3GPP TSG RAN WG1 Meeting #121	R1-2503231
St. Julian’s, Malta, May 19th – 23rd, 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 discuss our view on the progress on the Rel-19 further study on CSI compression (two-sided model), including observations drawn from results submitted by companies as of RAN1#120bis and our recommendation regarding how to move forward after Rel-19. Our observations and proposals are as follows.
For improving trade-off between performance and complexity/overhead:
Observation 1: For CSI compression using 2-sided model, regarding the trade-off between performance and complexity/overhead, when extending spatial-frequency (SF) domain to temporal-spatial-frequency (TSF) domain CSI compression and/or CSI compression plus CSI prediction, the majority of results indicated TSF Case 2 and Case 3 have further performance gain over Case 0 and/or the benchmark based on the results available as of RAN1#120bis.
Observation 2: For CSI compression using 2-sided model, regarding the trade-off between performance and complexity/overhead, using model(s) with lower complexity (e.g., ~10X less complexity than the most complex model) based on agreed scalable model structure can still achieve decent performance gain over the baseline (eType II) based on the results available as of RAN1#120bis, at least for Case 0.
Proposal 1: For CSI compression using 2-sided model, conclude that improving trade-off between performance and complexity/overhead is feasible.
For alleviating/resolving issues related to inter-vendor training collaboration:
Observation 3: For AI/ML-based CSI compression using two-sided model, performance loss concern identified by [Issue 4] for Direction A in inter-vendor training collaboration can be minimized/reduced with the help from dataset generation strategy and/or UE training alternatives.
Observation 4: For AI/ML-based CSI compression using two-sided model, based on the results and observations drawn, RAN1 has already concluded that [Issue 4] for Direction A in inter-vendor training collaboration can be addressed.
Observation 5: For AI/ML-based CSI compression using two-sided model, performance impact concern/issue identified by [Issue 9] for Direction C in inter-vendor training collaboration, there are relatively fewer results for each discussed scenario compared to Direction A; thus, it is uncertain about the performance Direction C would achieve in the field.
Observation 6: For AI/ML-based CSI compression using two-sided model, Direction C in inter-vendor training collaboration may require additional discussion/study to confirm the feasibility/practicality and address performance concern/issue.
Proposal 2: For AI/ML-based CSI compression using two-sided model, considering the agreements reached related to additional information to be shared from NW-side to UE-side, conclude that [Issue 1] identified for Direction A in inter-vendor training collaboration can be considered as addressed.
Proposal 3: For AI/ML-based CSI compression using two-sided model, considering the agreement reached related to proprietary information concern of Direction A, conclude that [Issue 2] identified for Direction A in inter-vendor training collaboration can be considered as addressed.
Proposal 4: For AI/ML-based CSI compression using two-sided model, based on LS reply from RAN2 for NW-side sharing dataset/model parameter transfer, conclude that [Issue 3] identified for Direction A in inter-vendor training collaboration can be considered as addressed.
Proposal 5: For AI/ML-based CSI compression using two-sided model, conclude that Direction A in inter-vendor training collaboration is feasible.
Proposal 6: For AI/ML-based CSI compression using two-sided model, conclude that Direction B in inter-vendor training collaboration is deprioritized as the evaluation results submitted are relatively limited and there is no agreement reached for majority of the issues identified for Direction B. 
Proposal 7: For AI/ML-based CSI compression using two-sided model, conclude that [Issue 8] identified for Direction C in inter-vendor training collaboration can be considered as addressed.
Proposal 8: For AI/ML-based CSI compression using two-sided model, conclude that Direction C in inter-vendor training collaboration is considered as lower priority given that some of the identified issues have not been fully addressed.
For other topics related to potential spec impact for CSI compression:
Observation 7: For AI/ML-based CSI compression using two-sided model, UE-side monitoring may be considered as supplementary or optional if NW-side monitoring is supported as its support depends on UE capability.
Observation 8: For AI/ML-based CSI compression using two-sided model, out of the two options for UE-side monitoring, the option: “based on intermediate KPI” is preferred as the other option requires additional study/discussion on monitoring KPI(s).
Proposal 9: 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 conclusion on Rel-19 further study of CSI compression:
Observation 9: Based on the evaluations/observations drawn and study of inter-vendor training collaboration options, consider the following on moving forward for AI/ML-based CSI compression using two-sided model:
Direction A: prioritized
Direction B: deprioritized
Direction C: lower priority than Direction A
Proposal 10: Conclude that for potential normative work on AI/ML-based CSI compression using two-sided model, RAN1 recommends specifying Direction A.

3GPP TSG-RAN WG1 Meeting #121	Tdoc R1-2503245
St Julian’s, Malta, May 19th – 23rd, 2025

Agenda Item:	9.1.4.1
Source:	Ericsson
Title:	AI/ML for CSI compression
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, but Option 3a-1 requires higher specification effort comparing to Option 4-1.
Observation 2	For Option 3a-1 and Direction C, there is lack of study on how to represent an AI model/model structure in 3GPP specifications.
Observation 3	For Direction C and Option 3a-1, evaluating and agreeing on a reference model structure require large effort, regardless of whether the work is done in RAN1 or RAN4.
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.
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.

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 recommending specifying reference model for direction C for potential normative work.
Proposal 2	For Option 3a-1 and Option 4-1, conclude that OTA signalling based dataset/model parameter sharing is not supported for CSI compression use case.
Proposal 3	For Option 3a-1 and Option 4-1, RAN1 should await confirmation from RAN3/SA2/SA5 on the feasibility of non-OTA solutions for dataset/model-parameter sharing before recommending any remaining option in Direction A for potential normative work.
Proposal 4	For Direction A, conclude that Option 3a-1 without {Target CSI} sharing from NW-side to UE-side is not supported for potential normative work.
Proposal 5	Support only inter-vendor training collaboration Option 4-1 and use case 0 for potential normative work, conditioned on that the feasibility of non-OTA signalling solutions is confirmed by RAN3/SA2/SA5 and the performance testing feasibility is confirmed by RAN4.
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/model parameter ID, where the mobile operator assigns unique dataset/model parameter ID ranges to different vendors. The bit-width of the dataset/model parameter ID can be further studied.
Proposal 11	In Direction C, the fully standardized reference model is associated with an ID for pairing related discussion, and the ID is specified together with the reference model.
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)
	Target CSI and CSI feedback pair reporting mechanisms.
	Signaling and configuration for event triggered and periodical monitoring 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	For CSI compression use case, for the aspect of CSI processing criteria for model inference, support the option where only dedicated APU is occupied. The number of occupied APU and the total number of APUs are reported as part of the UE capability.

3GPP TSG-RAN WG1 Meeting #121	  R1-2503254
St Julian’s, Malta, May 19 – 23, 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, the following observations and proposals are provided.
Observation 1: For the Rel-19 study, there is strong need to explore the necessary CSI compression solution to achieve high gains in widely deployed TDD scenarios.
Observation 2: For the approaches to generate DL precoding in legacy:
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 usage.
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 that incurs higher UE complexity, CSI overhead, and power consumption as the price.
Observation 3: For TDD scenario, fusion of SRS measurement and AI/ML compressed CSI feedback can be performed to optimize 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: 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 7: 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 8: 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 9: UE side proxy model (Case 2-1/2-2) is likely to operate under collaboration level x manner and can hardly be monitored or trusted by the NW side.
Observation 10: For the additional potential spec impact of temporal domain CSI compression Case 3, JPC is more realistic than SPC from training data collection perspective and simpler than SPC from inference/model management perspective.
For JPC,
Training data collection Option 2 is applicable, where the channel matrix is collected as the input CSI for training the joint model of prediction and compression.
Monitoring Option 2 is applicable, where the measured future CSI is collected for monitoring the joint model of prediction and compression.
For SPC,
Training data collection Option 1 is applicable. But it implies the CSI prediction part should be developed separately and earlier than the development of the CSI compression part by the UE side.
Monitoring Option 1 is applicable for monitoring the CSI compression model.
Observation 11: For the additional potential spec impact of temporal domain CSI compression, Case 3 may need more spec efforts from training data collection, inference/model management, and monitoring perspectives.

Proposal 1: For the spec impacts of inter-vendor collaboration Option 4-1 of Direction A, consider at least the following aspects:
Data sample format related aspects, 
Type of ground-truth CSI, including precoding matrix and channel matrix.
Dimension of Target CSI/CSI feedback.
CSI feedback related information.
Dataset construction related aspects, e.g., number of data samples, dataset ID, dataset split/segmentation information, association between ground-truth CSI and CSI feedback subject to a data sample, past CSI information for temporal domain Case 2, observation window/prediction window information for temporal domain Case 3.
Scalability information.
Performance target information.
Proposal 2: For the spec impacts of inter-vendor collaboration Option 3a-1 of Direction A, consider at least the following aspects:
Model structure related aspects, 
Standardized value(s) for the hyper parameters.
Scalability method.
Quantization method.
Model pairing related aspects, e.g., model ID and its related procedures such as data collection, inference, applicability report, etc., parameter/dataset split/segmentation.
Performance target information.
Proposal 3: For the spec impacts of inter-vendor collaboration Direction C, consider at least model pairing related aspects, e.g., model ID and its related procedures such as data collection, inference, applicability report, etc.
Proposal 4: 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 5: For NW side monitoring:
Support eT2-like high-resolution codebook for reporting format for ground-truth CSI subject to precoding matrix and raw channel matrix.
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 6: There is no strong motivation for specifying the UE side proxy model (Case 2-1/2-2) for monitoring.
Proposal 7: For quantization methods of the CSI report, further study potential specification impact on quantization alignment using standardized quantization scheme.
For vector quantization,
Configuration/reporting/updating of the quantization dictionary.
Segmentation of the CSI generation model output to map with short VQ vector.
For scalar quantization,
The configuration of the quantization granularity/range.
Proposal 8: For the study of CQI determination in inference, consider Option 1 (CQI is NOT calculated based on the output of CSI reconstruction part from the realistic channel estimation) as a starting point.
Proposal 9: Regarding the occupancy of AI/ML PU and legacy CPU for AI/ML-based CSI compression, for inference, consider the following two options with UE capability to report the selected option.
Option 1: Only legacy CPU is occupied.
Option 2: Both dedicated AI/ML PU and legacy CPU are occupied.
Equal occupancy duration can be assumed for AI/ML PU and legacy CPU.
Proposal 10: For A-CSI report subject to a compression model, the CSI processing timeline is impacted by the active/inactive state of the functionality. When the A-CSI report is triggered:
If the model is subject to “active state”, the CSI processing is subject to a shorter timeline (Zref).
If the model is subject to “inactive state”, the CSI processing is subject to a longer timeline (Zref+ΔT) as it additionally includes the activation delay of ΔT.
Proposal 11: For the additional potential spec impact of temporal domain CSI compression Case 2, consider NW signaling to reset of historical CSI information at UE with higher priority.
Other specification aspects of training data collection, inference, and monitoring are mostly common between Case 0 and Case 2.
Proposal 12: 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 13: Towards Rel-20, provide specification support for CSI compression for both FDD and TDD scenarios, encompassing:
Specify the sub-use cases of spatial-frequency domain CSI compression Case 0 and temporal-spatial-frequency domain CSI compression Case 2 (using past CSI).
Specify necessary signaling/mechanism(s) to enable inter-vendor collaboration with Direction A sub-option 4-1 (led by RAN1), Direction A sub-option 3a-1 (led by RAN4), and Direction C (led by RAN4).
Specify necessary signaling/mechanism(s) to facilitate LCM operations specific to the CSI compression, including at least training data collection, inference, and monitoring.

