the reference model structure is standardized, the feasibility, inter-vendor collaboration complexity, performance and interoperability/RAN 4 testing are summarized as:

Observation 6: Regarding Option 4-1, if the data / dataset format is standardized, the feasibility, inter-vendor collaboration complexity, performance and interoperability/RAN 4 testing are summarized as:
Observation 7: The performance of option 4-1 can be ensured with aligned backbone between NW-side decoder and UE-side encoder. 
Observation 8: Target CSI reporting overhead is not a critical concern when high layer signaling (e.g., MDT) is used for NW-side data collection.
Observation 9: For the performance of UE-side monitoring Case 2-1, the percentage of increases when more monitoring samples are averaged.
Observation 10: Using the same backbone between UE-side proxy decoder and NW-side actual decoder slightly outperforms UE-side proxy decoder with different backbone from NW-side actual decoder.
Observation 11: UE-side monitoring Case 2-1 with proxy decoder can provide sufficient generalization between UMa and UMi scenarios.
Observation 12: For UE-side monitoring Case 2-1, UE-side proxy decoder with the same or different backbone to NW-side actual decoder can achieve the performance with ,  and  for UMa and UMi scenarios.
Observation 13: Regarding NW-side monitoring Case 1-1, two concerns from overhead and UE capability aspects should be addressed.
Observation 14: Regarding UE-side monitoring Case 2-1, there is no overhead and UE capability concerns.
Proposal 1: Deprioritize Direction C (Option 1).
Proposal 2: One model structure with only one set of parameters should be the baseline to handle the scalability issue.
Proposal 3: Regarding the performance target exchange in Option 3a-1, both of SGCS and NMSE can be supported.
Proposal 4: Target CSI sharing from NW-side to UE-side is optional for option 3a-1 is UE/UE-side is capable of UE-side data collection.
Proposal 5: Regarding [issue 3] for Direction A and Direction B, overhead concern is more critical for over-the-air exchange compared to offline exchange, mainly from two aspects: 
Parameters exchange: can be alleviated by proper model compression methods
E.g., knowledge distillation, parameter quantization, pruning
Dataset exchange: can be alleviated by target CSI label quantization methods
E.g., float16, eType II-like codebook
Proposal 6: Regarding [issue 5] for Direction B, the feasibility should be considered from following two aspects:
Performance aspect: feasible to train multiple encoders for different UEs
UE capability aspects: feasible if supporting standardizing multiple encoder structures 
Proposal 7: Support to exchange the NW-side decoder backbone information as the additional information in option 4-1.
Proposal 8: Regarding [Issue 3] for Direction A (Option 4-1), overhead can be alleviated by target CSI label quantization methods, e.g., float16, eType II-like codebook.
Proposal 9: Regarding different options to resolve the inter-vendor collaborations and processing into normative work
Option 4-1 without model structure standardization should be treated in first priority.
Option 3a-1, Direction B and Direction C with model structure (and parameters) standardization should be treated in second priority
Proposal 10: Regarding the signaling of ground-truth reporting for NW-side data collection for training, high layer signaling (e.g., MDT) is preferred than L1 signaling.
Proposal 11: Regarding the NW-side data collection data format, float32 should be considered if high layer signaling is used for target CSI label reporting.
Proposal 12: Regarding the codebook-based data format in NW-side data collection, suggest to use Rel-18 eType2 or enhanced Rel-18 eType2 codebook for Case 2 and Case 3 with temporal domain correlations between target CSI labels.
Proposal 13: Regarding the CSI-RS configuration for UE-side and NW-side data collection, at least the follow aspects should be considered:
Data collection specific CSI-RS or not
Cell-specific or UE specific CSI-RS
Proposal 14: UE-side proxy decoder can be trained with the frozen UE-side actual encoder based on the dataset shared from NW-side.
Proposal 15: Regarding the performance of UE-side monitoring Case 2-1 and Case 2-2, use the following monitoring metric and candidate  value:
False alarm rate: 
Mis-detection rate: 
where  is suggested.
Proposal 16: Support to study and specify UE-side monitoring Case 2-1 and Case 2-2.
Proposal 17: Regarding rank > 1 option, support option 3-1 (layer-common and rank-common) and option 2-1 (layer-specific and rank-common):
Option 3-1 can be supported as the baseline 
Option 2-1 can be supported as a more flexible solution
E.g., Different CSI feedback payloads and temporal domain cases between layers

3GPP TSG-RAN WG1 #120bis	R1-2502338
Wuhan, China, April 7th – 11th, 2025
Agenda Item:	9.1.4.1
Source:	InterDigital, Inc.
Title:	On AI/ML-based CSI compression and other aspects
Document for:	Discussion and Decision

Conclusion
In this contribution, we discussed aspects of CSI compression performance, complexity, pre-processing, as well as two-sided model monitoring. We also provided simulation results for beam domain CSI compression and TSF-based CSI compression.
We provide the following observations and proposals.
Observation 1: Beam domain processing results in low-dimensional channel samples that may be compressed using low complexity Autoencoder models, resulting in reduction in both memory and computational requirements relative to spatial domain Autoencoders.
Observation 2: Beam domain processing exhibits good generalization capabilities as a single beam domain AE model can deal with multiple antenna configurations, potentially reducing the need for model switching.
Observation 3: While TSF compression may potentially result in further improvement in the compression performance relative to SF compression, it may require more complex AE models with higher computational and memory requirements relative to SF. 
Observation 4: Specification impacts of mitigation of missing UCI for CSI Compression Case 2 include signaling and timing information to maintain synchronization of the historical CSI information between the UE-side and NW-side.
Observation 5:	For UE-side monitoring, both time-based trigger and event-based trigger needs to be studied for reporting the monitoring metrics. 
Observation 6:	For UE-side monitoring, appropriate monitoring metrics need to be identified (e.g., other than SGCS) to avoid unnecessary model updating or switching considering that SGCS couldn’t reflect AI/ML model performance in a reasonable manner. 
Observation 7:	Out-of-distribution metrics for UE-side monitoring can be another candidate as monitoring metric but further study is necessary. 
Observation 8: 	For UE-side monitoring, monitoring metrics in terms of report size, metrics quantization, and report frequency also needs to be studied to investigate feedback overhead. 
Observation 9: 	UE-side monitoring based on the output of the NW-side CSI reconstruction model may increase the downlink signaling overhead, because the output CSI reconstructed at the NW needs to be indicated by the NW to the UE.
Observation 10: Precoded RS (CSI-RS, DM-RS) can be used for UE-side model monitoring, by comparing the expected precoding gain to actual precoding gain and determining the mismatch between actual CSI at UE and reconstructed CSI at NW.
Observation 11:  NW-side monitoring with lower signaling overhead options can be studied further in Rel-19 as alternative to the UE-side monitoring. 
Observation 12: It may be beneficial to use AI/ML Processing Units (APU) to indicate the number of supported simultaneous AI/ML functionalities.
Observation 13: AE with beam domain processing outperforms the AE with spatial domain processing in terms of intermediate KPI.
Observation 14: AE with beam domain processing has lower complexity and memory requirements compared to the AE with spatial domain processing.
Observation 15: For the scenarios investigated using 2 TD samples, TSF-based CSI compression can reduce the CSI feedback overhead by up to 33% compared to SF-based CSI compression, with a smaller reconstruction error.
Observation 16: For the scenarios investigated for multi-frame TSF architectures, increasing the number of frames for TSF compression improves the performance, as the same reconstruction error is achieved with a lower average number of CSI feedback bits. 
Observation 17: The gain of the multi-frame TSF performance appears to saturate as the number of TD CSI samples increases above a threshold. For example, for the scenarios investigated, the performance starts saturating for 4 frame TSF architectures.
Observation 18: Under FTP traffic with low, medium and high RU, TSF based compression achieves minor mean throughput gain over SF-based compression, and considerable mean throughput gain over Rel-16 Type II, at the same overhead of 80 bits. Specifically, for the 2 CSI frame scenario investigated, TSF achieves the following throughput gains,
1-2% mean throughput gain over SF compression under low, medium and high RU traffic
12-16% cell-edge throughput gain over Rel-16 Type II under low, medium and high RU traffic

Observation 19: Under FTP traffic with low, medium and high RU, TSF based compression outperforms SF compression and Rel-16 Type II in terms of cell-edge throughput, at the same overhead of 80 bits. Specifically, for the 2 CSI frame scenario investigated, TSF achieves the following cell-edge throughput gains,
12-17% cell-edge throughput gain over SF compression under low, medium and high RU traffic
25-45% cell-edge throughput gain over Rel-16 Type II under low, medium and high RU traffic

Observation 20: For low to moderate UE speeds, the overhead for reporting the selected beams (BDP overhead) is low, even under complex mobility models. 
Observation 21: CSI compression Case 2 with beam domain processing achieves better performance, e.g., overhead reduction, relative to Case 0 with spatial domain processing and beam domain processing.
Observation 22: TSF Case 3 shows only a small SGCS improvement over the AI/ML based CSI prediction with Rel-18 CB, and a small SGCS degradation as compared to the non-AI/ML prediction (Kalman) with Rel-18 CB.
Observation 23:  For Option 3a-1, it is feasible to perform UE-side off-line engineering to develop a UE-side encoder (including a lower complexity encoder) that is compatible with a proprietary NW-side actual decoder. There is a small performance impact (up to 0.6 dB NMSE degradation, and 3.3% SGCS degradation) relative to the reference encoder-actual decoder pair developed at the NW-side.
Observation 24:  For Option 3a-1, the additional information that needs to be exchanged from the NW-side to the UE-side to enable UE-side off-line engineering of an actual encoder includes: information on the exchanged dataset A (containing the target CSI), model backbone information (e.g., transformers, CNN, etc.), as well as minimum KPI (SGCS, NMSE). 
Observation 25:  For Issue 4, for the considered configuration, there is a small performance degradation due to UE-side / NW-side data distribution mismatch with respect to UE side additional conditions, specifically of up to 2.95% SGCS degradation and up to 0.54 dB NMSE degradation. 
Observation 26: SGCS performance may be improved while maintaining complexity, by increasing the number of transformer blocks while decreasing the model dimension (tf_dim). Increasing the number of transformer blocks beyond a threshold (in our example, increasing from 2 to 4 transformer blocks) shows diminishing returns. 
Observation 27: For the scenarios analysed, it can be seen that increasing the overall complexity above a threshold (e.g., above 5 million FLOPs) does not provide significant SGCS performance improvement.

Proposal 1: Study the specification impact of adapting the length of the historical CSI buffer to channel conditions, such as UE speed. 
Proposal 2:	Study further the following aspects for model monitoring in Rel-19:
Details of reporting mechanism for the monitoring metrics with both time/event-trigger based
Appropriate UE-side monitoring metric which reflects AI/ML model performance accurately
UE-side monitoring based on precoded RS (CSI-RS, DM-RS)
Reporting contents/structure of UE-side monitoring metric and its associated feedback overhead
NW-side monitoring with lower signaling overhead

Proposal 3: Support APU pool sharing by different AI/ML functionalities (e.g., AI/ML based CSI reporting, AI/ML based BM, AI/ML based POS).
Proposal 4: Support AI/ML Processing Units (APU) occupancy definition per AI/ML functionality (e.g., AI/ML based CSI reporting, AI/ML based BM, AI/ML based POS).
Proposal 5: For active models, the inference timeline needs to account for the AI/ML model readiness status. 
Proposal 6: TSF compression performance should be evaluated under multiple observation window lengths.
Proposal 7: For CSI compression Case 2 with beam domain processing, study long-term reporting of the selected beams associated with the BDP, e.g., as a function of the UE speed.

3GPP TSG RAN1 WG #120bis		                         R1-2502355
Wuhan, China, April. 7th– 11th, 2025

Agenda item:	9.1.4.1
Source:	Samsung
Title:	Views on additional study for AI/ML based CSI compression
Document for:	Discussion and Decision

Conclusion
In this contribution, the following observations and proposals are made:
Observation#1: Case 0 of AI/ML-based CSI compression using two-sided model does not consider the time-domain aspects in the CSI compression. Thus, it corresponds to the spatial-frequency domain compression studied in Rel-18. 

Observation#2: For the Case 2 of temporal aspects of AI/ML-based CSI compression using two-sided model, at least the following two options can be considered for the past CSI 
Case 2-1: Past CSI generated by AI/ML model at the UE and/or network 
Case 2-2: SD/FD basis vectors from preprocessing as past CSI and AI/ML-based CSI compression in angle-delay domain.  

Proposal#1: For the Case 2 of temporal aspects of AI/ML-based CSI compression using two-sided model, consider at least the following two options for the past CSI information
Case 2-1: Past CSI information generated by the UE-part and/or network-part of two-sided model
Case 2-2: Information on SD/FD basis vectors as past CSI information with angle-delay domain compression.  
 

Proposal#2: For the Case 2 of temporal aspects of AI/ML-based CSI compression using two-sided model: 
when past CSI information corresponds to SD/FD basis and AI/ML CSI compression is in the angle-delay domain, consider SD/FD basis reporting per N CSI reporting occasions, i.e., N times longer periodicity.
FFS on the values of N. 

Observation#3 For the Case 2 of temporal aspects of AI/ML-based CSI compression using two-sided model, at the low payload size (Rel-16 eType II PC1), reporting the SD/FD (W1/Wf) basis with longer reporting period than the angle-delay (W2) coefficients, e.g., N=2 and N=10 times longer period, reduces the CSI overhead, e.g., by 11.8% and 25.9%, while maintaining similar levels of SGCS gain, e.g., less than 1.8% for layer 1 and 4.8% for layer 4 SGCS degradation

Observation#4: For the Case 2 of temporal aspects of AI/ML-based CSI compression using two-sided model, at low payload size (Rel-16 eType II PC1), reporting the SD/FD (W1/Wf) basis with longer reporting period than the angle-delay (W2) coefficients, e.g., N=2 and N=5 times longer period,  improves the CSI accuracy, e.g., SGCS gain up to 30% for layer 3 and up to 10% for layer 4. 

Observation#5: For the Case 3 of AI/ML-based CSI compression using two-sided model with time-domain aspects. The reported CSI may correspond to  prediction instances (Doppler time intervals) 

Proposal#4: Among the identified six categories for AI/ML-based CSI compression using two-sided model, for Case 3, consider  prediction instances (Doppler time intervals). 
Option1: AI/ML-based CSI compression in spatial-frequency-time domain 
Option 2: The AI/ML-based CSI compression in angle-delay-time domain 
Option 3: The AI/ML-based CSI compression in angle-delay-Doppler domain 
Option1: AI/ML-based CSI compression in spatial-frequency-time domain 
Option 2: The AI/ML-based CSI compression in angle-delay-time domain 
Option 3: The AI/ML-based CSI compression in angle-delay-Doppler domain 

Observation#6: For the Case 2 of temporal aspects of AI/ML-based CSI compression using two-sided model, the CSI reconstruction error propagation can be resulted from the following
Imperfect past CSI generation ( representation)
Part 1 and/or Part II CSI dropping (depending on priority)
UCI transmission loss

Proposal#5: For the Case 2 of temporal aspects of AI/ML-based CSI compression using two-sided model, for the evaluation of performance impact resulting from CSI dropping 
Modelling for CSI dropping probabilities
x% Group 2 of Part II CSI dropping probability 
y% Group 1 of Part II CSI dropping probability 
z% Group 0 of Part II CSI dropping probability 
where  .
Note: Companies to report the partitioning of part I and part II CSI and the values for x, y, and z. 

Observation#7: In Angle-delay (W2)-domain AI/ML CSI compression, for small payload size (paramCombination-r16  1), ‘Heavy’ and ‘light’ models show comparable SGCS gain over Rel-16 eType II CSI.  

Observation#8: In Angle-delay (W2)-domain AI/ML CSI compression, for small payload size (paramCombination-r16 =1), increasing the number of basis vectors (increasing the dimension of the W2 matrix), e.g., paramCombination-r16=3, improves performance. 

Proposal#7: In Angle-delay (W2)-domain CSI compression, study the impact of the number of SD/FD basis vectors for performance-complexity tradeoff. 

Observation#9: For evaluation of the generalization performance of AI/ML-based CSI compression using two-sided model 
For precoding vectors in the spatial-frequency (W) domain as model input
For Case 2, when a model is trained based on dataset from Scenario#A and tested on dataset from Scenario#B, up to 37.9% performance degradation is observed as compared to Case 1 wherein the model is tested on dataset from the same Scenario#A
For Case 3, when a model is trained on training dataset constructed by mixing datasets from Scenario#A and Scenario#B and tested on dataset from Scenario#A or Scenario#B, up to 11.2% performance degradation is observed as compared to Case 1 wherein the model trained on dataset from Scenario#A and is tested on dataset from the same Scenario#A
For precoding vectors in the angle-delay (W2) domain as model input
For Case 2, when a model is trained based on dataset from Scenario#A and tested on dataset from Scenario#B, up to 9.0% performance degradation is observed as compared to Case 1 wherein the model is tested on dataset from the same Scenario#A
For Case 3, when a model is trained on training dataset constructed by mixing datasets from Scenario#A and Scenario#B and tested on dataset from Scenario#A or Scenario#B, up to 0.7% minor performance degradation is observed as compared to Case 1 wherein the model trained on dataset from Scenario#A and is tested on dataset from the same Scenario#A
Deployment scenarios such as InH, UMa 100% outdoor UEs and UMa with indoor:outdoor ration 80%:20% are considered for the Scenario#A and Scenario#B. 

