3GPP standardises AI/ML for 5G media along two complementary tracks, owned by different working groups. SA2 (system architecture) defines the Network Data Analytics Function (NWDAF), which collects data from network functions in the 5G core and produces analytics and predictions that other functions can consume. SA4 (media codecs and delivery) defines the Data Collection and Reporting framework, which standardises how data is gathered from User Equipment (UE) and media clients. The two tracks differ in where their data comes from and what they produce, but a media-aware AI/ML use case (for example, predicting a QoS drop before it affects a live stream) typically needs both: UE-side observations flowing in through SA4's framework, and network-side signals and inference flowing in through SA2's. For acronyms used here, see the Glossary.
Key specifications: 3GPP TS 23.288 (NWDAF architecture, SA2), TS 26.531 and TS 26.532 (Data Collection and Reporting, SA4), TR 26.847 (Evaluation of AI and ML in 5G media services).
Reference tools: the AI/ML Evaluation Framework provides benchmarks and datasets for evaluating AI/ML models applied to media processing, aligned with the SA4 study work below.
Go deeper
Technology & Analysis
NWDAF's internal structure, how it relates to SA4 data collection, and the SA4 AI/ML media study, below.
Standards Tracking
Normative 3GPP specs (TS 23.288, TS 26.531, TS 26.532) and 5G-MAG's contributions on AI/ML.
Software Tools
Benchmarks and datasets for evaluating AI/ML models for media quality, ABR and traffic classification.
NWDAF: network-side analytics (SA2)
The Network Data Analytics Function (NWDAF) is the 5G core function through which network functions and other data sources expose measurements, and through which consumers request analytics or predictions derived from that data. TS 23.288 splits the NWDAF's internal role in two:
- AnLF (Analytics Logical Function): produces the analytics or prediction output that a consumer (for example the PCF, for QoS policy, or an Application Function such as the 5GMS AF) subscribes to or requests.
- MTLF (Model Training Logical Function): trains the ML models that an AnLF uses for inference, and can expose those models to other NWDAF instances.
Splitting training from inference lets a model be trained once (potentially on a different NWDAF instance, closer to where the training data volume is largest) and reused for inference wherever it is needed. For a media session, a typical consumer-facing output is a load or QoS prediction that a function such as the 5GMS AF can act on ahead of time, for example by adjusting a bitrate ceiling before congestion actually hits. See Standards: 5G Media Streaming for where the 5GMS AF sits in that path.
SA4 data collection: the UE-side input
The SA4 Data Collection and Reporting framework (TS 26.531, TS 26.532) standardises how a UE or media client reports consumption and quality-of-experience (QoE) data, and how that data is exposed as events to a consuming function. It is a generic framework, reused inside 5G Media Streaming for QoE reporting and available standalone. The full architecture (the Data Collection Application Function, its R1-R6 reference points, and how it exposes events to consumers including the NWDAF) is covered in detail on Tech: UE Data Collection; this page does not repeat that detail.
For AI/ML purposes, the relevant point is where the two tracks meet: the Data Collection Application Function can expose its processed events to the NWDAF as one of its consumers (reference points R5/R6), so UE-side observations gathered under the SA4 framework can feed into SA2's network-side analytics rather than the two tracks running in isolation.
The SA4 AI/ML media study: TR 26.847
TR 26.847 (Evaluation of AI and ML in 5G media services, Release 19, completed June 2025) is SA4's study into applying AI/ML techniques directly to media processing and delivery, as distinct from the network-analytics use of AI/ML in NWDAF above. It defines evaluation methodology, use cases and metrics for assessing AI/ML models in areas such as adaptive bitrate optimisation and media quality enhancement; the AI/ML Evaluation Framework reference tooling is aligned with this study.
TR 26.847 was previously tracked under a different, now-superseded number (TR 26.927, last seen as an incomplete draft) with more detailed scope notes (split inference, model delivery, named logical functions). Those specifics have not been independently re-checked against TR 26.847's final published content. Verify before relying on them.
Related AI/ML studies that feed 6G
TR 22.874 (SA1, Study on traffic characteristics and performance requirements for AI/ML model transfer in 5GS) is a companion study, on the network-transport side of AI/ML rather than the media-processing side: it looks at how model-transfer traffic itself behaves on the 5G system. Both this study and TR 26.847 feed 5G-MAG's early input to 6G, where AI-native traffic management is one of the new IMT-2030 usage scenarios; see Towards 6G Media for that wider context.
The Release placement of TS 26.531 / TS 26.532 relative to TS 23.288, and the exact NWDAF consumer/producer role split described above, were not confirmed against a primary source (the 3GPP/ETSI portals block automated access). Verify against the 3GPP work plan before publication.
Related
- Tech: UE Data Collection, Reporting and Event Exposure: the DCAF architecture and reference points in full.
- Towards 6G Media: how these AI/ML studies feed 5G-MAG's 6G requirements input.
- Standards: AI/ML in 5G Media Standards: the standards-tracking view of this topic.
Refer to the Tech repository to contribute to this documentation.