3GPP TSG RAN WG1 #121 		R1- 2503350
St Julian’s, Malta, May 19th – 23th, 2025

Source:	vivo
Title:	Discussion on CSI compression 
Agenda Item:	9.1.4.1
Document for:	Discussion and Decision

Conclusion
In this contribution, we have the following observations and proposals:
Both NMSE and SGCS can be used as performance target shared as additional information along with the exchanged model parameters, for Option 3a-1 without target CSI.
For performance target, there is one to one mapping relation between NMSE and SGCS. The NW can well map the performance target to either NMSE and SGCS without ambiguity.
Without necessary model hyper-parameter information as additional information shared, there is performance degradation for 4-1.
It is necessary for NW to exchange model backbone/hyper-parameter in option 4-1 as additional information. 
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.
Understanding 1 (The pairing ID should be specified) is more aligned with direction A, but the potential for fine-tuning performance may be limited due to the lack of differentiation between scenarios or deployments.
Understanding 2 (The pairing ID Should be assigned to differential the scenarios /deployments) is incomplete and specified ID is still needed to indicate which reference model applies. 
Understanding 2 (The pairing ID Should be assigned to differential the scenarios /deployments) introduce new mapping relationship in which single model (i.e., one fully standardized reference model) can be associated with multiple pairing IDs. 
Following similarities for RAN1 reference model and RAN4 reference model are observed.
Similarity 1: RAN1 and RAN4 both use UMa to generate reference model, and RAN4 from testing perspective need to consider suitable model structure and scalability, the design would be similar from model structure perspective since all the evaluation assumptions are the same.
Similarity 2: No matter which option RAN1 take for inter-vendor collaboration, RAN4 specified model needs to take field performance into account. Otherwise, the testing would be meaningless. If RAN1 3a-1 or direction C is supported, the simplest way to design RAN4 model for testing is to use the same model for RAN1 reference model since the testing would guarantee the performance in the field. 
Similarity 3: Both RAN1 reference model and RAN4 reference model have low proprietary issue, since UE, NW and TE can develop their real models which are different from reference model. 
Similarity 4: For all RAN4 testing options, reference/testing encoder/decoder needs to be defined based on agreement and conclusions of RAN1 and RAN4. Then the same model structure can be used for 3a-1 for reference.

The understanding 1(The pairing ID should be specified) can be supported as starting point, and further study the necessity of an additional pairing ID which is used to develop different models tailored to various scenarios or deployments for the same reference model.
RAN1 and RAN4 can share the workload for defining the reference/testing model for potential recommended direction/sub-directions. RAN4 can take the lead for model structure/parameters specification and take RAN1 study/work output on model structure into account.
Model parameter [/dataset] and other additional information exchange signaling for 3a-1 [and 4-1]; (RAN2, RAN1, SA2)
Conduct further study on OTA v.s. non-OTA;
Including other assistant information exchange, e.g., performance metric;
Note: Consider to further coordinate between RAN and SA and put an objective in Rel-20 RAN AIML over air WID to finalize the whole picture of 5G-A UE side data collection, model transfer/delivery and two sided model/dataset exchange signaling together.
The following is recommended for Rel-20 AIML CSI feedback enhancement: 
CSI report enhancement for inference, data collection and monitoring (RAN1, RAN2)
Quantization codebook specification, configuration and codebook exchange; 
Payload size, reporting and other configuration and reporting details (e.g. CQI/RI determination);
AIML PU occupation rule and timeline; 
If needed, high resolution codebook parameters, subject to UE capability; 
If needed, UE side monitoring (option2, precoded RS based) related enhancement; 
Note: This applies for all direction/sub-directions.
Model structure specification, taking RAN1 scalability study outcome into account; (RAN4, RAN1)
This applies for all directions/sub-directions.
Model parameter [/dataset] and other additional information exchange signaling for 3a-1 [and 4-1]; (RAN2, RAN1, SA2)
Conduct further study on OTA v.s. non-OTA;
Including other assistant information exchange, e.g., performance metric;
Necessary signaling to facilitate LCM operations specific to the CSI compression, including global ID (e.g., model/dataset/pairing ID etc.), applicability reporting etc.; (RAN2, RAN1, SA2)
Necessary specification on testability and requirement for CSI compression; (RAN4).
Note: both TDD and FDD need to be considered AI/ML based CSI feedback.

3GPP TSG-RAN WG1 Meeting #121         						 	R1-2503508
St Julian's, Malta, 19 - 23 May, 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.
Proposal 1: For network side data collection, support enhanced Rel-16 eTypeII codebook design to achieve high-resolution CSI for model training.
Proposal 2: 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 3: For NW to collect data for training, enhancement on the design of CSI-RS is not needed.
Proposal 4: 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 5: 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.
Proposal 6: Normative work for AI-based CSI compression is supported at least with the following scope consideration:
Spatial/temporal/frequency compression with case 2,
Option 3a-1 and option 4-1 of direction A and direction C for inter-vendor training collaboration,
NW side monitoring based on the ground truth CSI reported by UE and UE side monitoring based on the recovery CSI identified by NW.

3GPP TSG RAN1 WG #121		                         R1-2503553
St. Julian’s, Malta, May 19th– 23th, 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 2 shows moderate UPT gain over Benchmark 1 (Rel-16 eType II) and smaller UPT gain over Benchmark 2 (Case 0 CSI compression). 
Issues from non-self-contained CSI report (dependency on past reports) may further diminish the gain and usability.
Higher gain observed for smaller payload sizes and gain decreases as payload size increases. 

Observation#2 Case 3 shows moderate UPT gain over the benchmark based on non-AI/ML based CSI prediction and Rel-18 DD codebook based reporting 
Higher gain is observed for cell-edge UEs. 

Observation#3: Based on Direction C and Direction A of inter-vendor training collaboration options, at least three deployment alternatives identified
Alt1: UE side directly deploys the specified reference encoder
UE-side may still adapt the encoder to its implementation using the specified reference decoder and dataset used to specify reference model. 
Alt2: UE side trains its encoder based on reference decoder and field data collected by the UE-side.
For field data collection, indication for consistency on NW-side additional conditions 
Alt3: UE side trains its encoder based on parameter/dataset shared by the network, additional field data collected by the UE-side 
For field data collection, indication for consistency on NW-side additional conditions 
For all alternatives, the network may train it decoder based on field data and/or specified reference encoder. 
Observation#4: Among the deployment alternatives of Direction C and Direction A in Observation#3, deployment Alt1 is the widely deployable alternative. 


Observation#5: 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#1: For Option 4-1 of Direction A, consider NW-side sharing the encoder backbone assumption associated with the dataset as additional information. 

Proposal#2: Conclude the Rel-19 study on two-sided model based CSI compression with recommendation on the following aspects for normative work in Rel-20
Support at least Direction C from the inter-vendor training collaboration options
CSI compression sub use Case 0 as first priority 
CSI compression sub use Case 3 as second priority 
Necessary specification support for measurement and report for inference, network-side and UE-side data collection, network-side and UE-side monitoring, timeline and processing unit
Pairing method to address consistency between training and inference regarding network-side additional conditions. 

3GPP TSG RAN WG1 Meeting #121                                                                             R1-2503628
Malta, Malta, May 19th – 23rd, 2025
Source:	TCL
Title:	Discussion  on CSI compression and other aspects on Al/ML air interface	
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: Study the triggering mechanism based on UCI loss event for CSI information reset or retransmission.
Option 1: UCI loss event is determined by NW
Option 2: UCI loss event is determined by UE


Proposal 4: Investigate the CSI information synchronization for mitigating the impact of UCI loss.
Option 1: Reset the CSI information to pre-defined values.
Option 2: NW configs or UE initials CSI information retransmission.
Proposal 5: 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 6: 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 7: 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 8: 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 9: Option 1 and 2 should not be standardized, at least they are out of the scope of R19.
Proposal 10: 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 11: Ran1 should consider whether/how to transmit the raw CSI for monitoring purpose.
FFS: The overhead reduction scheme based on spectral or temporal processing.
Proposal 12: The paring ID can be composed of multiple parts to indicate the model structure and scalar quantization trained at the network side.
Proposal 13: The paring ID can be derived by an upper-level ID.

3GPP TSG RAN WG1 Meeting #121	R1-2503651
St Julian’s, Malta, May 19th – 23th, 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:
Recommendation from RAN1
Observation 1: For Direction A Option 4-1, the overhead and resource consumption for the exchange of the {target CSI, CSI feedback} dataset would be much larger than that of the encoder parameter for Option 3a-1, and the feasibility of dataset sharing acorss UE vendors have not been well studied by other working groups.
Regarding the inter-vendor training collaboration options, prioritize Direction A Option 3a-1 over Direction A Option 4-1 for the normative work.
Observation 2: Temporal domain CSI compression Case 2 would lead to model performance degradation due to UCI missing, partial CSI omission, or aperiodic CSI triggering, and significantly increase the model complexity due to the extra module for time correlation extraction (e.g., LSTM).
Regarding the temporal domain aspects of AI/ML-based CSI compression using two-sided model, prioritize Case 3 over Case 2 for the normative work.
 Regarding the potential specification impacts of AI/ML-based CSI compression using two-sided model, considering the followings for the normative work: 
Data collection, e.g., ground-truth CSI reporting with higher resolution, quality based data omission
Model inference, e.g., enhancements on CSI report framework, CSI processing criteria
Performance monitoring, e.g., performance metrics, monitoring report mechanisms
Other issues, e.g., model pairing, APU occupation, model loading/processing timeline

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 3: 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 4: 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 5: 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 6: 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 7: 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.
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.
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 9: 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 10: 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 11: 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 12: 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

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 12: 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 #121			R1-2503710
Malta, MT, May 19th – 23rd, 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 Options 3a-1 and 4-1 in Direction A, it is proposed that a performance target may be shared as additional information along with the exchanged dataset or model parameters. Support is proposed for Option 1 regarding the format of the performance target:
Option 1: Average performance target, e.g., average SGCS and/or average NMSE computed over the evaluation dataset.