Proposal#8: For the cases of CSI compression with temporal aspects, consider the following for the network’s ground-truth CSI collection
For cases that require multiple time-domain samples for inference, cases 2/3/4/5, high resolution codebook quantization including temporal aspects, e.g., Rel-18 eType II-like method with new parameters. 
For cases with CSI prediction, e.g., cases 3/4, high resolution codebook quantization for explicit channel matrices, e.g., codebook to report the left and right eigenvectors of a channel matrix 
Specification impact on measurement and reporting for ground-truth CSI 

Proposal#9: For NW’s and UE’s Data collection, consider indication on the network-side additional condition.  
For two-sided models development, NW-part of two-sided model associated with a dataset can be considered as NW-side additional condition. 

Proposal#10: For performance monitoring, consider causes of performance loss that particularly affect the AI/ML based approach. 
Consider KPI for monitoring such as , where  and   are the SGCS for AI/ML-based CSI and baseline CSI, e.g., eType-II, respectively. 
To evaluate the monitoring options, apply the evaluation mechanism agreed in Rel-18 by replacing intermediate KPI by . 
FFS: baseline CSI for a payload size. 

Proposal#11: For NW-side monitoring of two-sided models, when the input CSI is the (W2) domain, consider the same SD, FD and/or DD basis vectors for the CSI report for inference and monitoring purposes. 

Proposal#12: For Direction C, wherein at least NW-part (CSI reconstruction model ) of two-sided model for CSI compression is standardized, the UE may calculate/determine the performance monitoring output/KPI based on the output of the fully standardized reference CSI reconstruction model at the UE

Observation#11: For Case-2A of Direction C, when UE and NW independently train/update the UE-part and NW-part of their actual models to be compatible with fully standardized reference two-sided models with training dataset (Dataset#B) collected from the same field, i.e., UE-side training dataset and NW-side training dataset are from the same distribution (deployment scenario):
When the model input is precoding vectors in the spatial-frequency (W) domain, 11.18% SGCS degradation is observed as compared to baseline models trained on Dataset#B.
When the model input is precoding vectors in the angle-frequency (W2) domain, a minor 1.51% SGCS gain is observed as compared to baseline model trained on Dataset#B.
Note: The training datasets Dataset#S for reference model and Dataset#B for proprietary models are generated with different deployment scenarios (UMa 80% indoor and InH, respectively). 

Observation#12: For Case-2A of Direction C, when the UE side and NW side independently train/update the UE-part and NW-part of their actual models to be compatible with fully standardized reference two-sided model with training dataset (Dataset#B-UE and Dataset#B-NW) collected from the different fields i.e., UE-side training dataset and NW-side training dataset are from different distributions (deployment scenarios):
When the model input is precoding vectors in the spatial-frequency (W) domain, up to 30.5% SGCS degradation is observed as compared to baseline models trained on Dataset#B-UE or Dataset#B-NW.
When the model input is precoding vectors in the angle-frequency (W2) domain, a minor 2.9% SGCS degradation is observed as compared to baseline models trained on Dataset#B-UE or Dataset#B-NW.
Note: The training datasets Dataset#S for reference model and Dataset#B-UE and Dataset#B-NW for actual models are generated with different deployment scenarios. 


Observation#13: It is feasible to separately train/fine-tune UE-part and NW-part of actual models to be compatible with fully standardized reference two-sided model with UE-side and network-side training/finetuning datasets, respectively, at least for the case wherein the model input is precoding vectors in the angular-delay (W2) domain. 
Observation#14:  For Case-2A of Direction C, when the UE side and NW side independently train/update the UE-part and NW-part of their actual models to be compatible with fully standardized reference two-sided model with training dataset (Dataset#B-UE and Dataset#B-NW) collected from the different fields i.e., training dataset at UE and NW are from different distributions (deployment scenarios), when Dataset#B-UE and Dataset#B-NW include the dataset for reference model training 
When the model input is precoding vectors in the spatial-frequency (W) domain, up to 3.6% SGCS degradation is observed as compared to baseline models trained on UE/NW-side training datasets.
Note: The training datasets Dataset#S for reference model and Dataset#B-UE and Dataset#B-NW for actual models are generated with different deployment scenarios (UMa 80% indoor, mix of UMa 80% indoor  InH,  mix of UMa 80% indoor  and UMa 100% outdoor, respectively). 


Proposal#13: For Issue 9 of Direction C, RAN1 to conclude that data distribution mismatch among datasets for reference model training (Dataset#C), UE-side training dataset (Dataset#B-UE) and NW-side training dataset (Dataset#B-NW)
causes minor performance degradation for model input in spatial-frequency (W) domain, when Dataset#B-UE and Dataset#B-NW are from the same field (deployment scenario). 
causes considerable performance degradation for model input in spatial-frequency (W) domain, when Dataset#B-UE and Dataset#B-NW are not from the same field (deployment scenario). 
Has negligible impact for model input in angular-delay (W2) domain.  
FFS: method to align UE-side training dataset (Dataset#B-UE) and NW-side training dataset (Dataset#B-NW) in terms of deployment scenario. 

Proposal#14: For Direction C, consider associated ID to align the distribution of UE-side training dataset (Dataset#B-UE) and NW-side training dataset (Dataset#B-NW) with respect to NW-side additional conditions, e.g., deployment scenario and/or NW-part of two-sided model. 
The associated ID is indicated with configuration for data collection (target CSI) for training and inference (pairing). 

Observation#15: For Case-2A of Direction C, when the UE side and NW side independently train/update the UE-part and NW-part of their proprietary models to be compatible with fully standardized reference two-sided model with training respective datasets, Dataset#B-UE and Dataset#B-NW with different UE antenna spacing assumptions:
When the model input is precoding vectors in the spatial-frequency (W) domain, a minor 0.21% SGCS performance gain is observed as compared to baseline model trained on Dataset#B-UE. 
When the model input is precoding vectors in the angle-frequency (W2) domain, a minor 0.89% SGCS degradation is observed as compared to baseline models trained on Dataset#B-UE.
Note: The training datasets Dataset#S for reference model and Dataset#B-UE and Dataset#B-NW for proprietary models are generated assuming UE side antenna spacing set as  (0.5, 0.5) λ , (0.5, 0.8) λ and (0.5, 0.5) λ, respectively.

Proposal#15: For Issue 10 of Direction C, RAN1 to conclude data distribution mismatch as a result of UE antenna spacing does not impact performance. 
UE side may train/update its UE-part of two-sided model independently

Proposal#16: For Direction A, consider associated ID to align the distribution of UE-side training dataset (Dataset#A-UE) and NW-side training dataset (Dataset#B-NW) with respect to NW-side additional conditions, e.g., deployment scenario and/or NW-part of two-sided model. 
For Option 4-1, the associated ID is indicated with configuration for data collection for training for UE’s measurement/reporting of target CSI and/or dataset transfer {target CSI, CSI feedback} and/or and inference (pairing). 
For Option 3a-1, the associated ID is indicated with configuration for data collection for training (target CSI measurement/reporting and/or reference model transfer and/or and inference (pairing). 
Note: The network’s indication of the associated ID for training data collection, reference model/dataset transfer, and inference ensures consistency on UE’s assumption on NW-side additional condition, e.g., NW-part of two-sided model, across training and inference. 

Proposal#17: For Issue 6 of Direction A, RAN1 to conclude that data distribution mismatch that may result from UE antenna spacing does not impact performance. 
UE side may train/update its UE-part of two-sided model independently.  

Observation#16: For Option 4-1 of Direction A, performance degradation is observed when the NW-side encoder backbone associated with the dataset is different from the UE-side encoder backbone. 

Proposal#18: For Option 4-1 of Direction A, consider NW-side sharing the encoder backbone assumption associated with the dataset as additional information. 

Observation#17: For evaluation of the scalability of specified model structure,  based on 
Subband as the token dimension and Tx port as a feature dimension
A common embedding layer with padding (e.g., zero-padding or other techniques for padding values) for model scalability
Negligible performance loss ( up to -1.91%) is performed by Case 1 (scalable structure) as compared to Case 2 (dedicated structure). 


Observation#18: For the transformer model in the spatial-frequency-domain compression, increase on both the model dimension and the number of attention blocks tends to have a monotonic but diminishing performance gain at the expense of increased model size and complexity. 

Observation#19: For the transformer model in the spatial-frequency-domain compression, for a given model dimension, increasing the number of attention blocks does not always guarantee performance gain.  

Proposal#19: For the transformer model in the spatial-frequency-domain compression, consider balanced model hyper-parameters selection based on the tradeoff between model performance and complexity/size. 

Observation#20: The UE may execute the configurations for AI/ML related features/functionalities partly or fully on a different hardware unit (processing resource) than legacy CSI.

Observation#21: The UE may execute the configurations for AI/ML related features/functionalities partly or fully on a different hardware unit (processing resource) than legacy CSI.

Proposal#20: For processing criteria and timeline of AI/ML feature related configurations, consider a separate AI/ML processing resource pool(s) than the legacy CPU, UE reports at least one pool of the number of simultaneously executed AI/ML processes it supports as 
FFS: how to address memory constraints, e.g., the number of simultaneously activated models/functionalities, buffering requirement for channel samples. 
FFS: whether the PU for AI/ML related CSI report are counted towards    only. 

Observation#21: It is beneficial if UE reports the supported AI/ML features/functionalities by mapping them (grouping them) to AI/ML [logical] models identified from UE side.  For example, AI/ML model activation may require longer delay than inference from activated AIML model. Thus, if the network has awareness on the UE’s mapping between the functionalities/configuration and AI/ML models, the network may provide model-level LCM assistance as per UE’s reporting. 

Proposal#22: Study mechanism for the UE to report the supported AI/ML functionalities by mapping them to its models identified to the network.   

3GPP TSG RAN WG1 #120bis                                                                                             R1-2502433
Wuhan, China, April 7th – 11th, 2025

Agenda item:	9.1.4.1
Source:	            		Xiaomi
Title:	Discussion on AI/ML model based CSI compression and other aspects on AI/ML model/data
Document for:	Discussion and Decision

Conclusions
In this contribution, the proposals and observations on two-side AI model based CSI compression and processing unit are summarised as follows:
Observation on model scalability
Observation 1: The scalable model over the feature dimension for Alt 1 and scalability over payload configurations for Alt1 obtain -2.05%~8.11% SGCS performance gain over legacy eType II codebook.
Observation 2: The scalable model over the feature dimension for Alt 1 and scalability over payload configurations for Alt1 obtain -0.88%~1.97% SGCS performance gain over Case 2.
Observation 3: Compared with eType II and Case 2, average gain for different feature dimension and payload configurations is 3.00% and -2.16 respectively.
Observation 4: The scalable model over the feature dimension for Alt 1 and scalability over payload configurations for Alt1 obtain -0.41%~4.82% SGC Sperformance gain over legacy eType II codebook.
Observation 5: The scalable model over the feature dimension for Alt 1 and scalability over payload configurations for Alt1 obtain -1.64%~6.02% SGCS performance gain over Case 2.
Observation 6: Compared with eType II and Case 2, average gain for different feature dimension and payload configurations is 2.04% and -3.13% respectively.

AI/ML-based CSI compression
Observation 7: NW-side monitoring could reduce UE-side computation complexity and need less specification efforts.
Observation 8: Computation complexity, model storage or downlink transmission resource overhead of UE-side monitoring is larger than NW-side monitoring, and more specification efforts are required to specify the performance metric and mechanism of monitoring results. 

Proposal 1: For Option 3a-1, sharing target CSI from NW side to UE side for training encoder is not necessary. 
Proposal 2: Backbone information of encoder should be shared to UE side.
Proposal 3: NW-side monitoring model performance based on target CSI reported by UE should be as a starting point of performance monitoring. 
Proposal 4: UE-side monitoring model performance based on precoded RS can be considered, and the performance metric based on legacy eType II codebook should be as a reference to evaluate the monitoring accuracy. 
Proposal 5: For training data collection for temporal domain aspect Case 3, Option 2, i.e., the target CSI is derived based on the measured CSI of the future slot(s) could be considered to train CSI compression model.
Proposal 6: For monitoring labels for temporal domain aspect Case 3, measured CSI of the future slot(s) is used as input to CSI generation part for monitoring purpose, could be considered to monitoring two-side AI/ML model.

AI processing unit
Proposal 7: Define a separate dedicated computation resource pool for AI-based use cases  
Proposal 8: Define AI CPU to quantize the AI processing amount and processing capability 
Proposal 9: 
The Functionality for the AI processor should be aligned 
Consider the AI processer on UE side could process legacy measurement processing and AI model inference
Proposal 10: The Occupation time for each use case should be defined 
It is assumed that Y time units would be consumed for one AI-based use case, the Y would be determined per use case
Proposal 11: Define the consumed AI CPUs for each AI use case in the specification by taking into the processing complexity of input measurement and AI model complexity
Proposal 12: 
Consider the impact of activation delay when determine the timeline of inference result-based report
Proposal 13: For the AI/ML model activation delay, further study the following two options 
Option 1: Fixed in the specification 
Option 2: Depends on UE capability report 
Proposal 14: UE should report the supported AI processing capability to facilitate the AI use case configuration on network side  

3GPP TSG RAN WG1 #120bis			R1-2502489
Wuhan, China, April 7th – 11th, 2025
Agenda item:		9.1.4.1
Source:		Mavenir 
Title:	AI/ML for CSI feedback enhancement
Document for:		Discussion and Decision
1  

Conclusions
In this contribution, we discuss the evaluations on AI/ML for CSI feedback enhancement and provide preliminary simulation results. We have the following observations and proposals.
Observation 1: For different scenarios, the SGCS degradation is about 1%~3% when training set and testing set are mismatching.
Observation 2: Training on mixing dataset with UMa and UMi can improve the generalization performance (about 1.4%/0.6%) of AI/ML model for UMa and UMi.
Proposal1: Mixed dataset should be considered as an effective way to increase AI model generation.
Proposal2: AI model performance should be evaluated on datasets with different scenario and configuration.
Proposal3: Prioritize the study of the specification impact of NW-side monitoring based on the target CSI reported by the UE.
Proposal4: It is recommended to define the target CSI report based on the enhanced eType-II format to ensure high-precision model performance monitoring at the NW end.
Proposal5: It is recommended to specify the signals for periodic data collection at the NW end and configure them for model performance monitoring.
Proposal6: Further study the data content and format for performance monitoring based on NW-side data collection.
4. 

3GPP TSG RAN WG1 #120bis		R1-2502493
Wuhan, China, April 7th – 11th, 2025

Source:	Panasonic
Title: 	Discussion on AI/ML for CSI compression
Agenda Item:		9.1.4.1
Document for:	Discussion

Conclusion
In this contribution, we provide our view on the AI/ML-based CSI compression. We made following proposals and observations.
Section 2: Inter-vendor collaboration
Proposal 1: {Target CSI} in Option 3a-1 should be target CSI exchanged from the NW-side.
Observation 1: Option 4-1 may be “feasible, but less reliable” without the assumption that the model structure is aligned based on offline inter-vendor collaboration.
Observation 2: Direction A Option 3a-1 can achieve almost the same performance as joint training if data distribution between NW-side offline training and UE-side offline training is aligned.
Observation 3: For Direction A Option 3a-1, in terms of UE-side additional condition of antenna layout/configuration or SVD phase normalization, the performance loss is small especially if Dataset A for NW-side decoder and nominal encoder training includes Dataset B.
Observation 4: For Direction A Option 4-1, in terms of UE-side additional condition of antenna layout/configuration the performance loss is small if Dataset A for NW-side decoder and nominal encoder training includes Dataset B.
Observation 5: For Direction A, Option 3a-1 and Option 4-1 provides almost the same performance at least if model backbone between NW-side and UE-side is aligned in Option 4-1.
Observation 6: For Option 3b (i.e., Direction B), more standardization effort compared with Option 1 (i.e., Direction C) may be needed to provide the same working environment, i.e., the parameters / conditions that shall be considered for inference encoder training should be aligned between NW and UE.
Observation 7: For Direction B, Issue 3 can be addressed with the combination of Direction C.
Observation 8: For Direction B, more standardization effort as UE-side additional condition may be needed to address Issue 5.
Observation 9: For Direction B, Issue 6 can be addressed with the mechanism for identifying the cause of the performance degradation.
Observation 10: For Direction B, there is less concern on NW’s proprietary disclosure, while there may be some risk for UE’s proprietary disclosure for addressing Issue 5.
Proposal 2: For the model structure and parameter specification in Direction C, train and specify model/parameters for one or a few combinations.
Proposal 3: RAN1 consider the combination of Direction A and Direction C in Rel.19.

Section 3: Temporal domain aspects of CSI compression
Observation 11: For the case of gNB failing to decode CSI, in order to provide specific input to AI/ML models for UCI missing, the mechanism to indicate the UCI missing situation from NW to UE is necessary.
Observation 12: For handing of UCI missing, the chain of the past CSI information is reset after certain period, such as every 10th or some configured value can be considered.
Proposal 4: Rank adaptation handling should be studied for handling rank > 1.

