Summary of S4-260179: Considerations for SA4 Terms of Reference

1. Introduction and Background

This contribution provides considerations for updating the SA4 Terms of Reference (ToR), following discussions at SA4#134 in Dallas where no agreement was reached on S4-252136. The document presents Huawei/HiSilicon's views on the ToR update with a revised version of the previous contribution.

2. AI/ML Related Traffic Characteristics

Main Position

The document argues that traffic characteristics input is the most urgent topic as it affects progress in other working groups.

Key Technical Points

• Related work on general AI traffic characteristics has already been addressed in SA1 and SA2 in Release 18 (e.g., TS 23.501 includes 5QI 88 for AI/ML traffic)
• SA4 should focus specifically on media service-related AI traffic characteristics to complement SA1/SA2 work
• Real-time media-related AI traffic is of particular interest to 3GPP working groups
• This includes representations and formats for AI-related media, such as multi-modal AI

Supporting Contribution

The source references companion contribution S4-260096 on AI-related formats used in state-of-the-art AI/ML systems with media components, indicating that TR 26.927 and the 6G study need updates.

3. AI/ML Quality Evaluation

Main Position

The document argues against including general AI/ML quality evaluation in SA4 ToR.

Rationale

• Quality evaluation of AI is highly task-dependent
• Many AI services are not under SA4 control
• Different tasks have different evaluation criteria
• Developing generic AI quality evaluation mechanisms is not trivial
• Could introduce numerous quality evaluation problems SA4 cannot address

Recommended Scope

SA4 should only address quality evaluation when:
- Very constrained to specific media-related use cases
- Valuable in the SA4 context (e.g., AI codec or other specific media-related AI use cases)

4. General Considerations

Scope Boundaries

• Most AI traffic (text/application data) is outside SA4 scope: MCP/A2A and agent protocols use text training data and application access, which have never been in SA4's domain
• Timed text differs from general unstructured web text and was traditionally in multimedia scope due to synchronization requirements

Media-Related AI Traffic

• For AI-related media traffic, multimodal AI requires attention
• Different representations are popular compared to previous considerations
• State-of-the-art updates needed in TR 26.927 or 6G study

AI Codec vs. AI Model Distinction

• AI codec/representation format evaluation differs from evaluating the AI model itself
• Different codecs/formats can compress/represent neural networks and AI models
• Codec development work is appropriate for SA4
• Work beyond codecs should be limited to well-defined media-related use cases

6G Considerations

• 6G AI/ML use cases should be considered for AI traffic characteristics
• Embodied AI identified as having the most stringent real-time requirements relevant to the network
• Current mobile networks meet most requirements of today's AI services (chatbots), making them less useful as anchors for 6G targeting 2030
• Previous ToR mentioned 5G; may be necessary to refer to 6G-specific aspects and use cases

General AI Usage in 3GPP

• General usage of AI in 3GPP (e.g., for network operation) should not be linked to codec working group SA4
• Such applications are outside SA4's expertise

5. Proposal

The document proposes incorporating these suggestions into any revision of the SA4 Terms of Reference, with suggested updates provided in an attachment to the paper.