Proposal 10: For inter-vendor collaboration, in direction B, use a common encoder across all UEs and standardize model parameters and pre-processing. 
Proposal 11: For inter-vendor collaboration, consider standardizing a model structure with configurable hyperparameters to adapt effectively to different deployment scenarios.
Proposal 12: For inter-vendor collaboration, for Direction C, consider standardizing both the encoder and decoder (1-3).
Proposal 13: For Standardizing Scalable Model Structure, Consider Training and Specifying Parameters for One or a Few Combinations.
Proposal 14: 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 15: In Direction C, an ID is required for monitoring purposes. the same ID used for UE-side data collection may be reused for monitoring configuration, to maintain consistency and simplify association with the reference model.
Proposal 16: Consider enabling the network to configure the UE to report ground truth CSI via L1 signalling or RRC.
Proposal 17: 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 18: 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 19: 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 20: 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 21: 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 22: 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 23: 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 #121		                                    R1-2503728
St Julian, Malta, May 19th – 23rd, 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 #121		R1-2503756
Malta, May 19th – May 23rd, 2024

Agenda Item:	9.1.4.1
Source:	Indian Institute of Technology Madras (IITM), IIT Kanpur
Title:	Discussion on Additional Study for AI/ML CSI Compression
Document for:	Discussion

Conclusion
In this contribution, we have provided our comments and discussion details on various aspects of inter-vendor collaboration along with simulations for comparing performance-complexity trade-off between different AI/ML models. Further discussion should consider the proposals and observations that we have provided above.
4 

3GPP TSG RAN WG1 #121		                             R1-2503773
St Julian’s, Malta, May 19th – 23rd, 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:
Proposal 1: Concluded in RAN1 that AI/ML based CSI compression is feasible and recommending AI/ML based CSI compression into Rel-20 WI.
Proposal 2: For AI/ML based CSI compression, only support Case 0 (SF domain) for normative work in Rel-20.
Proposal 3: For AI/ML based CSI compression, support inter-vendor training collaboration Option 3a-1 for normative work in Rel-20.
Proposal 4: For AI/ML-based CSI compression, conclude that the following specification impacts which are irrelevant to inter-vendor training collaboration directions/options have been identified:
LCM related signaling for two-sided model
Model/dataset pairing during data collection, training, inference and performance monitoring
Applicability reporting
Data collection
NW-side data collection
UE-side data collection
Inference and CSI reporting
Measurement resource configuration
AI/ML CSI report configuration and report content
Timeline, CSI processing unit and/or AI/ML processing unit occupation, and resource counting/memory
Performance monitoring
NW-side performance monitoring
UE-side performance monitoring
Testability, validation and requirements
Proposal 5: For AI/ML-based CSI compression, conclude that the following specification impacts which are relevant to inter-vendor training collaboration directions/options have been identified:
For Direction A Option 3a-1,
Model structure specification (including scalability) 
Model parameter exchange framework and format specification
Additional information specification: performance target, quantization configurations
For Direction A option 4-1,
Dataset exchange framework and format specification
Additional information specification: performance target, backbone/hyper parameter information, scalability configuration, quantization configurations
For Direction C,
Reference model (model structure+ model parameters) specification (including scalability)
Additional information specification: quantization codebook
Proposal 6: For AI/ML based CSI compression, reference model specification in inter-vendor training collaboration Direction C can adopt the model structure specification result in Direction A Option 3a-1 and specify additionally the model parameters.
Proposal 7: Provide specification support for AI/ML-based CSI compression in 5G-A Rel-20, encompassing:
Spatial-frequency domain CSI compression (Case 0)
Inter-vendor training collaboration with specified scalable model structure and exchanged model parameters (Direction A Option 3a-1) 
Specify necessary signalling/mechanism to facilitate AI/ML-based CSI compression
FFS the support of inter-vendor training collaboration with fully specified reference model (Direction C) and inter-vendor training collaboration with specified data/dataset format and exchange dataset (Direction A option 4-1)
Proposal 8: 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 9: 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 10: 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 11: 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 12: 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 13: 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.
Observation 1: For AI/ML CSI compression, one specification impact regardless of the inter-vendor training collaboration directions and options is the LCM related signaling for two-sided model, which includes:
Model/dataset pairing during data collection, training, inference and performance monitoring
Applicability reporting
Observation 2: For AI/ML CSI compression, specification of either model structure for inter-vendor training collaboration Direction A Option 3a-1 or reference model for Direction C does not require field data collection. Field data collection is needed for the training phase of inter-vendor training collaboration direction A Option 3a-1 and 4-1, with the following data collection schemes:
NW-side data collection
UE-side data collection
Observation 3: For AI/ML CSI compression, specification impact of training includes:
For Direction A Option 3a-1,
Model structure specification (including scalability) 
Model parameter exchange framework and format specification
Additional information specification: performance target, quantization configurations
For Direction A option 4-1,
Dataset exchange framework and format specification
Additional information specification: performance target, backbone/hyper parameter information, scalability configuration, quantization configurations
For Direction C,
Reference model (model structure+ model parameters) specification (including scalability)
Additional information specification: quantization codebook
Observation 4: For AI/ML CSI compression, one specification impact regardless of the inter-vendor training collaboration directions and options is the inference and CSI reporting, which includes:
Measurement resource configuration
AI/ML CSI report configuration and report content
Timeline, CSI processing unit and/or AI/ML processing unit occupation, and resource counting/memory
Observation 5: For AI/ML CSI compression, one specification impact regardless of the inter-vendor training collaboration directions and options is the performance monitoring, which includes:
NW-side performance monitoring
UE-side performance monitoring
Observation 6: For AI/ML CSI compression, one specification impact regardless of the inter-vendor training collaboration directions and options is the testability, validation and requirements, where RAN4 is primarily involved.
Observation 7: The scalable structure for SF domain CSI compression has obvious advantage in performance-complexity trade-off over the dedicated structure.
Observation 8: 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 9: Regarding the scalable structure for SF domain CSI compression, hyper-parameter config 2 has the best performance-complexity trade-off.
Observation 10: There are many state-of-the-art techniques in Machine Learning for complexity reduction that can be applied to AI/ML based CSI compression. These approaches are implementation specific and have little specification impact.
Observation 11: The performance variation of Temporal Domain aspects Case 2 is more complex than in Case 0. The calculation of performance metrics and the RS configuration for obtaining ground truth need to take into account factors such as UCI loss and the method of using historical CSI, which can cause instability in performance across multiple inferences.
Observation 12: 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.

3GPP TSG RAN WG1 #121			R1-2503823
St Julian’s, Malta, May 19th – 23rd, 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 are made:
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.

And we have the following proposals:
Proposal 1: 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 2: For inter-vendor-collaboration Options 3a-1 and 4-1 in Direction A, for the type of performance metric, prioritize SGCS.
Proposal 3: For UE-side monitoring, it is proposed to deprioritize UE side monitoring with proxy model / direct estimator of KPI / reference decoder.
Proposal 4: 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 5: 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.
Proposal 6: AI/ML based CSI compression using two-sided model is recommended into normative work from RAN 1 perspective:
Inter-vendor collaboration Direction C and Direction A with sub-option 4-1, 3a-1 with and without target CSI sharing is considered

3GPP TSG RAN WG1 #121                                                                                                  R1-2503875
St Julian’s, Malta, May 19th – 23rd, 2025

Agenda item:	9.1.4.1
Source:	            		Xiaomi
Title:	Further discussion on remained issues for AI/ML model based CSI compression
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:
Proposal 1: Support Case 2 and Case 3 on top of Case 0 to study as a normative work.
Proposal 2: Study and specify signalling and configuration on data collection for model training or monitoring.
Proposal 3: 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 4: 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.
Proposal 5: The three options on inter-vendor training collaboration, i.e., Option 1, Option 3a-1 and Option 4-1 could be further down selected in normative work if possible.
Proposal 6: Specify the signalling for dataset exchanged from NW side to UE side based inter-vendor collaboration training (Option 4-1).
Proposal 7: Specify model structure of Case 2 and Case 3 in normative work if Option 3a-1 or Option 1 for inter-vendor collaboration is supported. 
Proposal 8: Specify model identification for two-side AI/ML model in normative work.
Proposal 9: Specify signalling or mechanism for UE-side or NW-side AI/ML model performance monitoring. 
Proposal 10: Support CQI calculated by using two stage approach.
Proposal 11: Recommend two-side AI/ML model based CSI compression to study as a normative work in Rel-20, and the following scopes should be studied and specified if necessary.
Specify necessary signalling/mechanism to facilitate LCM of Case 2 and Case 3, e.g., data collection, performance monitoring
Study and specify signalling for dataset exchanged based inter-vendor collaboration (i.e., Option 4-1)
Study and specify signalling/mechanism for model identification
Study, and if necessary, specify CSI measurement and/or CSI reporting 

3GPP TSG RAN WG1 #121	R1-2503923
St Julian’s, Malta, May 19th – 23th, 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 necessary specification impacts and recommendations. Specifically, we have the following proposals:
Data collection:
Proposal 1: 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).
L1 signaling (e.g., the existing CSI report framework) can be used to report ground-truth CSI.
Model inference:
Proposal 2: For model inference, RAN1 to discuss potential spec impact about the following:
New report format for compressed bits.
New CSI processing, priority/omission rule.
Address the issue about misalignment of historical CSI information at NW side and historical CSI information at UE side caused by UCI loss or rank adaption.
Address the issue about outdated historical CSI information.
Performance monitoring:
Proposal 3: For NW-side monitoring (based on the target CSI reported by UE), RAN1 to discuss whether the compressed CSI and corresponding target CSI are reported in a same report, or separate reports.
Proposal 4: For UE-side monitoring (e.g., based on the output of the CSI reconstruction model at the UE), RAN1 to study potential specification impact about definition, calculation and reporting of performance metric.
Capability and applicability reporting:
Proposal 5: Study the capability and applicability reporting of UE-side model for Case 3.
Model identification:
Proposal 6: 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.
Model transfer:
Proposal 7: Support Alt. B for model transfer methodology z4. 
Proposal 8: 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 9: 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.
Recommendations:
Proposal 10: RAN1 to recommend normative work in Rel-20 on Case 2 and Case 3. And Case 0 can be regarded as a special case of Case 2.

3GPP TSG-RAN WG1 #121	R1-2503977
St Julian's, Malta, 19 - 23 May 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:	For UE-side monitoring, both time-based trigger and event-based trigger needs to be studied for reporting the monitoring metrics. 
Observation 2:	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 3:	Out-of-distribution metrics for UE-side monitoring can be another candidate as monitoring metric but further study is necessary. 
Observation 4:	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 5:	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 6: 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 7: NW-side monitoring with lower signaling overhead options can be studied further in Rel-19 as alternative to the UE-side monitoring. 
Observation 8: It may be beneficial to use AI/ML Processing Units (PU) to indicate the number of supported simultaneous AI/ML functionalities.
Observation 9:  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 10:  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 11:  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 12: 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 13: 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.
Observation 14: Option 3a-1 with Target CSI sharing has a comparable overhead to Option 4-1 and thus both options have the same challenge regarding the signalling overhead.
Observation 15: Temporal CSI compression Case 2 may have a significant specification impact and additional complexity due to methods for mitigating the impact of UCI loss. 
Observation 16: Both temporal CSI compression Case 2 and Case 3 may have a significant specification impact due to inter-vendor training collaboration.

Proposal 1:	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 2: Support PU pool sharing by different AI/ML functionalities (e.g., AI/ML based CSI reporting, AI/ML based BM, AI/ML based POS).
Proposal 3: Support AI/ML Processing Units (PU) occupancy definition per AI/ML functionality (e.g., AI/ML based CSI reporting, AI/ML based BM, AI/ML based POS).
Proposal 4: For active models, the inference timeline needs to account for the AI/ML model readiness status. 
Proposal 5: Inter-vendor collaboration Option 3a-1 without Target CSI should not be considered.
Proposal 6: Inter-vendor training collaboration, only one option (i.e., Option 4-1) will be considered if it proceeds to normative work.
Proposal 7: For CSI compression, only one case (i.e., Case 0) will be considered if it proceeds to normative work. Temporal CSI compression Case 2/Case 3 could be considered as an enhancement in a later release if needed.
Proposal 8: AI/ML-based CSI compression is not recommended for normative work. 

3GPP TSG RAN WG1 #121   	                              R1-2503983
St Julian’s, Malta, May 19th – 23st, 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 the performance-complexity trade-off and further potential aspects including inter-vendor training collaboration for CSI compression with two-sided AI/ML model. The following observations and proposals are provided.

Observation #1: Temporal domain Case 2 slightly outperforms Case 0 across all payload sizes at the expense of increased computational complexity. This demonstrates that additional temporal information provides more effectiveness for model learning.
Observation #2: Increasing model complexity does not lead to proportional improvements in SGCS gain. As the model becomes more complex, the rate of SGCS gain improvement gradually decreases. This shows a saturation effect where model complexity plateaus beyond a certain capacity.
Observation #3:  Model (or hyper-parameter) selection strategy balancing of complexity and performance is required, as excessive model scaling may result in inefficient resource utilization without significant performance gains.
Observation #4: For performance monitoring of CSI compression, signaling overhead mainly originates from ground-truth CSI report, and/or reconstructed CSI delivery.