Section 4: Potential specification impact
Observation 13: For complexity comparison to study the scalability of rank > 1 solutions, the total complexity with multiple models should be taken into account.
Observation 14: For CQI determination in CSI report, further study following options.
Option 1: CQI is NOT calculated based on the output of CSI reconstruction part from the realistic channel estimation, including
Option 1a: CQI is calculated based on target CSI with realistic channel measurement
Option 1b: CQI is calculated based on target CSI with realistic channel measurement and potential adjustment
Option 2: CQI is calculated based on the output of CSI reconstruction part from the realistic channel estimation, including
Option 2a: CQI is calculated based on CSI reconstruction output, if CSI reconstruction model is available at the UE and UE can perform construction model inference with potential adjustment
The CSI reconstruction part for CQI determination at the UE is a proxy model, which is different from the actual CSI reconstruction part at the network.
Observation 15: Data collection for model training and non-real time (slow) monitoring is not required to be real-time and then latency requirement can be relaxed.
Observation 16: Ground-truth CSI reporting could be realized through U-plane at least for data collection for model training and non-real time (slow) monitoring.
Observation 17: Assuming fast monitoring is 100s of ms order, U-plane, RRC or MAC-CE can be sufficient.
Observation 18: On data sample type / format for ground-truth CSI reporting, high resolution codebook-based format e.g., legacy codebook (e.g., eType II codebook) with potential enhancements such as extend more configurations in some parameters, should be studied.
Observation 19: For NW-side data collection, at least time stamps / situation of measurement, cell ID and UE location should be considered as the UE-side additional condition.
Observation 20: For NW-side data collection, the necessity and feasibility of UE reporting Rx filter assumption to network should be studied. Instead of informing actual configuration, UE-side associated ID is necessary.
Observation 21: For UE-side data collection for UE-side training, in order to identify the scenario / configuration, how to share the NW-side additional condition should be studied. Instead of informing actual configuration, some kind of configuration ID and /or change timing of NW-side additional condition is necessary.
Observation 22: There are at least two purposes for performance monitoring.  The monitoring purpose 1 is to check new untested model / parameter behavior. The monitoring purpose 2 is to check current model / parameters are suitable to current environment.
Observation 23: If AI/ML models are trained by UE-side and AI/ML model update cycle is non-real time, UE vendor specific monitoring in offline for monitoring purpose 1 sufficiently work.
Observation 24: If AI/ML models are trained by network side and AI/ML model update cycle is non-real time, A/B test can be used, and data collected for non-real time performance monitoring for monitoring purpose 1 can also be used for model training.
Observation 25: 
For monitoring purpose 1, the monitoring mechanisms to detect the degradation could be sufficient. 
For monitoring purpose 2, both monitoring mechanism to detect the degradation and to identify the cause of degradation would be needed.
Observation 26: 
As the monitoring mechanism to detect the degradation, following is useful.
NW-side or UE-side eventual KPI-based monitoring
UE-side intermediate KPI monitoring based on the output of the CSI reconstruction model at the UE 
As the monitoring mechanism to identify the cause of degradation, following useful.
NW-side intermediate KPI monitoring based on the target CSI reported by the UE via legacy eType II codebook or eType II-like codebook
UE-side intermediate KPI monitoring based on the output of the CSI reconstruction model indicated by the NW 

3GPP TSG RAN WG1 #120bis			R1-2502504
Wuhan, China, April 7th – 11th, 2025

Agenda item:	9.1.4.1
Source: 	ETRI
Title:	Discussion on AI/ML for CSI compression
Document for:	Discussion

Conclusion
In this contribution, ETRI’s views on AI/ML for CSI compression sub-use case were shown and the following observations and proposals were made:


Proposal 1: For AI/ML-based CSI compression using two-sided model, when UE and/or NW uses past CSI information, reuse the current specification on CSI-RS transmissions as much as possible.

Proposal 2: For AI/ML-based CSI compression using two-sided model, when the target CSI is Future slot(s), study following aspects, for performance monitoring operations:
Method to align whether prediction and compression occur in separate steps or simultaneously between UE and NW
Either UE-side or NW-side performance monitoring.

Proposal 3: For temporal domain aspects Case 3, for joint CSI prediction and CSI compression operation, study LCM aspects and specification impacts, consider following option for the monitoring labels:
Option 2c: The monitoring label is derived based on the measured CSI of the future slot(s) and the measured CSI of the past slot(s) is used as input to CSI generation part.
Note: This corresponds to monitoring of end-to-end monitoring of CSI prediction and compression.

Proposal 4: Regarding the direction A and B, to consider the following aspects:
Support both Direction A and Direction B
Regarding Option 3a-1, sharing NW-side target CSI from NW to UE for UE-side offline engineering may not always be necessary
Training the actual encoder may not necessarily require prior development of the nominal decoder, at least for Option 3a-1
Among options in direction A, Option 3a-1 is preferred over Option 4-1 due to its lower overhead.

Proposal 5: Regarding the direction C, to consider the following aspects:
Only UE-side offline engineering to the specified decoder (i.e., Option 1-2) is feasible without intervendor training collaboration.

Proposal 6: For AI/ML-based CSI compression using two-sided model, study functionality/model identification procedure for supporting both Direction A and B.

Proposal 7: For AI/ML based CSI compression using two-sided model, model identification process is necessary for inference and LCM procedures, both for Direction A and B.

Proposal 8: For signaling for inter-vendor training collaboration Direction A for AI/ML-based CSI compression using two-sided model, consider the over-the-air interface delivery of model parameters and/or datasets from NW to UE.

Proposal 9: For signaling for inter-vendor training collaboration Direction A for AI/ML-based CSI compression using two-sided model, further study a method to reduce the overhead by over-the-air interface delivery of model parameters and/or datasets.

Proposal 10: For inter-vendor training collaboration option 3a-1 for AI/ML-based CSI compression using two-sided model, consider optional sharing of the target CSI from NW to UE.

Proposal 11: For inter-vendor training collaboration option 3a-1 for AI/ML-based CSI compression using two-sided model, further study a method to validate the performance of UE-side AI/ML model without sharing of target CSI from NW. 

Proposal 12: For inter-vendor training collaboration option 4-1 for AI/ML-based CSI compression using two-sided model, consider NW sharing of backbone information (of the reference CSI generation model) to UE.

Proposal 13: For dataset delivery for training collaboration type 3, for CSI compression sub-use case using two-sided model, consider the limited number of dataset samples to assess the feasibility of incorporating standardized signaling.

Proposal 14: For the performance monitoring of CSI compression sub-use case using two-sided model, conclude that Case 2-1 is feasible with considerations of complexity, latency, and accuracy, when the actual, reference, or verified nominal CSI reconstruction model is available to UE.


Observation 1: 0.5% SGCS loss is observed in inference on UE-side data when the encoder is trained using NW-side data (i.e., Case 3) compared to the upper bound performance (i.e., Case 1). This reflects that the performance of Direction B is almost similar to the upper bound.

Observation 2: Under Case 2, the performance of Options 3a-1 and 4-1 was found to be comparable.

Observation 3: Under Case 2 and Option 3a-1, the performance of Alt. 1a (using NW-side data for UE-side offline engineering) was observed to be similar to the performance of Alt. 1b and Alt. 2 (which do not utilize NW-side data for UE-side offline engineering).

Observation 4: Under Case 2 and Option 3a-1, the performance of both Alt. 1a and Alt. 1b, and Alt. 2 were observed to be similar, indicating that training the actual encoder does not necessarily be trained on the nominal decoder.

Observation 5: Under Case 2 and Option 4-1, the performance of Alt. 1b was observed as 2.53% higher than Alt. 2. 

Observation 6: The performance of UE-side offline engineering (Direction A) is comparable to direct inference (Direction B), regardless of the availability of NW-side data of Direction A.

Observation 7: It was observed that, during inference on UE-side data (representing field data), the SGCS loss increased by 37.2% when the encoder/decoder pair was trained using NW-side data (representing synthetic data) in Case 2, compared to the upper bound in Case 1. This reflects the performance of Direction C, where the degradation is significantly more severe than in Directions A and B.

Observation 8: Adjusting the decoder model on the NW side may not be practical, as the NW may be unable or unwilling to manage multiple decoder models customized for each UE, UE vendor, or UE type.

Observation 9: Adjusting both the encoder on the UE side and the decoder model on the NW side may not be practical under Direction C, as the encoder-decoder pair cannot achieve proper alignment.

Observation 10: Offline engineering on the UE side improves inference performance by 21.4%, but the performance remains lower compared to Directions A and B.

Observation 11: One example of a method to reduce the overhead by over-the-air interface delivery is RAN1-managed offline delivery using an indicator (e.g., gNB shares a URL to UE, and the UE delivers the URL to the UE side training server to obtain model parameters and/or datasets).

Observation 12: One example of a method to validate the performance of UE-side AI/ML model without sharing of target CSI from NW is an online validation procedure (e.g., NW shares the validation dataset to UE side and the UE side delivers an inference output (CSI feedback) using the validation dataset to NW, then NW can compare the output of NW side inference and the original validation dataset.)

Observation 13: Regarding performance monitoring for CSI compression sub-use case using two-sided model, for Case 2-1:
Computation complexity depends on the CSI reconstruction model on UE
No communication complexity
Latency depends on the inference time of the CSI reconstruction model on UE
Accuracy depends on the difference between the reconstructed CSI and the actual output-CSI-NW. When UE uses the actual mode, the accuracy is the upper bound.

3GPP TSG RAN WG1 #120bis	R1-2502531
Wuhan, China, April 7th – 11th, 2025

Agenda item:			9.1.4.1
Source:	Nokia
Title:	AI/ML for CSI Compression
Document for:		Discussion and Decision

Conclusion
In this contribution, study points related to the CSI compression use case were discussed, and the observations and proposals are as follows:
Proposal 1: RAN1 to consider adding a non-linear layer (e.g., sigmoid or tanh function) between the output linear layer and the quantization layer on the encoder side. This non-linear layer would help restrict the output of the linear layer to a specific range, ensuring proper functioning of the quantization process.

Observation 1: Even with the same values of reference model structure variables, performance variations may still occur due to differences in activation functions or other hyperparameters.

Proposal 2: RAN1 needs to study the performance difference of reference model structure with the same specifications of structure variables, and how it affects the separate training Option 3a.  

Proposal 3: For the choice of token dimension and feature dimension, RAN1 to select Alt1:
Alt 1: Use subband as the token dimension and Tx port as a feature dimension

Observation 2: The unified-port-number model has a slight reconstruction accuracy degradation compared to the port-specific model.

Proposal 4: RAN1 to study performance degradation of Alt2, using padding or interpolation, for the choice of scalability over the feature dimensions (TX ports):
Alt2: a common embedding layer with padding (e.g., zero-padding or other techniques for padding values) 

Proposal 5: RAN1 to study performance degradation of Alt2, using padding or interpolation, for the choice of scalability over the token dimensions (bandwidth):
Alt2: Padding at the input

Proposal 6: RAN1 to study performance of Alt1 and Alt2 for scalability over payload sizes:
Alt1: specific output linear layer for each payload configuration
Alt2: truncation/masking of the output linear layer output

Proposal 7: RAN1 should provide input on
how many different payload sizes should be supported for CSI compression
how many external configurations should be supported for CSI compression, considering
antenna array geometry
antenna muting for power saving
subband configurations

Proposal 8: To support scalability methods based on padding, RAN1 should specify the padding convention to use and specify that subsampling at the network device use the same convention.

Proposal 9: RAN1 should specify a consistent ordering convention to use for mapping PMI coefficients to input vectors of AI/ML models for CSI compression.

Proposal 10: For scalability with respect to external configurations (antenna geometry and subband configuration), RAN1 should investigate the number of different configurations to be supported.

Proposal 11: For the scalability architecture of simple padding with padding-aware compression module, RAN1 should investigate:
How (Case 1) performance depends on the number of subset configurations supported
How training complexity depends on the number of subset configurations supported
How dataset set in Direction A options 4-1 or 3a-1 depends on the number of subset configurations supported.

Proposal 12: For the scalability architecture of padding with compression-aware interpolation and simple compression model, RAN1 should investigate Case1 performance.

Proposal 13: When defining two-sided AI/ML CSI compression model, include the following additional information:
Base configuration (including Tx antenna geometry and subband configuration)
List of supported latent vector sizes
Preferred quantization codebook for each latent vector size.

Proposal 14: For supporting AI/ML based feedback, RAN1 should provide means for specifying quantization codebooks, including, at least, scalar quantization codebooks.

Proposal 15: CSI reporting configuration messages for AI/ML feedback should include
An indication of which two-sided compression model to use
An indication of the latent vector size to use 
An indication of the quantization codebook to use
An indication of the external configuration to use, which must be a subset configuration of the base configuration associated with the model

Proposal 16: When CSI reporting has been configured using AI/ML feedback
If the configured external configuration matches the base configuration associated with the model in use, the estimated CSI coefficients should be obtained directly from the decoder output.
If the configured external configuration is a subset of the base configuration associated with the model in use, the estimated CSI coefficient should be obtained by subsampling the decoder output.

Observation 3: Under direction A options 3a-1 and 4-1, interoperability between a given decoder and encoder is determined by the information shared by the NW-side with the UE-side during inter-vendor training collaboration.

Observation 4: Assuming that models are network-operator-specific and that the UE can use knowledge of the operator of the serving ID, the pairing ID only needs to be unique within an operator’s network.
Proposal 17: Pairing IDs should be set by a NW-side entity, either the network operator or the network equipment vendor.
Further study the advantages and disadvantages of each entity assigning the pairing ID.

Proposal 18: Further study whether pairing ID’s can apply to all configurations of a scalable model or if they should be unique over some or all of the scalable parameters.

Proposal 19: RAN1 to consider adopting a recurrent quantization approach (Option 2) for temporal domain Case 2.

Observation 5: The complexity of an SFT model can be significantly reduced with a relatively small performance penalty using model pruning on a recurrent quantization-based model.  The complexity reduction is similar in percentage to the reduction seen with an SF model.

Observation 6: Regarding the performance target to be shared by the NW-side with the UE-side for inter-vendor collaboration Options 3a-1 and 4-1 in Direction A, the metric should be able to serve as a quality indicator for performance monitoring of the UE-side monitoring options (Case 2-1) and (Case 2-2).

Proposal 20: Regarding the performance target to be shared by the NW-side with the UE-side for inter-vendor-collaboration Options 3a-1 and 4-1 in Direction A, at least one of its types should be identical to the type of intermediate KPI which is adopted for UE-side performance monitoring options (Case 2-1) and (Case 2-2). It is recommended to take mean SGCS as a default metric, considering its wide acceptance among 3GPP participant companies/institutions.

Observation 7: Case 1 of the NW-side monitoring option requires a UE to support both AI/ML-assisted CSI compression and a legacy eT2 codebook-based scheme. This option would incur UL overhead increase due to the required ground-truth CSI feedback for monitoring.

Proposal 21: For Case 1 of the NW-side monitoring option,
Further study methods of handling incurred UL overhead/timeline issues for ground-truth target CSI reporting

Observation 8: Case 2-1 and Case 2-2 of the UE-side monitoring options do not incur significant increase of UL overhead, especially when an event-based reporting is considered. This can relax UE’s reporting timeline requirements as well.

Observation 9: Concern for additional LCM efforts as regards the proxy decoder model (Case 2-1) or the direct SGCS estimator (Case 2-2) needs justification.

Proposal 22: Consider the following model performance monitoring options for way-forward:
•	(Case 1) NW-side monitoring mechanism with ground-truth reporting for monitoring of performance degradation and monitoring root cause identification,
o	Provided that issues raised in Proposal 21, i.e., UL overhead/timeline issue for ground-truth target CSI reporting, are resolved.
o	Propose to deprioritize this option if the above-mentioned issues are not resolved.
•	(Case 2-1) UE-side monitoring mechanism based on the output of the CSI reconstruction (proxy) model at the UE,
o	Conditioned that appropriate minimum performance requirements are to be set in RAN4.
•	(Case 2-2) UE-side monitoring mechanism via direct estimation of intermediate KPI (e.g., SGCS) without reconstructing a target CSI,
o	Conditioned that appropriate minimum performance requirements are to be set in RAN4.

Proposal 23: As regards UE-side monitoring mechanism based on the proxy CSI reconstruction model or the direct intermediate KPI estimator, RAN1 needs to conclude which metric, e.g., SGCS, should be used for performance monitoring to move forward on this topic. The minimum performance requirement for the quality of this metric should be discussed in RAN WG4 as part of the inter-operability testing for two-sided model.

Observation 10: For proper derivation of KPIs for Alt1/2/3 of Proposal 73d (on performance metric for precoded RS-based UE-side monitoring option), NW needs to transmit a non-precoded RS for channel measurements concurrently with the precoded RS, to facilitate downlink channel estimation at the UE.

Proposal 24: In AI/ML-based CSI compression using two-sided model, for precoded RS-based UE-side monitoring option, conclude that it is necessary to transmit normal CSI-RS for channel measurements as well as precoded CSI-RS concurrently. Explore a method of assigning orthogonal CDM codes for precoded CSI-RS and for normal CSI-RS for channel measurements to serve this purpose.

