Meeting: TSGS4_135_India | Agenda Item: 5.2
LS on completion of Study on AI/ML consistency alignment
TSG SA
LS in
Information
Rel-19
| TDoc | S4-260028 |
| Title | LS on completion of Study on AI/ML consistency alignment |
| Source | TSG SA |
| Agenda item | 5.2 |
| Agenda item description | Other 3GPP groups |
| Doc type | LS in |
| For action | Information |
| Release | Rel-19 |
| Original LS | SP-251699 |
| download_url | https://www.3gpp.org/ftp/tsg_sa/WG4_CODEC/TSGS4_135_India/Docs/S4-260028.zip |
| To | TSG RAN, TSG CT,SA WG1, SA WG2, SA WG3, SA WG4, SA WG5, SA WG6,RAN WG1, RAN WG2, RAN WG3, RAN WG4, RAN WG5,CT WG1, CT WG3, CT WG4, CT WG6 |
| For | Information |
| Type | LS in |
| Contact | Andrijana Brekalo |
| Uploaded | 2026-01-21T15:41:08.193000 |
| Contact ID | 91743 |
| TDoc Status | noted |
| Reservation date | 21/01/2026 15:29:42 |
| Agenda item sort order | 7 |
The LS does not list:
- the actual terms added/changed in TS 21.905,
- the definitions,
- any scope notes (e.g., “AI model”, “ML model”, “training”, “inference”, “feature”, “dataset”, “federated learning”, “online learning”, “model drift”, etc.),
- or any non-goals (e.g., not standardizing algorithms).
For recipient WGs, “apply standardized terminology” is not actionable without at least a high-level list of the key terms and what changed.
A completion LS should ideally include:
- a list of impacted specifications (where terminology conflicts were found),
- whether any follow-up CRs are needed in those specs to align wording,
- guidance on transition (e.g., “replace term X with Y in future CRs; do not retroactively update legacy specs unless touched”).
None of that is present.
Even “just terminology” in TS 21.905 can have broad consequences:
- If definitions are normative (TS 21.905 is normative vocabulary), they can affect interpretation of requirements in other TS/TRs.
- If a term is redefined, it may create unintended reinterpretation of existing requirements where that term is used informally.
The LS does not assess:
- whether any existing specs become ambiguous under the new definitions,
- whether any existing usage becomes non-compliant with the vocabulary.
The LS should acknowledge this and propose a pragmatic approach (e.g., apply on touch, not mass retrofit).
Requesting all WGs to “apply” terminology without:
- a compliance mechanism,
- a checklist,
- or a mapping table,
may lead to partial adoption, increasing inconsistency rather than reducing it.
The LS says “No further work is planned” and the study is complete. That is feasible administratively, but technically it may be premature if:
- terminology changes require follow-up CRs in other specs,
- or if the study identified inconsistencies that were not resolved beyond vocabulary.
A study completion often triggers a WI (normative work) if alignment requires more than definitions.
The LS reads like a status note. For a topic titled “AI/ML Consistency Alignment,” it lacks:
- problem statement recap,
- key findings,
- what inconsistencies were found,
- what was decided (and why).
There is no argumentation showing:
- what inconsistencies existed,
- how the new definitions fix them,
- and how WGs should apply them in practice.
Without scope boundaries, recipients may interpret and apply inconsistently.
Include in the LS body (not only attachments):
- 5–10 bullet points of key terminology additions/changes (term + one-line definition or intent).
- A short summary of TR 22.850 conclusions:
- what inconsistencies were found,
- what principles were agreed (e.g., “AI vs ML usage”, “model lifecycle terms”, “training vs inference”, etc.).
Add an “Implementation guidance” section:
- “When drafting new specs/CRs, use terms X/Y/Z as defined in TS 21.905.”
- Provide a mapping table: “legacy term → preferred term”.
- Clarify whether existing specs should be updated:
- “Apply on touch” (recommended) vs “mass alignment CRs”.
If TR 22.850 identified inconsistencies beyond vocabulary (likely), propose:
- a small normative WI or a coordinated set of CRs to align key specs (SA2/SA5/CT/RAN as applicable),
- or at least a plan for cross-WG editorial alignment.
As an LS, the document is directionally fine (informing completion and requesting adoption), but it contains a clear technical error (TS vs TR 21.905) and is insufficiently detailed to enable consistent cross-WG adoption. Without a term list, mapping guidance, and impact notes, the request to “apply standardized terminology” is not practically actionable and risks inconsistent uptake.
[Technical] The LS asks groups to “apply the standardized AI/ML terminology defined in the updated TR 21.905,” but TR 21.905 is a definitions/abbreviations repository and not a normative “vocabulary specification”; the LS should clarify that the intent is consistent use of defined terms/abbreviations rather than imposing normative requirements.
[Technical] The document provides no summary of what terminology was actually added/changed by “CR 0128 on TR 21.905” (e.g., specific new definitions, deprecated terms, scope notes), making it hard for recipient groups to assess impact or identify where their specs may now be inconsistent.
[Technical] Claiming “No further work is planned” is potentially premature for a cross-cutting terminology alignment topic; at minimum, the LS should indicate whether any follow-on normative work (e.g., TS-level alignment, guidance for model lifecycle, data governance, or performance metrics) was explicitly deemed out of scope or deferred.
[Technical] The LS does not identify which existing 3GPP AI/ML-related terms are expected to be replaced (e.g., “AI,” “ML,” “model,” “training,” “inference,” “dataset,” “feature,” “drift”), nor does it provide a mapping/transition note, risking inconsistent adoption and parallel terminology across groups.
[Technical] There is no guidance on how to handle conflicts with already-established terminology in RAN/CT/SA specifications (e.g., if a term is already defined differently in a TS/TR); the LS should state precedence rules or at least recommend reconciliation via CRs.
[Technical] The deliverable list references “TR 22.850 v1.0.0” but does not state whether it is purely informative or contains recommendations that should be carried into normative specs; recipients may miss required follow-up if recommendations exist.
[Technical] The LS does not specify the version/date of TR 21.905 that includes the approved CR 0128, which is necessary for groups to reference the correct baseline and avoid applying terminology inconsistently across releases/versions.
[Editorial] The document number “S4-260028” conflicts with the stated “Document Number: SP-251699”; clarify whether this is an S4 LS draft, an SP document, or a re-numbered item to avoid traceability issues in meeting records.
[Editorial] The meeting/date information is inconsistent: it says “TSG-SA Meeting #110 (Baltimore, US, December 2025)” while the attachments are labeled “SP-251xxx,” which suggests a different numbering/timeframe; ensure the identifiers align with SA#110 outputs.
[Editorial] The “Attachments” list references SP-251698 and SP-251666 but the LS itself is SP-251699; include explicit titles and versions for the attachments (and confirm they are the final approved versions) to improve clarity for recipients.
[Editorial] The phrase “updated TR 21.905” is ambiguous; use “TR 21.905 (with CR 0128 approved at SA#110)” or provide the resulting version number to remove ambiguity.
[Editorial] “Apply the standardized AI/ML terminology … in their respective work” is vague; consider rewording to “use the terms/definitions/abbreviations as defined in TR 21.905 and avoid introducing conflicting definitions” to make the requested action unambiguous.