Proposal #1: Among temporal domain cases, consider to prioritize Case 0 over Case 2. (i.e., focus on temporal Case 0 in Rel-20), considering the performance-complexity trade-off.
Proposal #2: Consider NW-side monitoring approaches via direct estimation of intermediate KPI (e.g., SGCS) without ground-truth CSI reporting. 
Proposal #3: 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 #4: 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 #5: 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 #6: 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 #7: For temporal domain aspects Case 2, discuss the format of past CSI information and how to report it at least for performance monitoring perspective. 
Proposal #8: 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 #9: 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 #10: 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.
Proposal #11: Regarding inter-vendor training collaboration, recommend Direction A and Option 4-1 as the scope for normative work.

3GPP TSG RAN WG1 #121		R1-2503989
St Julian’s, Malta, May 19th – 23rd, 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
Observation 1: For Option 3a-1 without target CSI sharing, it is unclear whether UE-side development can be always aligned with NW side training. The model would be NW-side responsibility, and this is quite similar to just downloading the parameter to UE without offline engineering (i.e., Direction B).
Observation 2: For Option 3a-1 with target CSI sharing, the model responsibility is unclear since how much the parameters are adjusted by UE-side offline engineering is unknown to NW-side.
Observation 3: Option 4-1 may have less effort from NW perspective and as far as RAN4 test is available, no model backbone / structure related information sharing between NW-side and UE-side could be sufficient.

Section 3: Temporal domain aspects of CSI compression
Observation 4: 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 5: 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 1: Rank adaptation handling should be studied for handling rank > 1.

Section 4: Potential specification impact
Observation 6: For complexity comparison to study the scalability of rank > 1 solutions, the total complexity with multiple models should be taken into account.
Observation 7: 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 8: 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 9: 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 10: 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 11: 
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 12: 
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 Meeting #121	R1- 2503995
St Julian’s, Malta, May 19th – 23rd, 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
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.

3GPP TSG RAN WG1 #121		R1-2504027
St Julian’s, Malta, May 19th – 23th, 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 UE to report the APU occupancy rule for multiple functionalities based on the following options
Option 1: Dedicated APU pool for a functionality
Option 2: Shared APU pool for multiple functionalities
Option 2a: One inference takes 1 APU
Option 2b: Multiple inferences take 1 APU
Legacy CPU is also occupied with regard to the pre-processing and post-processing for inference
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 Direction B is deprioritized.
Proposal 14: For direction C, Option 2 (Latent adaptation) is preferred compared to Option 1 (Input/output adaptation with additional layers)
Proposal 15: 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 16: 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 17: 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 18: 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 #121	R1- 2504042      
Malta, MT, May 19 – 23, 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 specification of 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 of 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 specification of procedures/signaling enabling transmission of subset of CSI samples based on the experienced distortion level or quality indictor.
Proposal 4:	Support specification of 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 specification of 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, support procedures on   enabling it to  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.
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 7:	For option 4-1 and 3a-1 of Direction A, during the normative work, prioritize schemes based on first construction of the “local” decoder and then training of the encode model.
Proposal 8:	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 9:	During normative work, consider procedure/signaling enabling relating pairing information to 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 10:	When different models have been developed for different UE-NW vendor pairs, during normative work, consider procedure/signaling enabling model identification/selection, for inference, when the two sides does not have  the knowledge of who is the vendor of the other side.
Proposal 11:	Conclude that the local decoder model (or proxy model) that is trained using data shared in Direction A options 4 and 3a-1 with {target-CSI} and Direction C 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 12:	Conclude that UE-sided model monitoring can be used at least for inter-vendor collaboration options based on Direction A options 4, Direction A 3a-1 with {Target CSI} and Direction C.
Proposal 13:	Study mechanism for root-cause determination based on exchange of information regarding the NW-side trained encoder and/or decoder model.
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. Therefore, during normative work, it is desirable to consider procedure/signaling to determine the applicability of the encoder model for that inter-vendor collaboration option.
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:	Study mechanism for root-cause determination based on exchange of information regarding the NW-side trained encoder and/or decoder model.

References

3GPP TSG RAN WG1 Meeting #121	R1-2504083
St Julian’s, Malta, May 19th – 23rd, 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
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 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 or non-linear quantization if applied
Proposal 6:
For Direction C, test/validation dataset and the corresponding performance target should be shared from NW side to UE side.
Proposal 7:
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 could be considered as the performance metric. 
Performance and complexity trade-off
Observation 1:
The number of reference models for CSI compression could be significantly reduced by a scalability design.
Observation 2:
The model size of CSI generation part may be reduced by the retrain/redevelopment at UE side.
Observation 3:
For AI/ML-based CSI compression, the localized model (region-specific encoder and decoder) could improve the performance gain from 5% to 7.8% compared to the generalized model.

Proposal 8: 
For the issue of performance-complexity trade-off, it could be alleviated at least by the scalability design, the retrain/redevelopment of UE-sided encoder, localized models and new use cases (Case-2/3) studied in Rel-19.

Spec impact for two-sided mode
Observation 4:
For UE-side AI/ML model performance monitoring using a proxy model, the expectation of a simple structure and small size contradicts to the needs of a strong generalization capability for a proxy model to work well in various scenarios.
Observation 5:
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 6:
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 7:
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 9:
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 10:
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 11:
For CSI compression using two-sided AI/ML models, deprioritize Option 2 proposed for CQI determination.
Option 2: CQI is calculated based on the output of CSI reconstruction part from the realistic channel estimation.
Proposal 12:
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 13:
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 value corresponding to the precoding vectors.
Proposal 14:
For the performance monitoring of AI/ML-based CSI compression, RAN1 to prioritize the study of NW-sided monitoring.
Proposal 15:
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 16: 
For signaling support to mitigate the impact of UCI loss for Case-2, NW-triggered CSI retransmission should be deprioritized.
Proposal 17: 
For Case 2, RAN1 to study how to address layer ordering issue for training data collection, model inference and performance monitoring.

Recommendation for normative work
Proposal 18: 
For the issue of inter-vendor training collaboration, RAN1 to recommend Option 4-1 of Direction A, and Direction C for normative work.
Proposal 19: 
For the sub use cases of AI/ML-based CSI compression considering temporal domain aspects, RAN1 to recommend both Case 2 and Case 3 for normative work.

3GPP TSG RAN WG1 #121	R1-2504115
Malta, MT, May 19th – 23rd, 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:
Observation 1: 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 2: 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 1: Assignment of the pairing ID is controlled by the mobile network operator.
Observation 3: Whether a pairing ID is common to different configurations of a scalable model can be left as an implementation detail for the assigned of the ID as long as the number of possible ID’s is sufficient to cover unique pairing ID’s for each configuration of the feedback.
Proposal 2: Support allowing the pairing ID assigner the ability to assign unique or non-unique pairing ID’s across scalable model parameters as long as the combination of the pairing ID and other configuration parameters is sufficient to guarantee interoperability of the generation and reconstruction parts.
Observation 4: The correspondence between monitoring feedback and the pairing ID is known to the gNB since it has configured both the reporting of AI/ML-based CSI feedback and the monitoring feedback.
Proposal 3: Pairing ID is not explicitly required for performance monitoring configuration of the two-sided models for AI-ML CSI compression.Observation 5: It has been common practice in non-AI/ML CSI feedback codebooks to explicitly specify the NW-side mapping from feedback message bits to reconstructed CSI.
Observation 5: It has been common practice in non-AI/ML CSI feedback codebooks to explicitly specify the NW-side mapping from feedback message bits to reconstructed CSI.
Proposal 4: For quantization codebooks in AI/ML-enhanced CSI feedback, the information exchanged from NW-side to UE-side, including which codebook configuration to use and additional parameters defining the codebook of the specified configuration, should be sufficient to explicitly define the NW-side mapping from feedback message bits to latent vector.
Observation 6: Recurrent scalar quantization codebooks can facilitate achieving the performance benefits of Case 2 (encoder and decoder depend on past history) with low specification effort beyond Case 0.
Proposal 5: Include scalar quantization, vector quantization, and recurrent scalar quantization among the options for quantization codebooks to be considered in the work item phase.
Observation 7: A codebook for scalar quantization can be specified by a latent vector dimension, a number of bits per latent dimension, and a scalar codebook.
Observation 8: A codebook for vector quantization can be specified by a latent vector dimension, a number of bits per segment, a segment size, and a vector codebook table.
Observation 9: A codebook for scalar recurrent quantization can be specified by a latent vector dimension, a number of bits per latent dimension, a number of step size bits, a scalar stepsize codebook, parameters of a state machine, and a specification of the state machine operation.
Observation 10: When the NW-side operation of a quantization codebook is explicitly specified, it is not necessary to specify the UE-side operation to achieve interoperability.
Proposal 6: For normative work for AIML-assisted two-sided CSI compression, 3GPP needs to focus on specification of SF architecture/model (Case 0) with the 1st priority. Regarding Case 2 and Case 3 use cases, 3GPP needs to find a way to leverage the to-be-specified SF architecture as much as possible.
Observation 11: Temporal correlation, if any, between downlink channel instances should be observed at the output of SF encoder, i.e., on the latent space, as well. This temporal correlation at the latent space can be exploited for Case 2 or for Case 3 use cases without having to update SF encoder/decoder pair developed for Case 0.
Proposal 7: For normative work for AIML-assisted two-sided CSI compression, 3GPP needs to investigate methods to exploit temporal correlation at the latent space for efficient representation of the CSI feedback (Case 2) or for prediction of the future CSI (Case 3) as potential methods to adapt Case 0 SF encoder/decoder to Case 2 or 3 use cases.
Observation 12: Complexity of transformer-based CSI compression models can be significantly reduced without harming the performance using model pruning and knowledge distillation techniques.
Observation 13: The recurrent quantizer is a lightweight scheme that can be used instead of scalar/vector quantization to exploit time correlation of the CSI matrices to significantly increase the performance of CSI compression.
Observation 14: In Direction A, NW vendors do not have a meaningful way to generate a standalone encoder performance target.
Proposal 8: The performance target shared as additional information in Direction A should be an end-to-end performance target that compares the encoder input eigenvectors to the decoder output vectors, rather than an encoder performance target.
Observation 15: SGCS is a useful metric for end-to-end performance monitoring of CSI feedback methods that has been tested extensively in this RAN1 study on AI/ML-enhanced CSI feedback.
Proposal 9: The performance target shared as additional information in Direction A should be expressed using the SGCS metric.
Observation 16: The complexity and usefulness of providing performance targets for different configurations of a shared dataset or model parameters will depend on details of how scalability is addressed under Direction A, which have yet to be finalized.
Proposal 10: Revisit the question of multiple performance targets when the scalability approaches under Direction A are developed further.
Observation 17: Case 2-1 and Case 2-2 of the UE-side monitoring options do not incur significant increase of UL overhead, especially when event-based reporting is considered. This can relax UE’s reporting timeline requirements as well.
Observation 18: 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 11: Consider two-phase performance monitoring mechanism, i.e.,
Phase 1: UE-side monitoring via proxy decoder or direct intermediate KPI (e.g., SGCS) estimator for detection of performance degradation
Phase 2: NW-side monitoring via ground-truth target CSI reporting for root cause analysis and/or for validation of UE-side monitoring reports
Phase 2 is to be triggered by NW based on observation of the Phase 1 operation.