Observation 11: Provided that appropriate (CSI-)RS is configured for channel estimation as well as precoded (CSI-)RS for performance monitoring purposes, precoded RS based UE-side monitoring option can provide an UL resource efficient and gapless CSI acquisition procedure without incurring significant increase in computation complexity.

Proposal 25: Consider following model performance monitoring options for way-forward
•	UE-side monitoring mechanism via precoded RS transmitted from NW based on the output of the CSI reconstruction model,
o	Conditioned that means for concurrent RS for channel measurements / precoded RS for performance monitoring is agreed in 3GPP 

3GPP TSG-RAN WG1 Meeting #120-bis				                R1-2502571
Wuhan, China, April 7th – 11th, 2025

Source:	Continental Automotive
Title:                       CSI compression and other aspects on AI/ML model/data
Agenda Item:         9.1.4.1
Document for:	Discussion and Decision

Conclusions
In this contribution, we discussed open issues related to the addressed scope with previous agreements as well as the related aspects. Additionally, we ask RAN1 to discuss the following proposals: 
Proposal 1: Specify model identification LCM procedure for AI/ML-based CSI compression support.
Proposal 2: Specify mapping mechanism to associate models at the UE with corresponding models at the network for AI/ML-based CSI compression support.
Proposal 3: Specify pre-configuration of model identification for paired model ID.
Proposal 4: Support group-wise model identification for AI/ML-based CSI compression across multiple UEs with multi-vendor implementations.
Proposal 5: Specify configuration of model transfer types such as full/partial model transfers with adjustment of model complexity.
Proposal 6: Specify prioritization framework for alignment of model transfer importance with RRC state-awareness and LCM stage-dependency.
Proposal 7: Specify the unified measurement of AI/ML processing capability (e.g., CSI compression, CSI prediction, etc.) and its reporting.
Proposal 8: Specify configuration of a set of timeline types to support changing model operation behaviors, operation transitions across LCM stages, and multi-model operation scenarios.
Proposal 9: Define a separate processing unit to support AI/ML operation.

3GPP TSG RAN WG1 #120bis                                                              R1- 2502597	
Wuhan, China, Apr 7th – 11st, 2025

Agenda Item:	9.1.4.1 
Source:	Apple Inc.
Title:	Discussion on CSI compression and AI processing units  
Document for:	Discussion/Decision 

Conclusion
In this contribution, we discussed aspects on AI processing criterion and timeline, and remaining aspects of CSI compression. Based on the discussion, the following proposals have been proposed.

Proposal 1: For AI processing criteria and timeline discussion with UE side model, define separate AI processing unit pool(s) dedicated to AI-based processing.  

Proposal 2: For AI processing criteria and timeline discussion with UE side model, to enable flexible mapping of various AI use cases and hardware resources, define maximum simultaneous AI processing per CC and all CC for each use case, and maximum simultaneous AI processing per CC and all CC for all use cases. 

Proposal 3: For AI inferencing operation with UE side model, the NW should configure the AI- related use cases within the constraint of total AI processing units and the constraint of each AI use case processing units.

Proposal 4: Define different states for AI model based on the required time to start inference. 
De-activated state: This is the state where more than 10ms is required to begin inference. 
Activated state: This state corresponds to the scenario where inference can start within ms level delay. 
Inference state: In this state, the UE is actively performing inference. 

Proposal 5: Reuse the applicability report procedure for transitioning from the de-activated state to the activated state.  

Proposal 6: Define UE capability for additional processing delays when AI functionalities switch between the activated state and the inference state. 

Proposal 7: Additional processing time X1 between transition from “activated state” to “inference state” can be added to Z1/Z2/Z3 and Z1’/Z2’/Z3’ during AI functionalities switch/different CSI reports.

Proposal 8: Depending on the AI based use case, AI processing criterion and CPU criterion can be applied together or separately. 

Proposal 9: To enable unified approach for AI processing unit occupancy and resource counting regardless whether AI processing criterion and CPU criterion is applied jointly or separately,  CSI-RS resource and port counting like CPU can be applied. 

Proposal 10: Define UE capability report on the number of AI processing units used for each AI related inference report.  

Proposal 11: Apply CPU occupancy counting rule for AI based use case with network side model, including data collection for training, inferencing and performance monitoring. 
 
Proposal 12: Apply CPU occupancy counting rule for AI based use case with UE side model, for data collection for training and performance monitoring. 

Observation 1: The R18 adaptability study focused on adaptive model structure and a single parameter sets. Due to various available options and limited evaluation, a wide range of performance degradation was observed. 


Observation 2: With different parameter sets and/or dedicated model structure, baseline performance is achieved. The model parameter sets/data sets transfer overhead will linearly increase for different {number of Tx ports, CSI feedback payload size, bandwidth} combination, similar to option 4-1:
Option 4-1 requires different datasets for each {number of Tx ports, CSI feedback payload size, bandwidth} combination 
Option 3a-1 requires different model parameter sets and/or dedicated model structure for each {number of Tx ports, CSI feedback payload size, bandwidth} combination, with additional target CSI dataset.  

Observation 3: When traditional filter-based CSI prediction is used, no performance monitoring is performed in MIMO. When AI based CSI prediction is used, separate CSI prediction performance monitoring is specified.  

Proposal 13: For case 3 of time-frequency-spatial domain CSI compression, option 2a ensures E2E performance. The intermediate KPI or eventual KPI includes both compression and prediction performance. 

Observation 4: NW side performance monitoring with extended parameter sets for ground truth feedback has the following limitations: 
High feedback overhead
Additional UE complexity and power consumption due to the extended parameter sets. 
Limited guidance on whether fallback operation should be triggered

Observation 5: For UE side performance monitoring using a proxy model, since option 4-1 and 3a-1 does not provide output CSI, a new dataset consisting of {target CSI, output SGCS} is required. Additionally, the NW may need to perform performance monitoring on the proxy model.  

Observation 6: For UE side performance, NW implicitly transmit output CSI using precoded CSI-RS to the UE provides a simple and low-overhead solution. 

Observation 7: The KPI defined by precoding gain (Alt 3) can also compare the performance with legacy codebook with fall back configuration, which would assist gNB in making fallback decisions.  

Proposal 14: For CSI compression using two-sided model, for UE side performance monitoring, define monitoring KPI using precoding gain (Alt 3). Precoding gain calculation based on fall back legacy codebook configuration can be enabled. 
 
Proposal 15: For CSI compression using two-sided model, for UE side performance, further study RLF/BFD like mechanism for UE initiated report.

Proposal 16: For case 2, time-frequency-spatial domain CSI compression, for RI, consider longer term RI update across different CSI reports.   

Proposal 17: For time-frequency-spatial domain CSI compression, flexible CSI report configuration to support different cases should be studied.  

Agenda Item: 9.1.4.1
Source: MediaTek Inc.
Title:	Aditional study on AI/ML - CSI Compression
Document for: Discussion & Decision
 

Conclusion
We have the following proposals in this contribution: 
For Direction A, deprioritize option 3a-1 due to significant size and infeasibility of OTA transmission.
For Direction A, prioritize  option 3a-2 which includes transmission of encoder parameters and assistant info as a  performance guid for UE. 
For Offline engineering in direction A, discuss allowable operations and changes in performance of AI/ML models.
Specified AI/ML model in direction C can be used for initialization of training in other training solution.
Consider new options using both direction C and direction A jointly.
For NW-side AI/ML model training, NW can rely on UL CSI samples collected from SRS sent by UEs.

3GPP TSG RAN WG1 #120bis			R1-2502758
Wuhan, China, April 7th – 11th, 2025
Source:	NTT DOCOMO, INC.
Title:	Discussion on AI/ML for CSI compression
Agenda Item:	9.1.4.1
Document for: 	Discussion and Decision

Conclusion
In this contribution, we have the following observations and proposals,
Observation 1
CSI retransmission increases the occasions and overhead of UL reports, only transmitting aged information with little practical value.
Observation 2
Case 3 can be used to reduce the UL report occasions, which benefits NW’s UL transmission performance.
Observation 3
For Case 0, the input/output adaptation for a specified model structure can be solved by specifying the following,
Pre-/Post-processing (e.g., input padding schemes, output truncation schemes)
The specified TF (full model or model structure) can be designed with the largest input/output size (e.g., maximum subband number, Tx port number, the largest CSI payload size, etc.).
Input/Output adaptation layers, the structure of the adaptive layers, and their parameters if applicable, can be specified.
The specified TF (full model or model structure) can be designed based on the dimensions of the input/output adaptation layers.
For Case 2/3, the input/output adaptation can be used to extend the TF to capture temporal domain features,
The CSI of multiple time instances can be input to the TF, and the self-attention of the TF can be extended directly to capture the CSI features from multiple instances.
Observation 4
Option 2a is necessary to support both the joint and separate CSI prediction + compression for Case 3.
For separate CSI prediction + compression, Option 1 and Option 2a can be used successively for performance monitoring and identifying which part causes the performance degradation.
For joint CSI prediction + compression, Option 2a can be used.
Proposal 1
Deprioritize the approach that uses CSI retransmission for mitigating the impact of UCI loss.
Proposal 2
Prioritize the CSI compression with temporal domain aspects Case 3 for the normative work.
Proposal 3
For the scalability study purpose, no need to further discuss the structure of the decoder since it is up to NW’s implementation without scalability issues.
It is up to companies to implement scalable or configure-specific decoders based on existing agreements.
Proposal 4
Conclude in RAN1 that the adaptation for a specified model (Option 1) or model structure (Option 3a-1) can be solved by specifying,
Pre-/Post-processing schemes.
Adaptation layer structures (for Option 1 or 3a-1) and parameters (for Option 1).
For Case 2 and Case 3, support the input/output adaptation for the adaptation to the temporal domain with the following,
Input the CSI of multiple time instances to the TF and extend the self-attention of the TF to multiple time instances.
Proposal 5
Concluded in RAN1 that Directions A and C are needed for the normative work.
Direction C is used for the following objectives,
Provide a practical deployment option with minimized complexity.
Provide a baseline performance of AI/ML with simplified methods (e.g., model/layer switching) to adapt to the field data.
Alleviate the issues on the overhead/complexity/proprietary of additional information of Direction A (Option 3a-1/4-1).
Direction A is used to better adapt to the field data for further improvements.
Direction A should be supported if the user privacy or user consent issue can be addressed.
Proposal 6
If the data collection for training is started before the exchange of the dataset (for Option 4-1) or model parameters (for Option 3a-1), it can reduce the dataset exchange overhead by assigning the ID when the data collection is started.
Proposal 7
Option 2a for performance monitoring of Case 3 is necessary to support both joint and separate prediction and compression schemes.
The Rel-18 doppler codebook-like report can be used for predicted future CSI for performance monitoring Option 1. 
FFS the necessary to enhance the Rel-18 doppler codebook.
For the data collection for monitoring for Option 2a, an enhanced Rel-18 doppler codebook is necessary to support the report of historical CSI on multiple time instances.

3GPP TSG RAN WG1 #120-bis                                                                                                    R1-2502832
Wuhan, China, April 7th – 11th, 2025

Agenda item:		9.1.4.1
Source: 		Qualcomm Incorporated
Title: 			Additional study on CSI compression
Document for:		Discussion and Decision

Conclusions
In this document, we have discussed aspects related to AI/ML-based CSI compression using two-sided model. We have the following observations:
Observation 1: Important issues (issue 1, 2, 4 and 9) listed in RAN1 #118, including performance issues due to data mismatch, proprietary concern and additional information for Direction A and C have been answered positively. 
Observation 2: Remaining issues of Direction A and C are model (structure) standardization feasibility and scalability and signalling feasibility of exchanging dataset / parameters.
Observation 3: Sub-option 4-1 has less specification effort than sub-option 3a-1, while sub-option 3a-1 may yield less overhead in exchanging the model parameter than the dataset.
Observation 4: Based on observations captured in the last meeting, some companies observe 4-1 outperforms 3a-1, while some companies observe similar performance.
Observation 5: Specific linear embedding and zero-padding achieve good scalability performance for antenna port configurations with marginal loss relative to specific model development. Specific linear embedding slightly outperforms zero-padding.
Observation 6: Transformer with proper positional embedding and zero-padding achieve good generalization performance (with marginal loss relative to specific model development) if sufficient compression tasks are considered in the training phase.
Observation 7: Transformer with specific linear compression layer and specific vector quantization at the end of encoder achieve minor loss relative to specific model design for medium payload and achieve marginal loss for low payload regime.
Observation 8: Angle-delay approach may not be suitable for small payload regime considering that SD/FD bases selection would need 20+ bits.
Observation 9: Minor performance degradation (-3%) is observed by reducing the model size 3 times (from 300k + parameter to 100k) for spatial-frequency domain.
Observation 10: Complexity reduction by representing the precoder in angle-delay domain incurs non-trivial performance degradation in addition to extra payload of SD bases reporting.
Observation 11: Under similar model size, complexity reduction via NN architecture optimization yields less performance degradation than angle-delay domain approach
Observation 12: A case study on the FLOPs and latency for ML CSF and etypeII computations, shows that even though ML FLOPs count is 16x times higher, ML can achieve 20-30% latency reduction compared with etypeII. The reason is that matrix operations can be efficiently implemented, and it constitute most of the FLOPs in ML CSF.
Observation 13: There are spec-transparent performance monitoring mechanisms for legacy non-AI features. The same logic can be applied to two-sided AIML CSF.
Observation 14: SGCS of good encoder/decoder has a wide range, it is hard to make decision unless enough samples are gathered to get a reliable statistic.
Observation 15: High monitoring accuracy may not matter much in monitoring decision, the dominating factor is the natural SGCS variation.
Observation 16: Although SGCS CDF curve using eT2 as ground-truth is a bit biased from genie ground-truth, one can still make decision by eT2 curve. This is because the difference in dataset A curve (“in-distribution” CDF) and dataset B curve (“out-of-distribution” CDF) is relatively consistent no matter making decision using eT2 curve or genie curve.
Observation 17: Study is needed to determine monitoring accuracy requirement.
Observation 18: Increasing number of samples used in averaging the SGCS measurement improve the monitoring decision accuracy.
Observation 19: Using existing parameter-combinations of eType II codebook is sufficient to achieve good monitoring results.
Observation 20: UE side monitoring using SGCS estimator achieves similar monitoring results as those resulted by genie ground-truth and high-resolution eType II codebook.
Observation 21: Quantization codebook for direction C should be standardized. For Direction A, same codebook can be adopted or NW side can exchange their proprietary codebook to the UE side along with the dataset / model parameter exchange.
Observation 22: pre-determined codebook achieves similar performance to learned codebook for SQ.
Observation 23: For VQ, there is noticeable loss of pre-determined codebook compared to learned codebook as the VQ codebook should be designed specifically per data-distribution.
Observation 24: AIML inference may be performed at a dedicated AI engine (e.g., AI/ML accelerator) different from where non-AIML functionalities are executed.
Observation 25: Different AM/ML functions may share a common AI engine. Timeline and processing criteria may be different for different use cases.
Observation 26: There may be high-end UEs which have independent hardware for AIML than legacy algorithms, and also low-end UEs which may use common hardware for AIML and legacy algorithms.
Observation 27: CSI processing units depends on the model complexity, which may vary significantly across different use cases such as beam management, CSI prediction and CSI compression
Observation 28: To enable implementation flexibility, the number of APUs for each AI/ML enabled CSI processing feature can be indicated via UE capability.
Based on the observations, we propose:
Proposal 1: RAN1 should send LS to SA2 and SA5 to study other signalling approaches than using UE as relay (a.k.a., over-the-air approach).
Proposal 2: Lack of agreed signalling solution in R19 should not be the gating factor blocking the normative work because proprietary signalling can be assumed as default option in the absence of standardized signalling.
Proposal 3: For inter-vendor collaboration, recommend sub-option 4-1, 3a-1 of Direction A and Direction C for normative work.
Proposal 4: Model developed along Direction A and C will be identified with different IDs.
ID0 is assigned to the model that is designed compatible to the specified model in RAN4.
ID1, …, etc assigned to the models designed via dataset / model parameter sharing.