Observation 19: Aspects of the calculations required for AI/ML-based CSI compression are likely to consume computational resources used for legacy tasks as well as AI/ML-specific resources.
Proposal 12: RAN1 to consider the advantages and disadvantages of assigning both CPU’s and AI/ML PU’s to CSI compression according to the use of legacy and AI/ML-based computational resources, respectively.

3GPP TSG RAN WG1 #121			R1-2504132
St Julian’s, Malta, May 19th – 23th, 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 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 2: 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 3: 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 4: 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 5: 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 6: 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 7: 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 8: 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 9: 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.

Proposal 10: For AI/ML-based CSI reporting, consider an option that both dedicated AI/ML PU and legacy CPU are occupied.

Proposal 11: For AI/ML-based CSI reporting, the occupancy timeline of the AI/ML PU can be modeled differently from the CPU, while the CPU occupancy timeline can follow the legacy CPU occupancy.

Proposal 12: For AI/ML-based CSI reporting, the CSI reporting requires the model to be preloaded at the UE.

Proposal 13: To enable efficient AI/ML-based CSI reporting operation, it is necessary to study mechanisms that allow the gNB to predict or be informed of the model’s loading status at the UE.

Proposal 14: Recommend the CSI compression use case using two-sided AI/ML models as a normative work in Release-20.


Observation 1: 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 2: 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 3: Alternative 3 for scalability over payload configuration provides a feasible approach to achieving a scalable model structure, as an AI/ML model with a single trained parameter set and separate quantization parameters and codebooks for each payload configuration demonstrates good performance across various payload configurations. Moreover, it outperforms dedicated AI/ML models across the payload configurations.

Observation 4: 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 Meeting #121                                            	 R1-2504226
St Julian’s, Malta, May 19th-23rd, 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 results on the complexity-performance trade-off and recommendations as well as the scope of the normative work. Observations and proposals are summarized as follows:
Observation 1: Regarding the encoder complexity and performance trade-off for Case 0, the SGCS gain is 3.5%, 8.4% and 10.7% with encoder FLOPs <10M, 10M~100M and >100M.
Proposal 1: Potential spec impacts related to training stage:
Data collection:
Signaling to trigger NW-side and/or UE-side data collection
CSI-RS resource configuration
Mechanism, content and format of data reporting
Inter-vendor collaboration:
Signaling to facilitate information sharing from NW-side to UE-side
Information includes dataset/parameters, and possible additional information
Specified dataset/parameter format
Pairing mechanisms between NW-side and UE-side
Proposal 2: Potential spec impacts related to inference stage:
Signaling/mechanism to:
Trigger the AI/ML model inference
Pairing between NW-side and UE-side
Indicate the CSI feedback payload
CSI report:
CSI-RS resource configure
CSI reporting configure
CSI mapping 
CSI priority rule
Legacy CSI processing unit and/or AI/ML processing unit
Proposal 3: Potential spec impacts related to monitoring stage:
NW-side monitoring
Signaling/mechanism to trigger the NW-side monitoring
Mechanisms, content and format of target CSI reporting
CSI report 
CSI report configure
CSI mapping 
CSI priority rule
UE-side model monitoring
Signaling/mechanism to trigger the UE-side monitoring
Mechanisms, content and format of performance monitoring results reporting
Legacy CSI processing unit and/or AI/ML processing unit
Proposal 4: Recommend AI/ML based CSI compression into normative work in Rel-20. 
Any new CSI-related use case should be precluded
Proposal 5: Regarding the temporal domain Case 0, Case 2 and Case 3, support to specify common issue in higher priority and non-common issue in lower priority.
Proposal 6: Regarding Direction A and Direction C, only support sub-option 4-1 in Rel-20 5G-A WI phase.
Proposal 7: Regarding the performance monitoring for AI/ML based CSI compression, support 
UE-side monitoring Case 2-1
NW-side monitoring
and possible combination of UE-side monitoring Case 2-1 and NW-side monitoring

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 offline engineering in direction A, discuss allowable operations and changes in performance of AI/ML models.
For Direction A, discuss the boundary of AI/ML model (where the model starts and ends) and determine which entity is responsible for providing quantization codebook.
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.
Support separate discussions of PU for different AI/ML use cases (e.g., CSI, BM/POS) to adequately address their distinct requirements.
For NW-side AI/ML model training, NW can rely on UL CSI samples collected from SRS sent by UEs.
 Conclude RAN1 has not observed significant improvement in performance-complexity trade-off during R19 studies.
 Conclude RAN1 has provided feasible solutions to address complication of training two-sided AI/ML models.

3GPP TSG RAN WG1 #121                                                              R1- 2504311	
St Julian’s, Malta, May 19th – 23th, 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 the remaining aspects of CSI compression, and APU processing timeline across different use cases. Based on the discussion, the following proposals have been proposed.

Proposal 1: For inter-vendor training collaboration under direction C and direction A, option 3a-1, the reference model structure and parameters reuse RAN4 specified model and parameters.   
 
Proposal 2: Recommend RAN1 to specify 4-1 with non-OTA dataset delivery, with RAN4 specified model structure and parameters for Direction C.   

Observation 1: 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 3: 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. 

Proposal 4: For case 2, time-frequency-spatial domain CSI compression, for RI, consider longer term RI update across different CSI reports.   

Proposal 5: For time-frequency-spatial domain CSI compression, flexible CSI report configuration to support different cases should be studied.  

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.

Observation 2: For spatial-frequency domain CSI compression (case 0), by adjusting model hyper-parameter, lower FLOPs can be achieved with minor performance degradation.      

3GPP TSG RAN WG1 #121                                                                                                    R1-2504387
Malta, MT, May 19th – 23rd, 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 the summary of AI/ML-based CSI compression using two-sided model. We have the following observations:
Observation 1: Study of inter-vendor collaboration is complete from RAN1 perspective. Almost all issues of Direction A and C are addressed expect the following: The signaling feasibility needs RAN3/SA2/SA5 input; feasibility of model specification for performance testing and interoperability needs RAN4 conclusion. The details of data format, parameter format, performance target, ID assignment and quantization configuration can be discussed in normative work.
Observation 2: The study of performance monitoring is complete. NW side monitoring is necessary and there is no strong motivation to recommend UE side monitoring given that it needs NW side monitoring as prerequisite.
Observation 3: Complexity reduction for Case 0 is feasible with minor performance degradation. Case 2 and 3 provide improved performance with similar or slightly higher complexity compared to Case 0. The study of inter-vendor collaboration, model monitoring in R18 focuses on Case 0.
Observation 4: Sub-option 3a-1 shares similar specification impact as sub-option 4-1 in terms of standardized signaling for information exchange. Sub-option 3a-1 shares similar specification impact as Direction C in terms of model structure specification, and they both can reuse the model to be defined / specified in RAN4.
Observation 5: As long as standardized signalling is deemed feasible, any lack of agreed solution(s) for the standardized signaling should not be a blocking factor for the feasibility of dataset / model parameter exchange; vendors can still rely on proprietary inter-vendor collaboration until standardized signaling is introduced in the later releases.
Observation 6: Angle-delay approach may not be suitable for small payload regime considering that SD/FD bases selection would need 20+ bits.
Observation 7: Minor performance degradation (-3%) is observed by reducing the model size 3 times (from 300k + parameter to 100k) for spatial-frequency domain.
Observation 8: 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 9: Under similar model size, complexity reduction via NN architecture optimization yields less performance degradation than angle-delay domain approach.
Observation 10: For the same complexity (model-size / FLOPs count), case 2 provides higher SGCS gain compared to case 0.
Observation 11: 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.
Based on the observations, we propose
Proposal 1: Recommend 3a-1, 4-1 of Direction A and Direction C to normative work from RAN1 perspective.
Proposal 2: Recommend NW side monitoring to normative work.
Proposal 3: Recommend both NW side and UE side data collection for training to normative work.
Proposal 4: Recommend Case 0 to normative work as first priority, Case 2 and 3 as second priority.
Proposal 5: Capture the following conclusion for sub-option 3a-1:
RAN1 assumes that the model structure of the sub-option 3a-1 can be equivalent to the structure of the to-be-specified / defined testing / reference model, if it is specified / defined and the specified / defined testing / reference model includes the encoder, for performance requirements in RAN4.
Proposal 6: Following specification impacts are identified for inter-vendor collaboration directions:
Reference model (structure) specification
Note: It is assumed to be equivalent to RAN4 model if RAN4 defines a testing model and / or a reference encoder model / structure
Note: It is assumed that the defined / specified model in RAN4 may consider RAN1 study outcome of model structure and scalability
Data / model parameter / additional information (performance target, quantization codebook, ID) exchange format 
Proposal 7: Endorse the text proposal in section 2.3 for the summary of the study item of CSI feedback via two-sided model.
Proposal 8: Complexity reduction is possible in the Transformer-based design with spatial-frequency domain input.
Proposal 9: 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 10: Capture observation 7-11 in the TR.

3GPP TSG RAN WG1 #121			R1-2504494
St Julian’s, Malta, May 19th – 23rd, 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 reduce the UL report occasions, which benefits NW’s UL transmission performance.
Proposal 1
Deprioritize the approach that uses CSI retransmission for mitigating the impact of UCI loss.
Proposal 2
If the AI/ML-based CSI compression is supported for the normative work,
RAN1 recommends the temporal domain Case 3 (Case 0 can be treated as a special case of Case 3).

Proposal 3
If the AI/ML-based CSI compression is supported for the normative work,
RAN1 recommends at least Direction C Option 1 and Direction A Option 3a-1 for the normative work.
It is up to RAN4 to define reference model(s), which are based on the model structure of Option 3a-1, as RAN1 Direction C reference model(s).
Proposal 4
Defer the discussions of AI/ML PU and timeline for CSI compression to future normative work.

3GPP TSG-RAN WG1 Meeting #121				                           R1-2504569
Malta, May 19th – 23rd, 2025

Source:	Continental Automotive
Title:                       Discussion on AI/ML CSI compression
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 the unified measurement of AI/ML processing capability (e.g., CSI compression, CSI prediction, etc.) and its reporting.
Proposal 6: Support the option that only dedicated AI/ML PU is occupied for inference of CSI compression.

3GPP TSG RAN WG1 #121		   R1-2504587
Malta, May 19th – May 23th, 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, full standardization model of Direction C can be studied base on the deployment of Direction A models.
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:  Define VQ codebook identification methods and lifecycle management (versioning, loading, expiration).
Proposal 5: Defined triggers and procedures for fallback to Scalar Quantization (SQ) to ensure reliable operation.
Proposal 6: It is proposed to establish:A publicly available dataset structured according to 38.901 logging standards, covering representative channel conditions and deployment scenarios.

Proposal 7:A standardized reference model encapsulated in a common exchange format, with clearly defined performance metrics (e.g., SGCS/NMSE targets) to support consistent benchmarking across implementations.
Proposal 8: It is recommended to use eType-II W2 as the baseline precoding matrix during simulation and evaluation phases.
Proposal 9: 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 10: 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.
Proposal 11：Upon internal detection of anomalies (e.g., data mismatches, model uncertainty increase), the UE can autonomously trigger enhanced monitoring, reporting detailed metrics (e.g., SGCS, NMSE) for a short window, optionally informing the network via a monitoring flag.
Proposal 12：The network may select active encoder revisions via signaling or instruct rollback via dedicated MAC CE (e.g., ModelRollbackReq) based on performance evaluations (e.g., SGCS/NMSE/KPI degradation).
Proposal 13：Encourage fallback to Case 2-like compression if predicted CSI deviates significantly from measured CSI during inference.
Proposal 14：Validation failure within configured observation window after model activation.
Proposal 15：Define a finetuneTriggerLevel based on model degradation indicators, such as:
SGCS or NMSE degradation exceeding a threshold δ;
UCI loss ratio above predefined level (e.g., >10%);
Proposal 16：Evaluation window length (L) and pass/fail thresholds to validate retraining success.

3GPP TSG RAN WG1 Meeting #121						  R1-2504624
St Julian’s, Malta, May 19th – 23rd, 2025

Agenda Item:	9.1.4.1
Source:	IIT KANPUR, Indian Institute of Tech (M)
Title:	Discussion on AI/ML based CSI compression
Document for: Discussion and Decision

Conclusions
In this contribution, we presented our perspective on inter-vendor collaboration under Direction A, specifically discussing performance target sharing and ID-based model pairing for Options 3a-1 and 4-1. 
Our observations and proposals can be summarized as follows:
Observation 1: A performance target computed from the same data that trains the nominal encoder, and the actual decoder naturally reflects both the prevailing channel statistics and the decoder’s implementation losses. By calibrating against this “current” target figure, the UE optimises for the exact conditions it will encounter during transmission of CSI codewords, while the NW can apply the value as a reliable threshold for current CSI reports.
Observation 2: The interval with which the NW recomputes and transmits the target should depend on mobility: slower updates may be enough in low-mobility scenarios where fading is slow, whereas high-mobility scenarios may require recalculation faster to remain representative. Selecting the target is a trade-off: shorter intervals improve the accuracy of pass/fail decisions but increase signalling and decoder training computation, whereas longer intervals conserve resources at the cost of responsiveness. Choosing the performance target transmission interval, trades signalling and computing cost against the accuracy of switching decisions between AI-based and legacy feedback.
Observation 3: SGCS maps directly to beamforming gain and may reach its target earlier, during training the model, than NMSE. Using SGCS as the shared performance benchmark can therefore cut UE training time and computation load while still delivering near-optimal precoding. Any incremental benefit from prolonging training to improve NMSE remains to be quantified. 
Observation 4: The model ID for pairing can be generated with established network-ID generation procedures and its assignment can follow existing network ID reporting configuration field, minimising additional specification impact.

Proposal 1: RAN1 should further study the two parameters, SGCS and NMSE, interms of training convergence speed and performance of precoder, before deciding upon which to choose as performance target metric.

Proposal 2: Further details of indicating this model/dataset-paired ID using legacy configuration procedures can be studied.

	
3GPP TSG RAN WG1 #121                                                                                            R1-25xxxxx
Malta, MT, May 19th – 23rd, 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
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.

	
3GPP TSG RAN WG1 #121                                                                                            R1-25xxxxx
Malta, MT, May 19th – 23rd, 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
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.

	
3GPP TSG RAN WG1 #121                                                                                            R1-25xxxxx
Malta, MT, May 19th – 23rd, 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
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.

Agreements from RAN1 #121
Observation
The model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 can be same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.
	To reduce the workload of the potential normative work scope, the model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 is(are) same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.

	
3GPP TSG RAN WG1 #121                                                                                            R1-25xxxxx
Malta, MT, May 19th – 23rd, 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
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.

Agreements from RAN1 #121
Observation
The model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 can be same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.
	To reduce the workload of the potential normative work scope, the model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 is(are) same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.

	
3GPP TSG RAN WG1 #121                                                                                            R1-25xxxxx
Malta, MT, May 19th – 23rd, 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
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.

Agreements from RAN1 #121
Observation
The model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 can be same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.
	To reduce the workload of the potential normative work scope, the model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 is(are) same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.

Agreement
For addressing inter-vendor collaboration complexity, RAN1 identifies the following specification impacts for supporting sub-option 3a-1 (including with target CSI and without target CSI), sub-option 4-1, and Direction C 
Reference model / structure specification (for sub-option 3a-1 and Direction C)
Note: The so called fully specified reference model discussed in RAN1 for Direction C is assumed to be equivalent to the to-be-specified model, if specified, for performance requirements in RAN4.
Note: The model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 can be same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.
To reduce the workload of the potential normative work scope, the model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 is(are) same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.
Note: RAN1 has studied and agreed on a scalable model structure for feasibility study during Rel-19 study, and considers that can serve as a starting point for the model structure specification for the potential normative work.
Format and contents of parameter/dataset/information exchange: dataset (for 4-1 and 3a-1 with target CSI)/ encoder parameter (for 3a-1) / additional information (for 3a-1 and 4-1).
The workload can be reduced by identifying the common part among different aspects.


Observation
Case 0, encoder complexity (FLOPs) vs. SGCS gain (%) 




For temporal domain Case 0, the results on trade-off between complexity (FLOPs) and SGCS gain (%) over benchmark are summarized as follows
For using spatial-frequency domain eigen-vector as input,
For layer 1 and payload size X-bin:
9 sources [QC, CATT, vivo, Futurewei, OPPO, LG, Fujitsu, IITM, PengCheng Lab] observe minor to significant performance gain of 2.8%~20.68% over benchmark for FLOPs range of 1M to 10M.
11 sources [QC, CATT, vivo, Futurewei, Ericsson, Samsung, OPPO, LG, Fujitsu, PengCheng Lab, Spreadtrum] observes minor to significant performance gain of 1.2% ~21.7% over benchmark for FLOPs range of 10M to 100M.
5 sources [vivo, Huawei, Xiaomi, Spreadtrum, Nokia, OPPO] observe minor to significant performance gain of 4.8%~13.5% over benchmark for FLOPs > 100M
For layer 1 and payload size Y-bin:
3 sources [CATT, LG, Vivo] observe minor to significant performance gain of -1.8%~10.45% over benchmark for FLOPs range of 1M to 10M.
4 sources [CATT, LG, Vivo, Nokia] observe minor to significant performance gain of 0.4%~11.45% over benchmark for FLOPs range of 10M to 100M.
2 sources [Xiaomi, Nokia] observe minor to moderate performance gain of 1.63%~6.1% over benchmark for FLOPs > 100M
For layer 1 and payload size Z-bin:
2 sources [CATT, LG] observe minor to significant performance gain of -3.32%~10.62% over benchmark for FLOPs range of 1M to 10M.
2 sources [CATT, LG] observe minor to significant performance gain of 1.2%~11.99% over benchmark for FLOPs range of 10M to 100M.
2 sources [Xiaomi, Nokia] observe minor performance gain of -2.05%~2.1% over benchmark for FLOPs > 100M
For using angle-delay domain eigen-vector as input
For layer 1 and payload size X-bin:
2 sources [Ericsson, Samsung] observe minor to significant performance gain of 2.8%~19.29% over benchmark when FLOPs < 1M.
3 sources [Ericsson, QC, Vivo] observe minor to moderate performance gain of 1.4%~6.63% over benchmark for FLOPs range from 1M to 10M.
2 sources [Ericsson, QC] observe moderate performance gain of 5.3%~6.8% over benchmark for FLOPs > 10M.
For using spatial-frequency domain channel matrix as input
For layer 1 and payload size X-bin:
1 source [Huawei] observes 37.8% performance gain over benchmark when FLOPs is 100M.

Case 0, encoder model size (# parameters) vs. SGCS gain (%) 




For temporal domain Case 0, the results on trade-off between model size (# parameters) and SGCS gain (%) over benchmark are summarized as follows 
For using spatial-frequency domain eigen-vector as input,
For layer 1 and payload size X-bin:
10 sources [QC, CATT, vivo, Futurewei, Ericsson, OPPO, LG, Fujitsu, PengCheng Lab, IITM] observe minor to significant performance gain of 1.2%~18.5% over benchmark for model size less than 1M parameters.
12 sources [CATT, vivo, Futurewei, Ericsson, Xiaomi, OPPO, LG, Fujitsu, PengCheng Lab, IITM, Spreadtrum, Nokia] observe minor to significant performance gain of 1.3% ~21.7% over benchmark for model size in range of 1M to 10M parameters.
4 sources [Samsung, Huawei, OPPO, Spreadtrum] observe significant performance gain of 10.7%~27.9% over benchmark for model size > 10M parameters
For layer 1 and payload size Y-bin:
4 sources [CATT, LG, Vivo, Nokia] observe minor to significant performance gain of -1.8%~10.45% over benchmark for model size less than 1M parameters.
5 sources [CATT, LG, Vivo, Nokia, Xiaomi] observe minor to significant performance gain of 1.63% ~11.45% over benchmark for model size in range of 1M to 10M parameters.
For layer 1 and payload size Z-bin:
2 sources [CATT, LG] observe minor to significant performance gain of -3.32%~10.62% over benchmark for model size less than 1M parameters.
4 sources [CATT, LG, Nokia, Xiaomi] observe minor to significant performance gain of -2.05% ~11.99% over benchmark for model size in range of 1M to 10M parameters.
For using angle-delay domain eigen-vector as input,
For layer 1 and payload size X-bin:
4 sources [Ericsson, Samsung, QC, Vivo] observe minor to significant performance gain of 1.4%~19.29% over benchmark when model size < 1M parameters.
1 source [Ericsson] observe moderate performance gain of 6.5%~6.8% over benchmark for model size in range of 1M to 10M parameters.
For using spatial-frequency domain channel matrix as input,
For layer 1 and payload size X-bin:
1 source [Huawei] observes 37.8% performance gain over benchmark when model size is 12M params

Case 0, decoder complexity and model size
In most companies’ results, the encoder and the decoder have similar complexity, as shown in the following plots. Therefore, the performance-complexity trade-off for the decoder should be similar to that of the encoder.




Case 0, comparison of Rel-18 evaluations and Rel-19 evaluations 
The following plot shows the SGCS gain vs. encoder FLOPs, comparing the numbers from CSI_Table 1 (Rel-18) and the numbers from CSI_Table X9 (Rel-19).

It is observed that the performance/complexity trade-off has improved in Rel-19 compared to Rel-18 evaluations.

In summary,
For temporal domain Case 0, most of the SGCS gain is achievable using an encoder model with complexity less than 10M FLOPs. Use of more complex models provides limited additional SGCS gain. Similar trends are observed for the decoder complexity.
For temporal domain Case 0, most of the SGCS gain is achievable using an encoder model with size less than 1M parameters. Use of larger models provides marginal performance improvements. Similar trends are observed for the decoder model size.
For temporal domain Case 0, compared to Rel-18 evaluations, Rel-19 evaluations show improved performance/complexity trade-off. 
Reasons for the improved performance/complexity trade-off include the use more optimized AI/ML model structures and the use of different inputs.



Observation
Case 2, model complexity (FLOPs) vs. SGCS gain (%) 



For temporal domain Case 2, the results on trade-off between complexity (FLOPs) and SGCS gain (%) over benchmark are summarized as follows
For layer1, 
5 sources [Samsung, CATT, LG, QC, Vivo] observe minor to significant performance gain of 3.4%~20% over benchmark when FLOPs <= 10M.
6 sources [ZTE, Apple, QC, Fujitsu, LG, Nokia] observe moderate to significant performance gain of 4.11%~28% over benchmark when 10M< FLOPs <= 100M.
8 sources [Nokia, OPPO, CMCC, Xiaomi, FW, Spreadtrum, ETRI, HW, Nokia] observe moderate to significant performance gain of 4%~27.8% over benchmark when FLOPs > 100M.
where the complexity (FLOPs) is the average of the encoder FLOPs and the decoder FLOPs.