Proposal 5: Capture observation 5-7 into TR.
Proposal 6: Consider transformer-based model structure with positional embedding followed by series of transformer blocks
The hyper-parameters include number of transformer blocks = 6, dimension of transformer block = 64, number of attention heads = 8, dimension of attention head = 8.
The size of the positional embedding is based on max number of subbands (19) and dimension of the transformer (64)
Proposal 7: Recommend following methods for model structure scalability
For tokenization, consider subband as token and antenna port domain as feature dimension
To scale with antenna port configuration, employ specific linear embedding layers with zero-padding prior to the positional embedding
To scale with subband configuration, employ respective entries of positional embedding and perform zero-padding to the transformer output based on subband indices
To scale with payload configuration, employ specific linear compression layer and specific vector quantization after the transformer blocks
Proposal 8: Consider layer-common and rank common (Option 3-1) structure for CSI generation model and/or CSI reconstruction model for specified structures. Layer-common (Option 3-1) or layer-specific (Option 2-1) parameters can be upto vendor’s implementation choice.
Note: The standardized model structure is used to address inter-vendor collaboration complexity. The specification should be flexible to allow actual model for inference designed using all options 1-1, 1-2, 2-1, 2-2, 3-1 and 3-2.
Proposal 9: Complexity reduction is possible in the Transformer-based design with spatial-frequency domain input.
Proposal 10: Complexity reduction in spatial-frequency domain should be prioritized as it has better complexity/performance trade-off than using angular-delay domain as input.
Proposal 11: Capture observation 9-12 in the TR.
Proposal 12: Any standardized performance monitoring for two-sided AIML CSF should require additional capability than the feature itself.
Proposal 13: For NW side performance monitoring, additional UE capability includes
Concurrent processing of two-sided AIML CSF and non-AI CSI codebooks
Capability for supporting eType II CSI codebook
If higher-resolution eType II parameter combination is needed, it should be discussed and specified in the R20 NR MIMO agenda.  
The support of higher-resolution eType II parameter combination, as well as concurrent support of two-sided AIML CSF and higher-resolution eType II, should require additional UE capability.
Whether NW triggers the higher-resolution eType II parameter combination for two-sided AIML CSF performance monitoring is up to NW implementation and no further specification (other than concurrency aspects) is needed from AIML point of view.
Proposal 14: Consider existing parameter-combinations of eType II codebook for ground-truth reporting of NW side monitoring.
Proposal 15: RAN1 to perform evaluation study for monitoring. accuracy requirement, the following methodology is proposed for the study:
Generate a dataset A and develop an encoder-decoder pair E1D1 under this dataset A. Running inference on this dataset is considered to be the normal scenario where no underlying issue occurs.
A false-alarm event occurs if  where  is the average SGCS over N samples on dataset A. The probability of false-alarm event  is defined as number of false-alarm events divided by the total number of samples in dataset A.
Generate a dataset B from a different distribution than dataset A. Running inference on this dataset using E1D1 is considered to be the abnormal scenario where underlying issue (e.g., data drift) occurs.
A miss-detection event occurs if  where  is the average SGCS over N samples on dataset B using E1D1. The probability of miss-detection event  is defined as number of miss-detections events divided by the total number of samples in dataset B.
For various values of N, and for a given false-alarm probability requirement , companies to report the detection probability  corresponding to each chosen N, where the SGCS threshold  for deriving the detection probability is determined as the value that achieves the . Companies to also report the total feedback overhead over the N occasions.
Companies to also report the required value of N that achieves .
Monitoring accuracy requirement can be derived based on the value N and the accuracy of the ground truth report (i.e., eType2, higher-resolution eType2) that provides satisfactory monitoring decision performance.
Proposal 16: Consider standardized SQ codebook for the fully specified model in inter-vendor collaboration Direction C. Consider learned VQ codebook and exchange from NW to UE side for inter-vendor collaboration Direction A.
Proposal 17: Conclude that CQI calculation option 2a (where UE runs CSI reconstruction model and use its output for CQI calculation) can be employed with the consideration of potentially higher timeline and higher cost of processing unit and memory.
Proposal 18: Further study CQI calculation option 1b (CQI is calculated based on target CSI with realistic channel measurement and potential adjustment) considering adjustment measurement at UE side based on intermediate KPI or intermediate output of the CSI generation model
Proposal 19: With regards to the CPU pool for AI/ML-based CSI processing, support both of the following, as separate UE capabilities:
AI/ML shares the same CPU pool as legacy: the CPU pool for AI/ML-based CSI processing is shared with the legacy CPU pool.
Dedicated AI/ML PU pool separate from legacy: there’s a dedicated CPU pool for AI/ML-based CSI processing (namely AI/ML processing unit-APU), which is separate from legacy CPU pool.
The dedicated CPU pool for AI/ML-based use cases is shared among AI/ML-based CSI processing use cases.
Proposal 20: To address the FFS on whether the Processing Unit should be shared or separately counted among AI/ML related features/functionalities, support the following:
The AI/ML PU must be shared among AI/ML-based CSI processing use cases.
There will be no AI/ML PU counting for non-CSI-related AI/ML tasks
Proposal 21: For dedicated AI/ML PU pool, to run an AI/ML-enabled CSI processing task, the #CPUs required for that task should be counted towards both legacy CPU pool and AI/ML PU pool to account for the pre- and post-processing and memory needed to run the task, which would consume resources towards legacy CPU.

Proposal 22: For AI/ML processing units, consider the following solutions
Solution 1: Use a shared AI/ML processing unit budget and per-use-case processing unit count, reported in UE capability
Solution 2: Introduce tuples for supported combinations of concurrent AI/ML-enabled CSI processing across different features, as part of UE capability

Proposal 23: Study the latency requirements for concurrent and non-concurrent AIML processing.

3GPP TSG RAN WG1 #120bis				                                      R1-2502912
Wuhan, China, April 7th – 11th, 2025
Agenda item:        9.1.4.1
Source        :	CEWiT
Title             :	Discussion on AI/ML for CSI Compression
Document for:      Discussion 
__________________________________________________________________

Conclusion
In this contribution, we provided discussion on AI/ML based CSI compression. Based on the discussion, the following observations and proposals are provided.

Observation-3: Beyond a certain point, the number of basis components does not have a significant effect on the reconstruction.
Observation-4: In case of spatial-temporal-frequency domain based CSI compression, the channel noise and interference information obtained using legacy methods for the first instant, can be helpful in successive slots.  
Proposal-1: For evaluation of AI/ML based CSI compression for Option 3a-1, assume Alt.2 to be the starting point.  
Proposal-2: {Target-CSI} in Option 3a-1, is necessary and should be target CSI exchanged from NW side.
Proposal-3: Consider Option 4-1, as it requires less specification effort compared to Option 3a-1.
Proposal-4: For option 4-1, consider sending dataset related information like (Dataset ID) and NW sided additional condition for ensuring proper inference. 
Proposal-5: In case of Option 4-1, consider offline exchange of dataset from NW side to UE side to avoid overhead concerns. 
Proposal-6: For NW side monitoring, the existing format in high resolution codebook is sufficient enough.
Proposal-7: For model monitoring of AI/ML based CSI compression, reuse necessary signalling for transmitting the basis using legacy methods.
Proposal-8: For model monitoring of AI/ML based CSI compression, prioritize transmission of basis components-based monitoring technique.
Proposal-9: For model monitoring of AI/ML based CSI compression, prioritize  NW sided monitoring.
Proposal-10: In case of performance monitoring, consider UE side monitoring with low priority.
Proposal-11: In case of UE side monitoring, considering precoded CSI based monitoring as starting point.
Proposal-12: For CQI determination, consider Option-1(CQI is NOT calculated based on the output of CSI reconstruction part from the realistic channel estimation) to be the starting point.
Proposal-13: For CQI determination, Option 1c can be deprioritised.
Proposal-14: Study root cause with respect to Option 3a-1 and Option 4-1.




6.

3GPP TSG RAN WG1 #120-bis		   R1-2502934
Wuhan, China, April 7th – April 11th, 2025

Agenda Item:	9.1.4.1
Source:	Pengcheng Laboratory
Title:	Discussion on AI/ML-based CSI compression
Document for:	Discussion/Decision

Conclusion
In conclusion, we have the following observations and proposals on CSI compression with AI/ML for Rel-19 further study.
Observation 1: In the early phase of CSI compression with two-sided models, it is difficult to obtain a model covering diverse data distributions. Pre-trained models with encoder/decoder updates per deployment are therefore preferred. Direction C becomes more suitable once model coverage is sufficient.
Proposal 1: In the first phase of the inter-vendor collaboration study, it is recommended to prioritize Direction A, which supports dataset/model ID pairing and enables adaptive encoder/decoder updates based on deployment-specific data.
Proposal 2: Standardization of ID classification is essential. IDs should be utilized:
To identity datasets and model parameters being exchanged.
To identity encoder/decoder model pairs, encompassing distinctions in model structures (e.g., number of Tx ports, CSI feedback payload size, bandwidth) and data distributions characteristic.

Proposal 3: The UE should select an encoder based on measured CSI, include this ID within the CSI report to the network, which subsequently selects the corresponding decoder based on the reported ID.
Proposal 4: It is recommended to use eType-II W2 as the baseline precoding matrix during simulation and evaluation phases.
Proposal 5: It is proposed to initiate discussions and standardization efforts for a new CSI report format tailored explicitly for AI/ML applications. This ensures consistency between CSI reports generated with and without AI/ML methodologies.
Proposal 6: It should be the responsibility of the UE to determine the CSI-RS configuration based on its assessment of whether additional data collection is required for ongoing model training.

3GPP TSG RAN WG1 #120bis					                          R1-2502937 Wuhan, China, April 7th – 11th, 2025

						                          
Agenda Item:	9.1.4.1
Source:	IIT KANPUR
Title:	Discussion on AI/ML model-based CSI compression
Document for: Discussion and Decision

Conclusions
In this contribution, we presented simulation results showing the impact of model complexity reduction on performance, evaluated across both Case 0 and Case 2.
Our observations and proposals can be summarized as follows:
Observation 1: The lightweight model shows approximately a 21% degradation in NMSE and a 4–5% drop in SGCS performance when compared to the base model. Despite this performance decline, it offers significant efficiency gains, with a reduction of around 87–88% in FLOPs and a 70–80% decrease in the number of parameters for both Case 0 and Case 2, respectively.
Observation 2: The impact of model simplification on performance remains relatively consistent across both Case 0 and Case 2. The similar performance impact observed from applying the same model complexity reductions in both Case 0 and Case 2 suggests that the trade-off behaves consistently across different input domains.

3GPP TSG-RAN WG1 Meeting #120bis	Tdoc R1-2502952
Wuhan, China, April 7th – April 11st, 2025

Agenda Item:	9.1.4.1
Source:	Ericsson
Title:	AI/ML for CSI compression and other aspects on AI/ML model/data
Document for:	Discussion, Decision
1 

Conclusion
In the previous sections we made the following observations: 
Observation 1	Option 3a-1 with {Target CSI} sharing achieves similar performance as Option 4-1.
Observation 2	With two RAN1 meetings left for the Rel-19 SI, it is questionable whether RAN1 can complete the feasibility study on the remaining issues for option 3a-1 and Direction C.
Observation 3	It is questionable whether RAN1 can complete a feasibility study on specifying a scalable model structure during the remaining part of the Rel-19 SI.
Observation 4	Alt3 based token and feature dimension design (i.e., a fixed-size sub-block of Tx ports and subbands matrix as a token and represent the input as a sequence of tokens) is a promising technique that could allow fully scalable standardized model structure.
Observation 5	Performance target serves as a guidance for the UE-side on the achievable/expected performance during the encoder training phase.
Observation 6	SGCS is invariant to absolute phases, which makes it a better performance target compared to NMSE, if the phase of ground-truth target and/or encoder input is not standardized.
Observation 7	Eventual KPI based monitoring has low complexity, low overhead, and can capture network MU-MIMO performance. The NW can perform frequent monitoring of eventual KPIs via NW implementation-based solutions and use it as a first step for detecting potential AI/ML feature/functionality failure.
Observation 8	It is necessary to specify UE reporting high resolution target CSI to enable NW-side intermediate KPI based monitoring of the two-sided CSI-compression model performance.
-	Alternative 1: Evaluating end-to-end intermediate KPI(s).
-	Alternative 2: Evaluation the KPIs related to the encoder output.
Observation 9	NW-side monitoring of the two-sided CSI-compression model based on target CSI reporting is expected to be executed infrequently (e.g., event triggered or periodically with a large periodicity), hence, the monitoring data collection overhead for this model monitoring method is in general not an issue.
Based on the discussions above for Case 2-1, Case 2-2 and UE-side directly estimation of other performance metrics, we have the following observation and proposal:
Observation 10	The following three performance monitoring solutions can be categorized as UE-side intermediate KPI based performance monitoring:
a.	UE-side proxy-decoder based solution (Case 2-1)
b.	UE-side directly estimation of intermediate KPI (Case 2-2), and
c.	UE-side directly estimation of monitoring output, where the monitoring output is defined as intermediate KPI value range indicator
Observation 11	Among the three UE-side intermediate KPI based performance monitoring solutions, UE-side directly estimation of intermediate KPI value range indicator may provide the highest accuracy of monitoring results, with lowest computation complexity and smallest storage requirement at the UE.
Observation 12	The “target CSI” (testing dataset) and the end-to-end intermediate KPI based performance target shared from NW-side to UE-side during inter-vendor training collaboration are useful means for improving the accuracy of UE-side intermediate KPI based performance monitoring solutions.
Observation 13	To enable NW-side to check the quality of the UE reported estimated intermediate KPI in the field when needed, it requires the standard to support UE reporting the target CSI together with the encoder output (AI-CSI report) and its estimated intermediate KPI to the NW.
Observation 14	UE-side monitoring based on precoded RS transmission from NW introduces RS signaling overhead, latency and processing complexity at both UE and NW, without clear benefit.
Observation 15	In the AI-TSF compression Case 3, where UE performs prediction in a separate step before compression (i.e., joint UE-sided CSI prediction followed by two-sided TSF CSI compression), under ideal channel estimation assumption, a small to moderate SGCS gain compared to the UE-sided AI-based CSI prediction with Rel-18 MIMO eType II codebook for CSI feedback, is observed.
Observation 16	For AI CSI compression Case 3, the computational complexity ratio in terms of FLOPs between the AI model and legacy Rel-18 eType II is around 300 for payload Category X and around 200 for payload Categories Y and Z.
Observation 17	Compared to Rel-16 eType II benchmark, AI CSI compression Case 2 provides 8.5% and 20.4% SGCS gain at CSI payload X, for layer 1 and layer 2, respectively. The gain decreases as the CSI payload size increases.
Observation 18	Compared to a non-AI based benchmark (Rel. 16 eType II with W1, past Wf, past ), AI CSI compression Case 2 provides 5.9% and 17.1% SGCS gain at CSI payload X, for layer 1 and layer 2, respectively. The gain decreases as the CSI payload size increases.
Observation 19	Compared to the CSI compression Case 0, AI CSI compression Case 2 provides 1.7% and 2.4% performance gain at CSI payload X, for layer 1 and layer 2, respectively. The gain decreases as the CSI payload size increases.
Observation 20	It is up to UE implementation on whether to use dedicated AI hardware or reusing the hardware resources for legacy non-AI based CSI reporting to generate an AI based CSI report.
Observation 21	For AI/ML based CSI reporting features, the inference latency for generating CSI report may be shortened compared to legacy CSI reporting due to dedicated AI hardware and better parallelization for CSI processing, while model loading or model switching at UE, when needed, may introduce additional delay for CSI reporting.