Case 2, model size (# parameters) vs. SGCS gain (%) 



For temporal domain Case 2, the results on trade-off between model size (# parameters) and SGCS gain (%) over benchmark are summarized as follows 
For layer1, 
6 sources [QC, ZTE, Vivo, LG, CATT, Nokia] observe minor to significant performance gain of 3.4%~27.9% over benchmark when model size <= 1M parameters.
9 sources [Apple, Fujitsu, LG, Nokia, CMCC, Xiaomi, HW, ETRI, FW, Nokia] observe moderate to significant performance gain of 4%~27.8% over benchmark when 1M< model size <= 10M parameters.
2 sources [Spreadtrum, OPPO] observe moderate to significant performance gain of 8%~27.8% over benchmark when model size > 10M parameters.
where the model size (# parameters) is the average of the encoder size and the decoder size.

Comparison between Case0 and Case2 in terms of complexity vs performance: 




In summary,
Temporal domain Case 2 can achieve higher gain compared to Case 0 with similar or increased complexity
Some companies achieved gain using Case 2 AI/ML models having same/similar/lower complexity as their Case 0 AI/ML models, while some other companies achieved gain using Case 2 AI/ML models having higher complexity than their Case 0 AI/ML models.
Note: The Case 2 evaluations for this summary were done under no UCI loss.



Observation
Case 3, model complexity (FLOPs) vs. SGCS gain (%) 


For temporal domain Case 3, the results on trade-off between complexity (FLOPs) and SGCS gain (%) over benchmark (Rel-18 doppler eType II) are summarized as follows
For layer1, 
1 source [Ericsson] observes performance gain of 4.22% over benchmark when FLOPs <= 10M.
8 sources [Ericsson, CMCC, QC, DCM, ZTE, CATT, Vivo, MTK] observe minor to significant performance gain of 2.4%~28% over benchmark when 10M< FLOPs <= 100M.
4 sources [Fujitsu, Xiaomi, InterDigital, OPPO] observe minor to significant performance gain of -4%~39.76% over benchmark when FLOPs > 100M.
where the complexity (FLOPs) is the average of the encoder FLOPs and the decoder FLOPs.

Case 3, model size (# parameters) vs. SGCS gain (%) 


For temporal domain Case 3, the results on trade-off between model size (# parameters) and SGCS gain (%) over benchmark (Rel-18 doppler eType II) are summarized as follows
For layer1, 
4 sources [QC, Ericsson, CATT, DCM] observe minor to significant performance gain of 3.9%~22% over benchmark when model size <= 1M parameters.
7 sources [InterDigital, Ericsson, Vivo, Fujitsu, MTK, ZTE, CMCC] observe minor to significant performance gain of -4%~16.6% over benchmark when 1M< model size <= 10M parameters.
2 sources [Xiaomi, OPPO] observe significant performance gain of 20.8%~39.76% over benchmark when model size > 10M parameters.
where the model size (# parameters) is the average of the encoder size and the decoder size.

Comparison between Case0 and Case3 in terms of complexity vs performance: 
Note 1: Case 0 gain is w.r.t. Rel-16 eType II benchmark, in scenario of mixed indoor and outdoor
Note 2: Case 3 gain is w.r.t. Rel-18 Doppler eType II benchmark, in scenario of mixed indoor and outdoor, or outdoor only



In summary,
Under CSI prediction, temporal domain Case 3 can achieve better performance than Rel-18 Doppler eType II benchmark. The amount of gain of Case 3 over Rel-18 Doppler eType II benchmark is similar to the amount of gain of Case 0 over Rel-16 eType II.
On average, for each inference, FLOPs of Case 3 is higher than that of Case 0.

TDoc file unavailable

	
3GPP TSG RAN WG1 #121                                                                                            R1-2504743
Malta, MT, May 19th – 23rd, 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.

Agreements from RAN1 #121
Observation
The model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 can be same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.
	To reduce the workload of the potential normative work scope, the model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 is(are) same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.

Agreement
For addressing inter-vendor collaboration complexity, RAN1 identifies the following specification impacts for supporting sub-option 3a-1 (including with target CSI and without target CSI), sub-option 4-1, and Direction C 
Reference model / structure specification (for sub-option 3a-1 and Direction C)
Note: The so called fully specified reference model discussed in RAN1 for Direction C is assumed to be equivalent to the to-be-specified model, if specified, for performance requirements in RAN4.
Note: The model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 can be same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.
To reduce the workload of the potential normative work scope, the model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 is(are) same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.
Note: RAN1 has studied and agreed on a scalable model structure for feasibility study during Rel-19 study, and considers that can serve as a starting point for the model structure specification for the potential normative work.
Format and contents of parameter/dataset/information exchange: dataset (for 4-1 and 3a-1 with target CSI)/ encoder parameter (for 3a-1) / additional information (for 3a-1 and 4-1).
The workload can be reduced by identifying the common part among different aspects.


Observation
Case 0, encoder complexity (FLOPs) vs. SGCS gain (%) 




For temporal domain Case 0, the results on trade-off between complexity (FLOPs) and SGCS gain (%) over benchmark are summarized as follows
For using spatial-frequency domain eigen-vector as input,
For layer 1 and payload size X-bin:
9 sources [QC, CATT, vivo, Futurewei, OPPO, LG, Fujitsu, IITM, PengCheng Lab] observe minor to significant performance gain of 2.8%~20.68% over benchmark for FLOPs range of 1M to 10M.
11 sources [QC, CATT, vivo, Futurewei, Ericsson, Samsung, OPPO, LG, Fujitsu, PengCheng Lab, Spreadtrum] observes minor to significant performance gain of 1.2% ~21.7% over benchmark for FLOPs range of 10M to 100M.
5 sources [vivo, Huawei, Xiaomi, Spreadtrum, Nokia, OPPO] observe minor to significant performance gain of 4.8%~13.5% over benchmark for FLOPs > 100M
For layer 1 and payload size Y-bin:
3 sources [CATT, LG, Vivo] observe minor to significant performance gain of -1.8%~10.45% over benchmark for FLOPs range of 1M to 10M.
4 sources [CATT, LG, Vivo, Nokia] observe minor to significant performance gain of 0.4%~11.45% over benchmark for FLOPs range of 10M to 100M.
2 sources [Xiaomi, Nokia] observe minor to moderate performance gain of 1.63%~6.1% over benchmark for FLOPs > 100M
For layer 1 and payload size Z-bin:
2 sources [CATT, LG] observe minor to significant performance gain of -3.32%~10.62% over benchmark for FLOPs range of 1M to 10M.
2 sources [CATT, LG] observe minor to significant performance gain of 1.2%~11.99% over benchmark for FLOPs range of 10M to 100M.
2 sources [Xiaomi, Nokia] observe minor performance gain of -2.05%~2.1% over benchmark for FLOPs > 100M
For using angle-delay domain eigen-vector as input
For layer 1 and payload size X-bin:
2 sources [Ericsson, Samsung] observe minor to significant performance gain of 2.8%~19.29% over benchmark when FLOPs < 1M.
3 sources [Ericsson, QC, Vivo] observe minor to moderate performance gain of 1.4%~6.63% over benchmark for FLOPs range from 1M to 10M.
2 sources [Ericsson, QC] observe moderate performance gain of 5.3%~6.8% over benchmark for FLOPs > 10M.
For using spatial-frequency domain channel matrix as input
For layer 1 and payload size X-bin:
1 source [Huawei] observes 37.8% performance gain over benchmark when FLOPs is 100M.

Case 0, encoder model size (# parameters) vs. SGCS gain (%) 




For temporal domain Case 0, the results on trade-off between model size (# parameters) and SGCS gain (%) over benchmark are summarized as follows 
For using spatial-frequency domain eigen-vector as input,
For layer 1 and payload size X-bin:
10 sources [QC, CATT, vivo, Futurewei, Ericsson, OPPO, LG, Fujitsu, PengCheng Lab, IITM] observe minor to significant performance gain of 1.2%~18.5% over benchmark for model size less than 1M parameters.
12 sources [CATT, vivo, Futurewei, Ericsson, Xiaomi, OPPO, LG, Fujitsu, PengCheng Lab, IITM, Spreadtrum, Nokia] observe minor to significant performance gain of 1.3% ~21.7% over benchmark for model size in range of 1M to 10M parameters.
4 sources [Samsung, Huawei, OPPO, Spreadtrum] observe significant performance gain of 10.7%~27.9% over benchmark for model size > 10M parameters
For layer 1 and payload size Y-bin:
4 sources [CATT, LG, Vivo, Nokia] observe minor to significant performance gain of -1.8%~10.45% over benchmark for model size less than 1M parameters.
5 sources [CATT, LG, Vivo, Nokia, Xiaomi] observe minor to significant performance gain of 1.63% ~11.45% over benchmark for model size in range of 1M to 10M parameters.
For layer 1 and payload size Z-bin:
2 sources [CATT, LG] observe minor to significant performance gain of -3.32%~10.62% over benchmark for model size less than 1M parameters.
4 sources [CATT, LG, Nokia, Xiaomi] observe minor to significant performance gain of -2.05% ~11.99% over benchmark for model size in range of 1M to 10M parameters.
For using angle-delay domain eigen-vector as input,
For layer 1 and payload size X-bin:
4 sources [Ericsson, Samsung, QC, Vivo] observe minor to significant performance gain of 1.4%~19.29% over benchmark when model size < 1M parameters.
1 source [Ericsson] observe moderate performance gain of 6.5%~6.8% over benchmark for model size in range of 1M to 10M parameters.
For using spatial-frequency domain channel matrix as input,
For layer 1 and payload size X-bin:
1 source [Huawei] observes 37.8% performance gain over benchmark when model size is 12M params

Case 0, decoder complexity and model size
In most companies’ results, the encoder and the decoder have similar complexity, as shown in the following plots. Therefore, the performance-complexity trade-off for the decoder should be similar to that of the encoder.




Case 0, comparison of Rel-18 evaluations and Rel-19 evaluations 
The following plot shows the SGCS gain vs. encoder FLOPs, comparing the numbers from CSI_Table 1 (Rel-18) and the numbers from CSI_Table X9 (Rel-19).

It is observed that the performance/complexity trade-off has improved in Rel-19 compared to Rel-18 evaluations.

In summary,
For temporal domain Case 0, most of the SGCS gain is achievable using an encoder model with complexity less than 10M FLOPs. Use of more complex models provides limited additional SGCS gain. Similar trends are observed for the decoder complexity.
For temporal domain Case 0, most of the SGCS gain is achievable using an encoder model with size less than 1M parameters. Use of larger models provides marginal performance improvements. Similar trends are observed for the decoder model size.
For temporal domain Case 0, compared to Rel-18 evaluations, Rel-19 evaluations show improved performance/complexity trade-off. 
Reasons for the improved performance/complexity trade-off include the use more optimized AI/ML model structures and the use of different inputs.



Observation
Case 2, model complexity (FLOPs) vs. SGCS gain (%) 



For temporal domain Case 2, the results on trade-off between complexity (FLOPs) and SGCS gain (%) over benchmark are summarized as follows
For layer1, 
5 sources [Samsung, CATT, LG, QC, Vivo] observe minor to significant performance gain of 3.4%~20% over benchmark when FLOPs <= 10M.
6 sources [ZTE, Apple, QC, Fujitsu, LG, Nokia] observe moderate to significant performance gain of 4.11%~28% over benchmark when 10M< FLOPs <= 100M.
8 sources [Nokia, OPPO, CMCC, Xiaomi, FW, Spreadtrum, ETRI, HW, Nokia] observe moderate to significant performance gain of 4%~27.8% over benchmark when FLOPs > 100M.
where the complexity (FLOPs) is the average of the encoder FLOPs and the decoder FLOPs.