Based on the discussion in the previous sections we propose the following:
Proposal 1	For Direction C, RAN1 should await conclusions from RAN4 feasibility study before further discussion on model specification.
Proposal 2	For Option 3a-1, conclude that {Target CSI} sharing from NW-side to UE-side is needed.
Proposal 3	For inter-vendor training collaboration, RAN1 should prioritize further study and try to make conclusions on option 4-1 during the remaining meetings for the Rel-19 SI.
Proposal 4	Conclude that Option 3a-1 and Direction C cannot be concluded feasible without a thorough study of the feasibility of representing a standardized reference model (structure + parameters) and/or a standardized reference model structure in 3GPP specifications.
Proposal 5	Study the feasibility of representing a standardized reference model (structure + parameters) and/or a standardized reference model structure in 3GPP specifications.
Proposal 6	For the performance target sharing, at least the end-to-end (encoder-decoder model pair) based performance target is supported.
Proposal 7	For the end-to-end (encoder-decoder model pair) based performance target sharing, support only SGCS-based type of performance metric.
Proposal 8	Support multiple SGCS statistics (e.g., SGCS values at X-percentiles) as the type of performance target instead of using only a single mean SGCS value across all samples.
Proposal 9	For the case of a single dataset or model parameter set containing/supporting multiple configurations (payload sizes, number of layers, max rank values, subbands, etc.), multiple performance targets are supported.
–	FFS: the association between the dataset/model-parameter ID, different configurations, testing dataset, and performance targets.
Proposal 10	The global dataset ID can be created by a combination of PLMN ID and a dataset ID, where the mobile operator assigns unique dataset ID ranges to different vendors. The bit-width of the dataset ID can be further studied.
Proposal 11	In Options 3a-1 and 4-1, the associated pairing ID can be optionally used for NW to collect NW-side target CSI for NW-side training.
Proposal 12	Conclude that it is necessary to specify UE reporting high resolution target CSI to enable NW-side intermediate KPIs based performance monitoring and performance degradation error cause detection for two-sided CSI-compression use case.
Proposal 13	In CSI compression using two-sided model use case, capture in TR that ground-truth CSI report based on enhancements of the eType-II format with new parameters shall be defined to ensure high-accuracy model performance monitoring and error cause detection at the NW-side. Potential specification impact include:
	Define the target-CSI format (e.g., Rel16 eType II CB with new parameters) for NW-side data collection (can reuse the ground truth defined for model training data collection)
	Mechanisms (e.g., RRC-message based methods) to support UE reporting the target CSI together with the encoder output for NW-side data collection for performance monitoring.
	Signaling and configuration for event triggered and periodical data collection at the NW-side.
Proposal 14	For two-sided CSI-compression use case, supported UE-side performance monitoring based on directly estimation of an intermediate-KPI based performance metric (e.g., a SGCS value range indicator), at least the following spec impact are identified:
	The format of the performance monitoring metric (e.g., a SGCS/NMSE range indicator)
	Signaling and mechanisms for UE reporting monitoring metric
	RAN4 testing of the quality of UE reported monitoring metric
	Mechanisms (e.g., RRC-message based methods) to support UE reporting target CSI, encoder output, together with the associated intermediate-KPI based performance metric to the NW.
Proposal 15	For two-sided CSI-compression use case, conclude that UE-side performance monitoring based on precoded RS transmission from NW is not supported.
Proposal 16	Further study methods to reduce the AI computational complexity for AI CSI compression Case 3.
Proposal 17	Deprioritize AI CSI compression use case 2.
Proposal 18	Limit Processing Unit and timeline discussion to AI/ML based Layer 1 CSI reporting features.
Proposal 19	RAN1 discuss and align on the assumptions of UE implementation for AI based CSI reporting first, before discussing the detailed solutions for the CSI processing criteria and timeline issues.
Proposal 20	If dedicated AI hardware will be used at least for some AI-based CSI reporting, then, support separate processing unit pools (i.e., legacy CPU pool and AI-CPU pool) between CSI reporting generated using legacy CPU and CSI reporting generated using AI-CPU.
Proposal 21	If a UE can use different types of process unit pools for different AI based CSI reporting features, then, support UE indicating the type of process unit pool (i.e., legacy CPU pool and AI-CPU pool) for CSI processing for an AI-based CSI reporting feature.
Proposal 22	If dedicated AI hardware will be used at least for some AI-based CSI reporting, Specify a new type of CSI processing unit, AI-CPU, whose design principle follows that of the legacy CPU, including
	The AI-CPU resource pool is shared among AI based CSI reporting features
	UE indicates the number of supported simultaneous AI-based CSI calculations via UE capability reporting
	Define the number of occupied AI-CPU and the occupation time for an AI based CSI report based on the report configuration (e.g., reportQuantity, CSI report time domain behavior, measurement resource configuration, etc.)
	Priority based CSI report dropping due to lack of sufficient un-occupied AI-CPUs
Proposal 23	If dedicated AI hardware will be used at least for some AI-based CSI reporting, for defining the CSI processing timeline of an AI-based CSI reporting feature, support the following
	A shortened processing timeline for the CSI report (except for the initial CSI reporting occasion) of an AI-based CSI reporting feature, comparing to the corresponding legacy non-AI based CSI report feature.
	An additional delay, , can be introduced in the CSI processing timeline for the initial CSI reporting occasion of an AI-based CSI reporting feature to account for latency introduced by model switching.
o	FFS: Definition of the initial CSI report of an AI-based CSI reporting feature.

3GPP TSG RAN WG1 Meeting #120bis	R1-2502996
Wuhan, China, April 7th – 11th, 2025
Agenda Item:	9.1.4.1
Source:	Futurewei
Title:	Discussion of CSI compression on AI/ML for NR air interface
Document for:	Discussion

Conclusions
In this contribution, we discussed our evaluation results for scalable model structure and complexity-performance tradeoff by using different model structures. In addition, we shared our views regarding remaining issues and potential specification impact. Our observations and proposals are as follows.
For evaluation of scalable model structure:
Observation 1: In model structure scalability study for temporal domain Case 0 of CSI compression using two-sided model, when handling scalability over payload configurations using Alt2 (truncation/masking of the output linear layer output) and/or Alt3 (by varying quantization parameters), performance evaluation shows there is small SGCS performance loss compared to payload size-specific models.
Alt2 has 1.5% - 3.5% SGCS performance degradation compared to the payload size-specific models.
Observation 2: In model structure scalability study for temporal domain Case 0 of CSI compression using two-sided model, using scalable model structure can potentially save significant overhead in model training, the associated LCM burden and the storage concern on both UE-side and NW-side while achieving comparable SGCS performance compared to the benchmark (payload size-specific models), at least for Alt2 and Alt3.
For evaluation of tradeoff between complexity and performance:
Observation 3: For CSI compression using two-sided model, when comparing SGCS performance across models with different space and computational complexities, minor to medium performance degradation is observed on the less complex models (from ~0.8% to ~6.3% from our study), compared to the benchmark model with much higher complexity.
Observation 4: For CSI compression using two-sided model, it is feasible for less complex models to achieve reasonable performance that is comparable with a more complex benchmark model.
Proposal 1: For CSI compression using 2-sided model, regarding the trade-off between performance and complexity/overhead, consider adopting model structure with less complexity to reduce both space and computational overhead.
For other topics in the study:
Observation 5: If Option 3b in Direction B is supported, it is preferrable for UE-side to start using the encoder or encoder parameters as soon as possible, especially for model (parameter) update case.
Observation 6: For Option 3a-1 and Option 4-1 in Direction A, the latency requirement for model parameter/dataset exchange/transfer can be more relaxed compared to Option 3b in Direction B if supported.
Proposal 2: If inter-vendor training collaboration option 3b in Direction B is supported, consider adopting over-the-air delivering method(s).
Proposal 3: If inter-vendor training collaboration option 3a-1 and/or 4-1 in Direction A are/is supported, consider adopting standardized signalling in upper layers or other offline method(s) as delivery options.
Proposal 4: For AI/ML-based CSI compression using two-sided model, further study the following potential specification impact related to quantization of CSI feedback at least for Option 3a-1/4-1 in Direction A and Option 3b in Direction B in alleviating/resolving the issues related to inter-vendor training collaboration:
Vector quantization:
Exchange of vector quantization codebook(s).
Segmentation information (if segmentation is used) of the CSI output.
Scalar quantization:
Configuration of quantization granularity and the corresponding range values.
Exchange of scalar quantization dictionary.
Proposal 5: In AI/ML-based CSI compression using two-sided model, for CQI determination, if the actual or reference CSI reconstruction model is available at UE, adopt Option 2a to determine CQI at UE. 
For CSI processing unit:
Observation 7: AI/ML-based CSI reporting needs both legacy CSI processing resources and AI/ML-specific processing resources.
Observation 8: For AI/ML-based CSI reporting, the needed legacy CSI processing part can share CPU with legacy CSI reporting.
Observation 9: The AI/ML engines / hardware on the UE are likely shared among different AI/ML-based features on the UE.
Observation 10: Separating the CPU counting for legacy and AI/ML-based CSI reporting provides a more accurate, flexible, and manageable approach to handling the computational demands of this new technology. It aligns with the potential for dedicated AI/ML processing hardware, facilitates better resource management and prioritization, enables more informative UE capability reporting, and supports efficient sharing of AI/ML resources within the UE.
Observation 11:  Supporting the sharing of AI/ML-based computing resources among different AI/ML features and functionalities is a more efficient, flexible, and scalable approach. It optimizes resource utilization, potentially reduces power consumption, simplifies UE capability reporting, and aligns with industry trends in AI hardware design. Counting the resources of this shared pool separately from legacy CPU usage provides a clearer understanding of the processing demands of AI/ML-based operations, facilitating better network management and configuration.
Proposal 6: Support to separate CPU Counting for Legacy and AI/ML-based CSI Reporting, i.e., legacy CPU and AI/ML-based CPU are from different resource pools.
Proposal 7: Support the sharing of AI/ML-based computing resources among different AI/ML features and functionalities.
Proposal 8: Adopt a new/updated timeline for model inference for the AI/ML-based counting approach of CPU and model inference processing resource. 

	
3GPP TSG RAN WG1 #120-bis                                                                                            R1-250xxxx
Wuhan, China, April 7th – 11th, 2025

Agenda item:	9.1.4.1
Source: 	Moderator (Qualcomm)
Title: 		Draft summary of Additional study on AI/ML for NR air interface: CSI compression
Document for:	Discussion and Decision

Conclusion
For NW-first training, in inter-vendor training collaboration Direction A and C, identification of root cause for performance issues can be achieved at least by the following 
Data distribution mismatch that is identified by either NW side or UE side
Via target CSI and CSI feedback pair report, e.g., with the understanding that NW side runs inference using the NW side reference CSI generation part and compare with the inference using UE side CSI generation part.
Note that this conclusion does not preclude other methods.
Note that triggering of root cause detection may be after detecting performance degradation


Agreement
For studying the standardized model structure, for temporal domain Case 0, in case of spatial-frequency domain input, adopt the following model structure as one example structure for study purpose,

Encoder description:

Decoder description:
The decoder has a mirroring design as the encoder. Details are to be discussed.



Agreement
For model structure scalability study for temporal domain Case 0,
For the choice of token dimension and feature dimension,
Alt 1: Use subband as the token dimension and Tx port as a feature dimension
The number of tokens varies with the number of subbands.
Alt 2: Use Tx port as the token dimension and subband as a feature dimension
The number of tokens varies with the number of Tx ports.
Alt 3: Use a fixed-size sub-block of Tx ports and subbands matrix (e.g., n_Tx_ports*m_Subbands) as a token and represent the input as a sequence of tokens.
The number of tokens varies with the number of Tx ports and the number of subbands.
For scalability over the feature dimension, 
Alt1: specific embedding layer for each feature size
Alt2: a common embedding layer with padding (e.g., zero-padding or other techniques for padding values) 
For scalability over the token dimension, 
Alt1: positional embedding specific to each token index
 tokens out of   token positions are used as input.
Alt2: Padding at the input
For scalability over payload configurations,
Alt1: specific output linear layer for each payload configuration
Alt2: truncation/masking of the output linear layer output
Alt3: by varying quantization parameters
Notes
Other Alternatives are not precluded.
Different Alternatives may be used in combination.
Same/similar approach is applied at the decoder side.
Evaluations to consider:
Case 1 (scalable structure): Scalable model structure described above
Using model structure as indicated in above diagram with fixed hyperparameters, companies may train a single parameter set or different parameter sets across different {number of Tx ports, CSI feedback payload size, bandwidth} assuming a common model structure.
To report whether a single parameter set or different parameter sets were used across different {number of Tx ports, CSI feedback payload size, bandwidth}. (e.g., single parameter set across different payload sizes and bandwidths, different parameter set across different number of Tx ports)
Case 2 (dedicated structure): Using model structure as indicated in above diagram with different hyperparameters, where the input and the output related hyperparameters are chosen optimally corresponding to each specific {number of Tx ports, CSI feedback payload size, bandwidth} without scalability considerations.
Different parameter sets are trained across different number of Tx ports, CSI feedback payload sizes, and bandwidths.
For each scalable model structure choice, to evaluate the SGCS performance of the non-AI/ML benchmark (e.g., Rel-16 eType2), Case 1, and Case 2, for each of {number of Tx ports, CSI feedback payload size, bandwidth}, and report the average gain (%) in SGCS of Case 1 and Case 2 over the non-AI/ML benchmark, as well as the loss (%) in the average gain of Case 1 w.r.t. Case 2. The average is performed by first calculating the SGCS gain (%) for each {number of Tx ports, CSI feedback payload size, bandwidth} and then averaging the SGCS gain (%) values over {number of Tx ports, CSI feedback payload size, bandwidth}.

Agreement
For model structure scalability study for temporal domain Case 0, for the choice of {number of Tx ports, CSI feedback payload size, bandwidth}, take the following as baseline values.
Number of Tx ports 
[N1, N2, P] = [2, 8, 2] (32 CSI-RS ports)
[N1, N2, P] = [4,4,2] (32 CSI-RS ports)
[N1, N2, P] = [2, 4, 2] (16 CSI-RS ports)
CSI feedback payload size
Multiple values, encouraged to pick at least one value from each of small, medium, and large payload regions X, Y, and Z.
Bandwidth & subband size
All subbands
A subset of all the subbands.


Agreement
Capture the following observations into TR:
For the evaluation of performance impact due to UE-side / NW-side data distribution mismatch with respect to UE-side additional condition (issue 4 and 6), 
When dataset A include dataset B, for case 2, option 3a-1, with NW-side target CSI sharing
1 source [Panasonic] observes minor performance loss (-0.22% ~ -1.09%) relative to case 1 with NW side target CSI sharing.

When dataset A include dataset B, for case 2, option 3a-1, without NW-side target CSI sharing,
1 2 source [ZTE,vivo] observes minor performance loss (-0.67 -0.006% ~ -0.97%) related to case 1 without NW side target CSI sharing.
1 source [OPPO] observes moderate performance loss (-3.3% ~ -4%) relative to case 1 without NW side target CSI sharing

When dataset A include dataset B, for case 3 (Direction B),
1 2 source [ZTE,vivo] observes minor performance loss  (-0.6% ~ -0.9%  -0.9% ~ +0.01%) relative to case 1.

When dataset A does not include dataset B, for case 2, option 3a-1, with NW-side target CSI sharing
3 sources [QC, Apple, Samsung] observe similar performance (-0.4% ~ +0.24%) as case 1 for Alt1 UE training.
3 sources [vivo, Xiaomi, ETRI] observe minor performance loss up to -2.5% relative to case 1 for Alt1 training.
1 2 source [Ericsson, ZTE] observe moderate performance loss of (-3.8 ~ -8.3%) relative to case 1 for Alt1 training.
1 source [OPPO] observe significant performance loss (-51% ~ -62.5%) relative to case 1 for Alt1 UE training.
2 sources [CATT, Futurewei] observe minor performance loss (-1.62% -2.78%  ~ -3.2%) relative to case 1 for Alt2 UE training
1 source [Apple] observe moderate loss of -6.7% relative to case 1 for Alt2 UE training.

When dataset A does not include dataset B, for case 2, option 3a-1, without NW-side target CSI sharing
4  5 sources [CATT, ZTE, Xiaomi, ETRI, vivo] observe minor to moderate performance loss (-0.4% ~ -3.9%) relative case 1.
2  3 source [QC, Ericsson, Apple] observe moderate performance loss (-6.7% ~ -8.6%) relative to case 1.
1 source [OPPO] observes significant performance loss of -62.1% relative to case 1.
 
When dataset A does not include dataset B, for case 2, option 4-1,
4 sources [QC, Apple, Xiaomi, ETRI] observe zero to minor performance loss (-2.4% ~ 0%) relative to case 1 for Alt1 UE training.
1 source [Ericsson] observe minor to moderate performance loss (-2.9% ~ -7.9%) relative to case 1 for Alt1 UE training depending on whether dataset A applies augmentation using various phase normalization methods.
4 sources [CATT, Xiaomi, Futurewei, ETRI] observe minor performance loss (-1.41% ~ -3.52%) relative to case 1 for Alt2 UE training.
1 source [Apple] observes moderate loss of -7.9% relative to case 1 for Alt2 UE training.
1 source [vivo] observes moderate loss of -7.7% relative to case 1 for 3a-1 for Alt1 UE training when backbone are different.

When dataset A does not include dataset B, for case 3, Direction B,
4 sources [CATT, vivo, ZTE, ETRI] observe minor loss to positive gain (-3.7% ~ 1%) relative to case 1.
1 source [QC] observes significant loss (-20%) relative to case 1.



Agreement
Capture the following observations into TR:
For the evaluation of performance impact due to mismatch between the distribution of the dataset used for reference model(s) training, UE-side data distribution, and NW-side data distribution (Issue 9), 
For case 2 (model trained on dataset S but tested on dataset B), 
2 sources [vivo, QC] observe significant performance loss (-7.2% ~ -17.4%) relative to case 1, where dataset B consists of actual field data.
1 source [ETRI] observes significant performance loss of -37.3% relative to case 1, where dataset S and B are different by TxRU mapping.
3 resources [ZTE, Panasonic, Ericsson] observe moderate performance loss (-2.12% ~ -4.75%) relative to case 1, where dataset S and B are generated from different scenarios, antenna layout or UE location. E.g., Uma and InH respectively, or dataset S and B are different by antenna layout and indoor/outdoor ratio.
For case 2A (model further finetuned on dataset B),
If finetune at NW and fix encoder at UE
For using actual field data as dataset B
When dataset A / B mismatch is considered but dataset B does not contain A, 1 source [vivo] observe moderate loss (-4.63% ~ -7.66%) relative to case 1. The improvement compared to case 2 is minor to moderate (2.58% ~ 9.17%).
When dataset A / B mismatch is not considered, 2 1 sources [vivo, QC] observes minor to moderate loss of (-3.5% ~ -7.66%) loss relative to case 1. The improvement compared to case 2 is moderate to significant (2.6% ~ 13.9%) depending on the specific scenario where the field data is collected.
For using synthetic data as dataset B
When dataset A / B mismatch is considered but dataset B does not contain A, 1 source [ZTE] observes moderate loss to case 1 (-4.32% ~ -4.75%). The improvement compared to case 2 is negative (-0.68% ~ -1.02%).
When dataset A includes or is the same as dataset B, 3 sources [ZTE, Panasonic, Ericsson] observe minor to moderate loss relative to case 1 (-0.97% ~ -4.67%). The improvement compared to case 2 is negative to minor (-0.61.01% ~ 1.6149%). Dataset S and B are Uma and InH respectively, or dataset S and B are different by antenna layout and indoor/outdoor ratio.
If finetune at UE and fix decoder at NW
For using actual field data as dataset B, 
2 sources [vivo, QC] observe moderate to significant loss (-6.8% ~ -12%) relative to case 1. The improvement compared to case 2 is minor moderate (0.4% ~ 7.1%) depending on the specific scenario where the field data is collected.
For using synthetic data as dataset B
2 source [Panasonic, Ericsson] observe minor to moderate loss (-1.56% ~ -3.81%) relative to case 1. The improvement compared to case 2 is marginal (0.03% ~ 0.75%). Dataset S and B are different by antenna layout and indoor/outdoor ratio.
1 source [ETRI] observes significant loss (-23.8%) relative to case 1. The improvement compared to case 2 is significant (13.5%). Dataset S and B are different by TxRU mapping
If finetune at both UE and both NW sides
For using actual field data as dataset B
When dataset A / B mismatch is considered and dataset A does not contain B, 1 source [vivo] observes significant loss (-8.5% ~ -44.85%) relative to case 1 depending on specific scenarios for data collection. The improvement compared to case 2 is negative (-28% ~ -1.3%). The loss relative to finetune at one side is negative (-1.72% ~ -37.19%).
When dataset A / B mismatch is not considered and A is equal to B, 1 2 sources [vivo, QC] observes moderate to significant loss of (-7.3% ~ -56.1%) relative to case 1. The improvement compared to case 2 is negative to significant (-39.3% ~ 10.1%). The loss relative to finetune at one side is minor to significant (-1.7 ~ 48.4%) negative to moderate (-3.82~ 4.67%).
For using synthetic data as dataset B
When dataset A / B mismatch is modelled by different scenarios (i.e., by NW side condition),
If the input is in spatial-frequency domain, 1 source [Samsung] observes minor to significant loss (-29.0% ~ -30.5%) relative to case 1 depending on whether data Set S is used in the finetune.
If the input is in angle-delay domain, 1 source [Samsung] observes similar performance to case 1 (-0.3% ~-1.3%)
When dataset A includes B or is equal to B or has same distribution as B, 
If the input is in spatial-frequency domain, and dataset A does not contain B, 2 sources [Panasonic, Ericsson] observe minor to moderate loss (-1.1% ~ -3.69%) relative case 1 depending on the modelling of synthetic data. The improvement compared to case 2 is minor (0.57% ~ 1.3%). Compared to finetune at one side, the performance improvement is similar to minor (-0.35% ~ 1%).
If the input is in spatial-frequency domain, 1 source [Samsung] observes minor to significant loss (-3.67% ~ -11.18%) relative to case 1 depending on whether data Set S is used in the finetune.
If the input is in angle-delay domain, 1 source [Samsung] observes similar performance to case 1 (-0.89% ~1.51%).
When dataset A / B mismatch is modelled by Rx antenna spacing and dataset A does not contain B and dataset A has same distribution as S (i.e., synthetic/field data distribution mismatch modeled only at the UE side (via UE-side additional condition) but not modeled at the NW-side), 1 source [Samsung] observes similar performance (-0.89% ~ 0.21%) as case 1.


Agreement
Study performance-complexity trade-off by comparing different AI/ML models, e.g. by optimizing existing designs, and/or by comparing different precoder representation in (spatial-frequency and angle-delay) or (spatial-frequency-time and angle-delay-doppler), by considering the following aspects 
Performance comparison between different AI/ML designs and benchmark schemes
Complexity numbers (FLOP, calculated/projected latency or power consumption if available, measured latency or power consumption if available) of different AI/ML designs and benchmark schemes

	
3GPP TSG RAN WG1 #120-bis                                                                                            R1-250xxxx
Wuhan, China, April 7th – 11th, 2025

Agenda item:	9.1.4.1
Source: 	Moderator (Qualcomm)
Title: 		Draft summary of Additional study on AI/ML for NR air interface: CSI compression
Document for:	Discussion and Decision

Conclusion
For NW-first training, in inter-vendor training collaboration Direction A and C, identification of root cause for performance issues can be achieved at least by the following 
Data distribution mismatch that is identified by either NW side or UE side
Via target CSI and CSI feedback pair report, e.g., with the understanding that NW side runs inference using the NW side reference CSI generation part and compare with the inference using UE side CSI generation part.
Note that this conclusion does not preclude other methods.
Note that triggering of root cause detection may be after detecting performance degradation


Agreement
For studying the standardized model structure, for temporal domain Case 0, in case of spatial-frequency domain input, adopt the following model structure as one example structure for study purpose,

Encoder description:

Decoder description:
The decoder has a mirroring design as the encoder. Details are to be discussed.



Agreement
For model structure scalability study for temporal domain Case 0,
For the choice of token dimension and feature dimension,
Alt 1: Use subband as the token dimension and Tx port as a feature dimension
The number of tokens varies with the number of subbands.
Alt 2: Use Tx port as the token dimension and subband as a feature dimension
The number of tokens varies with the number of Tx ports.
Alt 3: Use a fixed-size sub-block of Tx ports and subbands matrix (e.g., n_Tx_ports*m_Subbands) as a token and represent the input as a sequence of tokens.
The number of tokens varies with the number of Tx ports and the number of subbands.
For scalability over the feature dimension, 
Alt1: specific embedding layer for each feature size
Alt2: a common embedding layer with padding (e.g., zero-padding or other techniques for padding values) 
For scalability over the token dimension, 
Alt1: positional embedding specific to each token index
 tokens out of   token positions are used as input.
Alt2: Padding at the input
For scalability over payload configurations,
Alt1: specific output linear layer for each payload configuration
Alt2: truncation/masking of the output linear layer output
Alt3: by varying quantization parameters
Notes
Other Alternatives are not precluded.
Different Alternatives may be used in combination.
Same/similar approach is applied at the decoder side.
Evaluations to consider:
Case 1 (scalable structure): Scalable model structure described above
Using model structure as indicated in above diagram with fixed hyperparameters, companies may train a single parameter set or different parameter sets across different {number of Tx ports, CSI feedback payload size, bandwidth} assuming a common model structure.
To report whether a single parameter set or different parameter sets were used across different {number of Tx ports, CSI feedback payload size, bandwidth}. (e.g., single parameter set across different payload sizes and bandwidths, different parameter set across different number of Tx ports)
Case 2 (dedicated structure): Using model structure as indicated in above diagram with different hyperparameters, where the input and the output related hyperparameters are chosen optimally corresponding to each specific {number of Tx ports, CSI feedback payload size, bandwidth} without scalability considerations.
Different parameter sets are trained across different number of Tx ports, CSI feedback payload sizes, and bandwidths.
For each scalable model structure choice, to evaluate the SGCS performance of the non-AI/ML benchmark (e.g., Rel-16 eType2), Case 1, and Case 2, for each of {number of Tx ports, CSI feedback payload size, bandwidth}, and report the average gain (%) in SGCS of Case 1 and Case 2 over the non-AI/ML benchmark, as well as the loss (%) in the average gain of Case 1 w.r.t. Case 2. The average is performed by first calculating the SGCS gain (%) for each {number of Tx ports, CSI feedback payload size, bandwidth} and then averaging the SGCS gain (%) values over {number of Tx ports, CSI feedback payload size, bandwidth}.

Agreement
For model structure scalability study for temporal domain Case 0, for the choice of {number of Tx ports, CSI feedback payload size, bandwidth}, take the following as baseline values.
Number of Tx ports 
[N1, N2, P] = [2, 8, 2] (32 CSI-RS ports)
[N1, N2, P] = [4,4,2] (32 CSI-RS ports)
[N1, N2, P] = [2, 4, 2] (16 CSI-RS ports)
CSI feedback payload size
Multiple values, encouraged to pick at least one value from each of small, medium, and large payload regions X, Y, and Z.
Bandwidth & subband size
All subbands
A subset of all the subbands.


Agreement
Capture the following observations into TR:
For the evaluation of performance impact due to UE-side / NW-side data distribution mismatch with respect to UE-side additional condition (issue 4 and 6), 
When dataset A include dataset B, for case 2, option 3a-1, with NW-side target CSI sharing
1 source [Panasonic] observes minor performance loss (-0.22% ~ -1.09%) relative to case 1 with NW side target CSI sharing.

When dataset A include dataset B, for case 2, option 3a-1, without NW-side target CSI sharing,
1 2 source [ZTE,vivo] observes minor performance loss (-0.67 -0.006% ~ -0.97%) related to case 1 without NW side target CSI sharing.
1 source [OPPO] observes moderate performance loss (-3.3% ~ -4%) relative to case 1 without NW side target CSI sharing

When dataset A include dataset B, for case 3 (Direction B),
1 2 source [ZTE,vivo] observes minor performance loss  (-0.6% ~ -0.9%  -0.9% ~ +0.01%) relative to case 1.

When dataset A does not include dataset B, for case 2, option 3a-1, with NW-side target CSI sharing
3 sources [QC, Apple, Samsung] observe similar performance (-0.4% ~ +0.24%) as case 1 for Alt1 UE training.
3 sources [vivo, Xiaomi, ETRI] observe minor performance loss up to -2.5% relative to case 1 for Alt1 training.
1 2 source [Ericsson, ZTE] observe moderate performance loss of (-3.8 ~ -8.3%) relative to case 1 for Alt1 training.
1 source [OPPO] observe significant performance loss (-51% ~ -62.5%) relative to case 1 for Alt1 UE training.
2 sources [CATT, Futurewei] observe minor performance loss (-1.62% -2.78%  ~ -3.2%) relative to case 1 for Alt2 UE training
1 source [Apple] observe moderate loss of -6.7% relative to case 1 for Alt2 UE training.

When dataset A does not include dataset B, for case 2, option 3a-1, without NW-side target CSI sharing
4  5 sources [CATT, ZTE, Xiaomi, ETRI, vivo] observe minor to moderate performance loss (-0.4% ~ -3.9%) relative case 1.
2  3 source [QC, Ericsson, Apple] observe moderate performance loss (-6.7% ~ -8.6%) relative to case 1.
1 source [OPPO] observes significant performance loss of -62.1% relative to case 1.
 
When dataset A does not include dataset B, for case 2, option 4-1,
4 sources [QC, Apple, Xiaomi, ETRI] observe zero to minor performance loss (-2.4% ~ 0%) relative to case 1 for Alt1 UE training.
1 source [Ericsson] observe minor to moderate performance loss (-2.9% ~ -7.9%) relative to case 1 for Alt1 UE training depending on whether dataset A applies augmentation using various phase normalization methods.
4 sources [CATT, Xiaomi, Futurewei, ETRI] observe minor performance loss (-1.41% ~ -3.52%) relative to case 1 for Alt2 UE training.
1 source [Apple] observes moderate loss of -7.9% relative to case 1 for Alt2 UE training.
1 source [vivo] observes moderate loss of -7.7% relative to case 1 for 3a-1 for Alt1 UE training when backbone are different.

When dataset A does not include dataset B, for case 3, Direction B,
4 sources [CATT, vivo, ZTE, ETRI] observe minor loss to positive gain (-3.7% ~ 1%) relative to case 1.
1 source [QC] observes significant loss (-20%) relative to case 1.



Agreement
Capture the following observations into TR:
For the evaluation of performance impact due to mismatch between the distribution of the dataset used for reference model(s) training, UE-side data distribution, and NW-side data distribution (Issue 9), 
For case 2 (model trained on dataset S but tested on dataset B), 
2 sources [vivo, QC] observe significant performance loss (-7.2% ~ -17.4%) relative to case 1, where dataset B consists of actual field data.
1 source [ETRI] observes significant performance loss of -37.3% relative to case 1, where dataset S and B are different by TxRU mapping.
3 resources [ZTE, Panasonic, Ericsson] observe moderate performance loss (-2.12% ~ -4.75%) relative to case 1, where dataset S and B are generated from different scenarios, antenna layout or UE location. E.g., Uma and InH respectively, or dataset S and B are different by antenna layout and indoor/outdoor ratio.
For case 2A (model further finetuned on dataset B),
If finetune at NW and fix encoder at UE
For using actual field data as dataset B
When dataset A / B mismatch is considered but dataset B does not contain A, 1 source [vivo] observe moderate loss (-4.63% ~ -7.66%) relative to case 1. The improvement compared to case 2 is minor to moderate (2.58% ~ 9.17%).
When dataset A / B mismatch is not considered, 2 1 sources [vivo, QC] observes minor to moderate loss of (-3.5% ~ -7.66%) loss relative to case 1. The improvement compared to case 2 is moderate to significant (2.6% ~ 13.9%) depending on the specific scenario where the field data is collected.
For using synthetic data as dataset B
When dataset A / B mismatch is considered but dataset B does not contain A, 1 source [ZTE] observes moderate loss to case 1 (-4.32% ~ -4.75%). The improvement compared to case 2 is negative (-0.68% ~ -1.02%).
When dataset A includes or is the same as dataset B, 3 sources [ZTE, Panasonic, Ericsson] observe minor to moderate loss relative to case 1 (-0.97% ~ -4.67%). The improvement compared to case 2 is negative to minor (-0.61.01% ~ 1.6149%). Dataset S and B are Uma and InH respectively, or dataset S and B are different by antenna layout and indoor/outdoor ratio.
If finetune at UE and fix decoder at NW
For using actual field data as dataset B, 
2 sources [vivo, QC] observe moderate to significant loss (-6.8% ~ -12%) relative to case 1. The improvement compared to case 2 is minor moderate (0.4% ~ 7.1%) depending on the specific scenario where the field data is collected.
For using synthetic data as dataset B
2 source [Panasonic, Ericsson] observe minor to moderate loss (-1.56% ~ -3.81%) relative to case 1. The improvement compared to case 2 is marginal (0.03% ~ 0.75%). Dataset S and B are different by antenna layout and indoor/outdoor ratio.
1 source [ETRI] observes significant loss (-23.8%) relative to case 1. The improvement compared to case 2 is significant (13.5%). Dataset S and B are different by TxRU mapping
If finetune at both UE and both NW sides
For using actual field data as dataset B
When dataset A / B mismatch is considered and dataset A does not contain B, 1 source [vivo] observes significant loss (-8.5% ~ -44.85%) relative to case 1 depending on specific scenarios for data collection. The improvement compared to case 2 is negative (-28% ~ -1.3%). The loss relative to finetune at one side is negative (-1.72% ~ -37.19%).
When dataset A / B mismatch is not considered and A is equal to B, 1 2 sources [vivo, QC] observes moderate to significant loss of (-7.3% ~ -56.1%) relative to case 1. The improvement compared to case 2 is negative to significant (-39.3% ~ 10.1%). The loss relative to finetune at one side is minor to significant (-1.7 ~ 48.4%) negative to moderate (-3.82~ 4.67%).
For using synthetic data as dataset B
When dataset A / B mismatch is modelled by different scenarios (i.e., by NW side condition),
If the input is in spatial-frequency domain, 1 source [Samsung] observes minor to significant loss (-29.0% ~ -30.5%) relative to case 1 depending on whether data Set S is used in the finetune.
If the input is in angle-delay domain, 1 source [Samsung] observes similar performance to case 1 (-0.3% ~-1.3%)
When dataset A includes B or is equal to B or has same distribution as B, 
If the input is in spatial-frequency domain, and dataset A does not contain B, 2 sources [Panasonic, Ericsson] observe minor to moderate loss (-1.1% ~ -3.69%) relative case 1 depending on the modelling of synthetic data. The improvement compared to case 2 is minor (0.57% ~ 1.3%). Compared to finetune at one side, the performance improvement is similar to minor (-0.35% ~ 1%).
If the input is in spatial-frequency domain, 1 source [Samsung] observes minor to significant loss (-3.67% ~ -11.18%) relative to case 1 depending on whether data Set S is used in the finetune.
If the input is in angle-delay domain, 1 source [Samsung] observes similar performance to case 1 (-0.89% ~1.51%).
When dataset A / B mismatch is modelled by Rx antenna spacing and dataset A does not contain B and dataset A has same distribution as S (i.e., synthetic/field data distribution mismatch modeled only at the UE side (via UE-side additional condition) but not modeled at the NW-side), 1 source [Samsung] observes similar performance (-0.89% ~ 0.21%) as case 1.


Agreement
Study performance-complexity trade-off by comparing different AI/ML models, e.g. by optimizing existing designs, and/or by comparing different precoder representation in (spatial-frequency and angle-delay) or (spatial-frequency-time and angle-delay-doppler), by considering the following aspects 
Performance comparison between different AI/ML designs and benchmark schemes
Complexity numbers (FLOP, calculated/projected latency or power consumption if available, measured latency or power consumption if available) of different AI/ML designs and benchmark schemes

Agreements from RAN1 #120bis

Agreement
Introduce a dedicated AI/ML PU for AI/ML features for UE,
The AI/ML PU is used at least for quantifying the simultaneous processing of multiple CSI reports subject to CSI-related AI/ML use case(s), e.g., CSI compression (if supported), CSI prediction, BM spatial prediction, BM temporal prediction.