Case 2, model size (# parameters) vs. SGCS gain (%) 



For temporal domain Case 2, the results on trade-off between model size (# parameters) and SGCS gain (%) over benchmark are summarized as follows 
For layer1, 
6 sources [QC, ZTE, Vivo, LG, CATT, Nokia] observe minor to significant performance gain of 3.4%~27.9% over benchmark when model size <= 1M parameters.
9 sources [Apple, Fujitsu, LG, Nokia, CMCC, Xiaomi, HW, ETRI, FW, Nokia] observe moderate to significant performance gain of 4%~27.8% over benchmark when 1M< model size <= 10M parameters.
2 sources [Spreadtrum, OPPO] observe moderate to significant performance gain of 8%~27.8% over benchmark when model size > 10M parameters.
where the model size (# parameters) is the average of the encoder size and the decoder size.

Comparison between Case0 and Case2 in terms of complexity vs performance: 




In summary,
Temporal domain Case 2 can achieve higher gain compared to Case 0 with similar or increased complexity
Some companies achieved gain using Case 2 AI/ML models having same/similar/lower complexity as their Case 0 AI/ML models, while some other companies achieved gain using Case 2 AI/ML models having higher complexity than their Case 0 AI/ML models.
Note: The Case 2 evaluations for this summary were done under no UCI loss.



Observation
Case 3, model complexity (FLOPs) vs. SGCS gain (%) 


For temporal domain Case 3, the results on trade-off between complexity (FLOPs) and SGCS gain (%) over benchmark (Rel-18 doppler eType II) are summarized as follows
For layer1, 
1 source [Ericsson] observes performance gain of 4.22% over benchmark when FLOPs <= 10M.
8 sources [Ericsson, CMCC, QC, DCM, ZTE, CATT, Vivo, MTK] observe minor to significant performance gain of 2.4%~28% over benchmark when 10M< FLOPs <= 100M.
4 sources [Fujitsu, Xiaomi, InterDigital, OPPO] observe minor to significant performance gain of -4%~39.76% over benchmark when FLOPs > 100M.
where the complexity (FLOPs) is the average of the encoder FLOPs and the decoder FLOPs.

Case 3, model size (# parameters) vs. SGCS gain (%) 


For temporal domain Case 3, the results on trade-off between model size (# parameters) and SGCS gain (%) over benchmark (Rel-18 doppler eType II) are summarized as follows
For layer1, 
4 sources [QC, Ericsson, CATT, DCM] observe minor to significant performance gain of 3.9%~22% over benchmark when model size <= 1M parameters.
7 sources [InterDigital, Ericsson, Vivo, Fujitsu, MTK, ZTE, CMCC] observe minor to significant performance gain of -4%~16.6% over benchmark when 1M< model size <= 10M parameters.
2 sources [Xiaomi, OPPO] observe significant performance gain of 20.8%~39.76% over benchmark when model size > 10M parameters.
where the model size (# parameters) is the average of the encoder size and the decoder size.

Comparison between Case0 and Case3 in terms of complexity vs performance: 
Note 1: Case 0 gain is w.r.t. Rel-16 eType II benchmark, in scenario of mixed indoor and outdoor
Note 2: Case 3 gain is w.r.t. Rel-18 Doppler eType II benchmark, in scenario of mixed indoor and outdoor, or outdoor only



In summary,
Under CSI prediction, temporal domain Case 3 can achieve better performance than Rel-18 Doppler eType II benchmark. The amount of gain of Case 3 over Rel-18 Doppler eType II benchmark is similar to the amount of gain of Case 0 over Rel-16 eType II.
On average, for each inference, FLOPs of Case 3 is higher than that of Case 0.


Agreement
Replace the following figure in the agreed observation

with the following figure.
 



Observation
Localized models, model complexity (FLOPs) and model size (# parameters) vs. SGCS gain (%) 
Based on the evaluations results in table X3 (Typo correction: Layer 0 in plot 2 needs to be corrected to Layer 1)




For localized models, 
5 sources [ZTE, Vivo, Oppo, Intel, Fujitsu] observed that local models can improve the complexity-performance trade-off compared to global models.



Agreement
For inter-vendor collaboration, RAN1 concludes that both Direction A and Direction C are feasible. For Direction A, RAN1 concludes that the feasible sub-options are sub-option 4-1 and sub-option 3a-1, and in case of sub-option 3a-1, with and without target CSI sharing from NW side. 
Performance: Direction C with standardized reference model (structure + parameters) ensures a minimum performance requirement, while Direction A with standardized reference model structure + parameters (sub-option 3a-1), or standardized dataset exchange (sub-option 4-1) from NW-side to UE-side may achieve better performance. 
Direction C addresses inter-vendor collaboration complexity: From RAN1 perspective, Direction C (a.k.a., option 1) eliminates the inter-vendor collaboration complexity if it is feasible to standardize the model. The fully standardized reference model (structure + parameters) should be trained using synthetic data. The model structure and parameters are assumed to be equivalent to the to-be-specified model, if specified, for performance requirements in RAN4. RAN1 concludes that the scalable model structure specification and model parameters specification is feasible, but the conclusion on interoperability and RAN4 testing feasibility is a separate study in RAN4.
For Direction A with sub-option 4-1 and 3a-1, RAN1 concludes that standardized signaling, if feasible and specified, can be used for model parameter / dataset exchange to alleviate / resolve the inter-vendor collaboration complexity. The standardized signaling, if needed, may be over-the-air, or other approaches. For OTA and non-OTA based standardized signaling, as well as the feasibility of standardized signaling, related status have been provided in [R2-2503169]. For sub-option 3a-1, the specified encoder (a.k.a., CSI generation part) structure should be determined using synthetic data, and the scalable encoder structure specification is feasible.

For NW-side performance monitoring, RAN1 concludes that target CSI reporting via legacy codebooks can be used. Target CSI reporting with CSI codebook enhancement via higher-resolution parameter combination may be beneficial for improving NW-side performance monitoring with additional cost of complexity and overhead at UE side.
For UE-side side monitoring, RAN1 concludes that it is feasible, but 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.
NW side data collection for training is studied including data format, configuration of rank / layer, number of subbands and mechanisms for ground-truth reporting, but not all aspects are concluded. UE side data collection for training is studied including NW configuration or UE request, configuration for temporal aspects, but not all aspects are concluded. Aspects that were not concluded can be discussed in the normative phase.

The study of CSI feedback using two-sided model is complete from RAN1 perspective. 
Based on the study, for addressing inter-vendor collaboration complexity, RAN1 identifies the following specification impacts for supporting sub-option 3a-1 (including with target CSI and without target CSI), sub-option 4-1, and Direction C 
Reference model / structure specification (for sub-option 3a-1 and Direction C)
Note: The so called fully specified reference model discussed in RAN1 for Direction C is assumed to be equivalent to the to-be-specified model, if specified, for performance requirements in RAN4.
Note: The model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 can be same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.
To reduce the workload of the potential normative work scope, the model structure(s) for the reference encoder in inter-vendor collaboration sub-option 3a-1 is(are) same as/equivalent to the structure of the RAN4 defined/specified (if defined/specified for testability) reference encoder.
Note: RAN1 has studied and agreed on a scalable model structure for feasibility study during Rel-19 study, and considers that can serve as a starting point for the model structure specification for the potential normative work.
Format and contents of parameter/dataset/information exchange: dataset (for 4-1 and 3a-1 with target CSI)/ encoder parameter (for 3a-1) / additional information (for 3a-1 and 4-1).
The workload can be reduced by identifying the common part among different aspects.
Based on the study, RAN1 recommends the following with potential RAN1 specification impacts common to all inter-vendor collaboration options.
Recommend Case 0 
Case 2 and 3 can be considered for further enhancement built upon the specification of case 0.
Recommend following specification impacts for the two-sided CSI compression use case
Model pairing procedure including ID and applicability reporting
Inference aspects including target CSI type, measurement and report configuration, CQI / RI determination, payload determination, quantization configuration / codebook, UCI mapping, CSI processing criteria and timeline, priority rules for CSI reports
NW and UE side data collection for training,
Target CSI format 
Note: The framework defined in BM and CSI prediction use cases could be reused.
Specify performance monitoring, if needed
--------------------------------------- END OF TP ------------------------------------

3GPP TSG RAN WG1 #121                                                                                                       R1-2505019
Malta, May 18th – 22nd, 2025

Agenda item:		9.1.4.1
Source: 		Moderator (Qualcomm)
Title: 			Text proposal to capture the output of Rel-19 study of Agenda Item 9.1.4.1 into TR 38.843
Document for:		Discussion and Decision

Conclusions
The following aspects have been studied for the general framework of AI/ML over air interface for one-sided models and two-sided models:
-	Various Network-UE Collaboration Levels
-	Functionality-based LCM and model-ID-based LCM
-	Functionality/model selection, activation, deactivation, switching, fallback
-	Functionality identification and model identification
-	Data collection
-	Performance monitoring
-	Various model identification Types and their use cases
-	Reporting of applicable functionalities/models
-	Method(s) to ensure consistency between training and inference regarding NW-side additional conditions (if identified) for inference at UE
-	Model delivery/transfer and analysis of various model delivery/transfer Cases
The above studied aspects for General Framework can be considered for developing/specifying AI/ML use cases and common framework (if needed for some aspects) across AI/ML use cases.
CSI feedback enhancement:
CSI compression sub use case: 
The performance benefit and potential specification impact were studied for AI/ML based CSI compression sub use case. 
Evaluation has been performed to assess AI/ML based CSI compression from various aspects, including performance gain over non-AI/ML benchmark, model input/output type, CSI feedback quantization methods, ground-truth CSI format, monitoring, generalization, training collaboration types, etc. Some aspects were studied but not fully investigated, including the options of CQI/RI calculation, the options of rank>1 solution.
Performance gain over baseline and computational complexity in FLOPs are summarized in clause 6.2.2.8. 
Potential specification impact on NW side/UE side data collection, dataset delivery, quantization alignment between CSI generation part at the UE and CSI reconstruction part at the NW, CSI report configuration, CSI report format, pairing information/procedure and monitoring approach were investigated but not all aspects were identified. 
The pros and cons are analysed for each training collaboration types, and each training collaboration type has its own benefits and limitations in different aspects. The study has investigated the feasibility of the studied training collaboration types and necessity of corresponding potential RAN1 specification impact. However, not all aspects have been concluded.
Both NW side and UE side performance monitoring were studied, some but not all aspects were concluded.
From RAN1 perspective, there is no consensus on the recommendation of CSI compression for normative work.
At least the following aspects are the reasons for the lack of RAN1 consensus on the recommendation of CSI compression for normative work:
-	Trade-off between performance and complexity/overhead.
-	Issues related to inter-vendor training collaboration.
Other aspects that require further study/conclusion are captured in the summary above.















































CSI prediction sub use case: 
---------------------------------------------------------------Text omitted ---------------------------------------------------------
Beam management:
---------------------------------------------------------------Text omitted ---------------------------------------------------------
Positioning accuracy enhancements: 
---------------------------------------------------------------Text omitted ---------------------------------------------------------

------------------------------- END OF TP ------------------------------ 


Comments (before/during RAN1 #121 meeting)


Comments (1st round)



Comments (2nd round)