	
3GPP TSG RAN WG1 #120-bis                                                                                            R1-250xxxx
Wuhan, China, April 7th – 11th, 2025

Agenda item:	9.1.4.1
Source: 	Moderator (Qualcomm)
Title: 		Draft summary of Additional study on AI/ML for NR air interface: CSI compression
Document for:	Discussion and Decision

Conclusion
For NW-first training, in inter-vendor training collaboration Direction A and C, identification of root cause for performance issues can be achieved at least by the following 
Data distribution mismatch that is identified by either NW side or UE side
Via target CSI and CSI feedback pair report, e.g., with the understanding that NW side runs inference using the NW side reference CSI generation part and compare with the inference using UE side CSI generation part.
Note that this conclusion does not preclude other methods.
Note that triggering of root cause detection may be after detecting performance degradation


Agreement
For studying the standardized model structure, for temporal domain Case 0, in case of spatial-frequency domain input, adopt the following model structure as one example structure for study purpose,

Encoder description:

Decoder description:
The decoder has a mirroring design as the encoder. Details are to be discussed.



Agreement
For model structure scalability study for temporal domain Case 0,
For the choice of token dimension and feature dimension,
Alt 1: Use subband as the token dimension and Tx port as a feature dimension
The number of tokens varies with the number of subbands.
Alt 2: Use Tx port as the token dimension and subband as a feature dimension
The number of tokens varies with the number of Tx ports.
Alt 3: Use a fixed-size sub-block of Tx ports and subbands matrix (e.g., n_Tx_ports*m_Subbands) as a token and represent the input as a sequence of tokens.
The number of tokens varies with the number of Tx ports and the number of subbands.
For scalability over the feature dimension, 
Alt1: specific embedding layer for each feature size
Alt2: a common embedding layer with padding (e.g., zero-padding or other techniques for padding values) 
For scalability over the token dimension, 
Alt1: positional embedding specific to each token index
 tokens out of   token positions are used as input.
Alt2: Padding at the input
For scalability over payload configurations,
Alt1: specific output linear layer for each payload configuration
Alt2: truncation/masking of the output linear layer output
Alt3: by varying quantization parameters
Notes
Other Alternatives are not precluded.
Different Alternatives may be used in combination.
Same/similar approach is applied at the decoder side.
Evaluations to consider:
Case 1 (scalable structure): Scalable model structure described above
Using model structure as indicated in above diagram with fixed hyperparameters, companies may train a single parameter set or different parameter sets across different {number of Tx ports, CSI feedback payload size, bandwidth} assuming a common model structure.
To report whether a single parameter set or different parameter sets were used across different {number of Tx ports, CSI feedback payload size, bandwidth}. (e.g., single parameter set across different payload sizes and bandwidths, different parameter set across different number of Tx ports)
Case 2 (dedicated structure): Using model structure as indicated in above diagram with different hyperparameters, where the input and the output related hyperparameters are chosen optimally corresponding to each specific {number of Tx ports, CSI feedback payload size, bandwidth} without scalability considerations.
Different parameter sets are trained across different number of Tx ports, CSI feedback payload sizes, and bandwidths.
For each scalable model structure choice, to evaluate the SGCS performance of the non-AI/ML benchmark (e.g., Rel-16 eType2), Case 1, and Case 2, for each of {number of Tx ports, CSI feedback payload size, bandwidth}, and report the average gain (%) in SGCS of Case 1 and Case 2 over the non-AI/ML benchmark, as well as the loss (%) in the average gain of Case 1 w.r.t. Case 2. The average is performed by first calculating the SGCS gain (%) for each {number of Tx ports, CSI feedback payload size, bandwidth} and then averaging the SGCS gain (%) values over {number of Tx ports, CSI feedback payload size, bandwidth}.

Agreement
For model structure scalability study for temporal domain Case 0, for the choice of {number of Tx ports, CSI feedback payload size, bandwidth}, take the following as baseline values.
Number of Tx ports 
[N1, N2, P] = [2, 8, 2] (32 CSI-RS ports)
[N1, N2, P] = [4,4,2] (32 CSI-RS ports)
[N1, N2, P] = [2, 4, 2] (16 CSI-RS ports)
CSI feedback payload size
Multiple values, encouraged to pick at least one value from each of small, medium, and large payload regions X, Y, and Z.
Bandwidth & subband size
All subbands
A subset of all the subbands.


Agreement
Capture the following observations into TR:
For the evaluation of performance impact due to UE-side / NW-side data distribution mismatch with respect to UE-side additional condition (issue 4 and 6), 
When dataset A include dataset B, for case 2, option 3a-1, with NW-side target CSI sharing
1 source [Panasonic] observes minor performance loss (-0.22% ~ -1.09%) relative to case 1 with NW side target CSI sharing.

When dataset A include dataset B, for case 2, option 3a-1, without NW-side target CSI sharing,
1 2 source [ZTE,vivo] observes minor performance loss (-0.67 -0.006% ~ -0.97%) related to case 1 without NW side target CSI sharing.
1 source [OPPO] observes moderate performance loss (-3.3% ~ -4%) relative to case 1 without NW side target CSI sharing

When dataset A include dataset B, for case 3 (Direction B),
1 2 source [ZTE,vivo] observes minor performance loss  (-0.6% ~ -0.9%  -0.9% ~ +0.01%) relative to case 1.

When dataset A does not include dataset B, for case 2, option 3a-1, with NW-side target CSI sharing
3 sources [QC, Apple, Samsung] observe similar performance (-0.4% ~ +0.24%) as case 1 for Alt1 UE training.
3 sources [vivo, Xiaomi, ETRI] observe minor performance loss up to -2.5% relative to case 1 for Alt1 training.
1 2 source [Ericsson, ZTE] observe moderate performance loss of (-3.8 ~ -8.3%) relative to case 1 for Alt1 training.
1 source [OPPO] observe significant performance loss (-51% ~ -62.5%) relative to case 1 for Alt1 UE training.
2 sources [CATT, Futurewei] observe minor performance loss (-1.62% -2.78%  ~ -3.2%) relative to case 1 for Alt2 UE training
1 source [Apple] observe moderate loss of -6.7% relative to case 1 for Alt2 UE training.

When dataset A does not include dataset B, for case 2, option 3a-1, without NW-side target CSI sharing
4  5 sources [CATT, ZTE, Xiaomi, ETRI, vivo] observe minor to moderate performance loss (-0.4% ~ -3.9%) relative case 1.
2  3 source [QC, Ericsson, Apple] observe moderate performance loss (-6.7% ~ -8.6%) relative to case 1.
1 source [OPPO] observes significant performance loss of -62.1% relative to case 1.
 
When dataset A does not include dataset B, for case 2, option 4-1,
4 sources [QC, Apple, Xiaomi, ETRI] observe zero to minor performance loss (-2.4% ~ 0%) relative to case 1 for Alt1 UE training.
1 source [Ericsson] observe minor to moderate performance loss (-2.9% ~ -7.9%) relative to case 1 for Alt1 UE training depending on whether dataset A applies augmentation using various phase normalization methods.
4 sources [CATT, Xiaomi, Futurewei, ETRI] observe minor performance loss (-1.41% ~ -3.52%) relative to case 1 for Alt2 UE training.
1 source [Apple] observes moderate loss of -7.9% relative to case 1 for Alt2 UE training.
1 source [vivo] observes moderate loss of -7.7% relative to case 1 for 3a-1 for Alt1 UE training when backbone are different.

When dataset A does not include dataset B, for case 3, Direction B,
4 sources [CATT, vivo, ZTE, ETRI] observe minor loss to positive gain (-3.7% ~ 1%) relative to case 1.
1 source [QC] observes significant loss (-20%) relative to case 1.



Agreement
Capture the following observations into TR:
For the evaluation of performance impact due to mismatch between the distribution of the dataset used for reference model(s) training, UE-side data distribution, and NW-side data distribution (Issue 9), 
For case 2 (model trained on dataset S but tested on dataset B), 
2 sources [vivo, QC] observe significant performance loss (-7.2% ~ -17.4%) relative to case 1, where dataset B consists of actual field data.
1 source [ETRI] observes significant performance loss of -37.3% relative to case 1, where dataset S and B are different by TxRU mapping.
3 resources [ZTE, Panasonic, Ericsson] observe moderate performance loss (-2.12% ~ -4.75%) relative to case 1, where dataset S and B are generated from different scenarios, antenna layout or UE location. E.g., Uma and InH respectively, or dataset S and B are different by antenna layout and indoor/outdoor ratio.
For case 2A (model further finetuned on dataset B),
If finetune at NW and fix encoder at UE
For using actual field data as dataset B
When dataset A / B mismatch is considered but dataset B does not contain A, 1 source [vivo] observe moderate loss (-4.63% ~ -7.66%) relative to case 1. The improvement compared to case 2 is minor to moderate (2.58% ~ 9.17%).
When dataset A / B mismatch is not considered, 2 1 sources [vivo, QC] observes minor to moderate loss of (-3.5% ~ -7.66%) loss relative to case 1. The improvement compared to case 2 is moderate to significant (2.6% ~ 13.9%) depending on the specific scenario where the field data is collected.
For using synthetic data as dataset B
When dataset A / B mismatch is considered but dataset B does not contain A, 1 source [ZTE] observes moderate loss to case 1 (-4.32% ~ -4.75%). The improvement compared to case 2 is negative (-0.68% ~ -1.02%).
When dataset A includes or is the same as dataset B, 3 sources [ZTE, Panasonic, Ericsson] observe minor to moderate loss relative to case 1 (-0.97% ~ -4.67%). The improvement compared to case 2 is negative to minor (-0.61.01% ~ 1.6149%). Dataset S and B are Uma and InH respectively, or dataset S and B are different by antenna layout and indoor/outdoor ratio.
If finetune at UE and fix decoder at NW
For using actual field data as dataset B, 
2 sources [vivo, QC] observe moderate to significant loss (-6.8% ~ -12%) relative to case 1. The improvement compared to case 2 is minor moderate (0.4% ~ 7.1%) depending on the specific scenario where the field data is collected.
For using synthetic data as dataset B
2 source [Panasonic, Ericsson] observe minor to moderate loss (-1.56% ~ -3.81%) relative to case 1. The improvement compared to case 2 is marginal (0.03% ~ 0.75%). Dataset S and B are different by antenna layout and indoor/outdoor ratio.
1 source [ETRI] observes significant loss (-23.8%) relative to case 1. The improvement compared to case 2 is significant (13.5%). Dataset S and B are different by TxRU mapping
If finetune at both UE and both NW sides
For using actual field data as dataset B
When dataset A / B mismatch is considered and dataset A does not contain B, 1 source [vivo] observes significant loss (-8.5% ~ -44.85%) relative to case 1 depending on specific scenarios for data collection. The improvement compared to case 2 is negative (-28% ~ -1.3%). The loss relative to finetune at one side is negative (-1.72% ~ -37.19%).
When dataset A / B mismatch is not considered and A is equal to B, 1 2 sources [vivo, QC] observes moderate to significant loss of (-7.3% ~ -56.1%) relative to case 1. The improvement compared to case 2 is negative to significant (-39.3% ~ 10.1%). The loss relative to finetune at one side is minor to significant (-1.7 ~ 48.4%) negative to moderate (-3.82~ 4.67%).
For using synthetic data as dataset B
When dataset A / B mismatch is modelled by different scenarios (i.e., by NW side condition),
If the input is in spatial-frequency domain, 1 source [Samsung] observes minor to significant loss (-29.0% ~ -30.5%) relative to case 1 depending on whether data Set S is used in the finetune.
If the input is in angle-delay domain, 1 source [Samsung] observes similar performance to case 1 (-0.3% ~-1.3%)
When dataset A includes B or is equal to B or has same distribution as B, 
If the input is in spatial-frequency domain, and dataset A does not contain B, 2 sources [Panasonic, Ericsson] observe minor to moderate loss (-1.1% ~ -3.69%) relative case 1 depending on the modelling of synthetic data. The improvement compared to case 2 is minor (0.57% ~ 1.3%). Compared to finetune at one side, the performance improvement is similar to minor (-0.35% ~ 1%).
If the input is in spatial-frequency domain, 1 source [Samsung] observes minor to significant loss (-3.67% ~ -11.18%) relative to case 1 depending on whether data Set S is used in the finetune.
If the input is in angle-delay domain, 1 source [Samsung] observes similar performance to case 1 (-0.89% ~1.51%).
When dataset A / B mismatch is modelled by Rx antenna spacing and dataset A does not contain B and dataset A has same distribution as S (i.e., synthetic/field data distribution mismatch modeled only at the UE side (via UE-side additional condition) but not modeled at the NW-side), 1 source [Samsung] observes similar performance (-0.89% ~ 0.21%) as case 1.


Agreement
Study performance-complexity trade-off by comparing different AI/ML models, e.g. by optimizing existing designs, and/or by comparing different precoder representation in (spatial-frequency and angle-delay) or (spatial-frequency-time and angle-delay-doppler), by considering the following aspects 
Performance comparison between different AI/ML designs and benchmark schemes
Complexity numbers (FLOP, calculated/projected latency or power consumption if available, measured latency or power consumption if available) of different AI/ML designs and benchmark schemes

Agreements from RAN1 #120bis

Agreement
Introduce a dedicated AI/ML PU for AI/ML features for UE,
The AI/ML PU is used at least for quantifying the simultaneous processing of multiple CSI reports subject to CSI-related AI/ML use case(s), e.g., CSI compression (if supported), CSI prediction, BM spatial prediction, BM temporal prediction.


Agreement
For inter-vendor-collaboration Options 3a-1 and 4-1 in Direction A, confirm SGCS and NMSE as the type of performance metric that may be used for the performance target shared as additional information along with the exchanged dataset or the model parameters.
FFS: when to use SGCS, NMSE, and which one to use or both, and relationship with the inter-vender collaboration sub-options.
FFS: details of the format of the performance target
Option 1: Average performance target, e.g. average SGCS and/or average NMSE
Option 2: distribution of the performance target, e.g., SGCS / NMSE for 5, 10, 20, 30 percentiles, etc.
FFS: whether multiple performance targets should be exchanged for different configurations, such as antenna ports configuration, subband configuration and payload configuration, etc., along with each exchanged dataset or model parameters 

Agreement
In Options 3a-1 and 4-1, the exchanged dataset or the model parameters can be associated with an ID for pairing related discussion, then
The same ID can be used for UE to collect UE-side target CSI for UE-side training 
The same ID can be used for applicability inquiry and reporting
The same ID can be used for inference configuration
The same ID can be used for NW-side data collection
FFS: whether ID/even same ID is needed for monitoring configuration
FFS: where the ID is assigned
Note: whether the purpose for pair will be specified will be discussed separately.

Agreement
In Direction C, the fully standardized reference model is associated with an ID for pairing related discussion, then
The same ID can be used for UE to collect UE-side target CSI for UE-side training 
The same ID can be used for applicability inquiry and reporting
The same ID can be used for inference configuration
The same ID can be used for NW-side data collection
FFS: whether ID/even same ID is needed for monitoring configuration
FFS: where the ID is assigned or how the ID is specified
Note: whether the purpose for pair will be specified will be discussed separately.

Agreement
For inter-vendor collaboration Direction C, 
Use standardized quantization codebook
For inter-vendor collaboration Direction A Options 4-1, 3a-1 (with or without NW-side target CSI sharing), 
Standardize configuration(s) of quantization codebook, e.g., scalar or vector quantization, segment size of VQ, codebook size.
FFS: applicability of the above for Case 2 
Exchange quantization codebook of (the selected) standardized configuration(s) from NW-side to UE-side along with each exchanged dataset or model parameters.
FFS: whether quantization codebook may be different across different payload size configurations 

	
3GPP TSG RAN WG1 #120-bis                                                                                            R1-250xxxx
Wuhan, China, April 7th – 11th, 2025

Agenda item:	9.1.4.1
Source: 	Moderator (Qualcomm)
Title: 		Draft summary of Additional study on AI/ML for NR air interface: CSI compression
Document for:	Discussion and Decision

Conclusion
For Direction C, confirm that the specified model should be trained using synthetic data (answer to issue 8).
For inter-vendor collaboration option 3a-1 of Direction A, confirm that the specified model structure should be determined using synthetic data.

	
3GPP TSG RAN WG1 #120-bis                                                                                            R1-250xxxx
Wuhan, China, April 7th – 11th, 2025

Agenda item:	9.1.4.1
Source: 	Moderator (Qualcomm)
Title: 		Draft summary of Additional study on AI/ML for NR air interface: CSI compression
Document for:	Discussion and Decision

Conclusion
For NW-side monitoring with target CSI reporting
Target CSI reporting via legacy CSI codebooks can be used for NW-side monitoring
Target CSI reporting with CSI codebook enhancement via higher-resolution parameter combination may be beneficial for improving NW-side monitoring with additional cost of complexity and overhead at UE side.

TDoc file unavailable

	
3GPP TSG RAN WG1 #120-bis                                                                                            R1-2503034
Wuhan, China, April 7th – 11th, 2025

Agenda item:	9.1.4.1
Source: 	Moderator (Qualcomm)
Title: 		Final summary of Additional study on AI/ML for NR air interface: CSI compression
Document for:	Discussion and Decision

Conclusion
UE-side monitoring is feasible.
Observation
Some companies think that, at least in some UE-side monitoring options, NW-side monitoring with target CSI reporting is needed to check the reliability of UE-side monitoring reports.