REVIEW 3 major objections 4 minor 24 references
Position Paper: Rethinking AI/ML for Air Interface in Wireless Networks
T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read The blockers to AI at the wireless air interface are data governance and lifecycle management, not model accuracy.
desk verdict A clean, well-organized position paper that maps 3GPP Release 18/19 AI/ML air-interface discussions onto a useful challenge taxonomy, but its central priority claim is asserted rather than established. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The organizing machinery is the AI/ML model pipeline as viewed through 3GPP's air-interface work: data governance (collection, cleaning, privacy, security), training and testing, and deployment with monitoring and management, run in both centralized and distributed settings. The paper maps six research directions—multi-task learning, conditional architectures, root-cause analysis, opportunistic data collection, reinforcement learning/optimization, and efficient labeling—onto the specific gaps in that pipeline. These directions are the mechanism by which the paper expects the deployment barriers to be overcome.
What would settle it
A documented field deployment or 3GPP decision in which an AI/ML air-interface function meets operator KPIs without solving cross-vendor interoperability, data ownership, or continuous monitoring would contradict the paper's claim that these are the main barriers. Concretely, a work item that standardizes a high-accuracy model while leaving data governance and lifecycle management unaddressed—and still sees commercial adoption—would falsify the position.
Extended reading notes
Core claim
The paper's central claim is that fully realizing AI/ML at the air interface requires a hybrid approach that combines modular, adaptive algorithms with robust data governance and continuous model monitoring, and that the community should reduce its reliance on hard-to-acquire labeled data. It contends that 3GPP has laid the foundations for AI/ML adoption in Releases 18 and 19, but the decisive challenges for 6G are lifecycle management (training, testing, monitoring, KPIs), data governance, and interoperability between UE-side and network-side models. The paper does not present new measurements; it offers a synthesis of the standardization landscape and a research agenda.
Load-bearing premise
The position rests on the assumption that the paper's reading of 3GPP Release 18/19 discussions is complete and representative, and that its six research directions are the right levers; if the standards work is driven by different constraints, the proposed priorities would be misdirected.
Editorial extensions
If this is right
- If multi-task and conditional architectures are adopted, UE-side computational load and signaling overhead drop while monitoring becomes more tractable.
- If root-cause analysis matures, an operator can isolate whether a failure comes from the encoder, the decoder, or the radio environment, enabling partial retraining instead of full model replacement.
- If opportunistic data collection is used, labeling and testing resources are spent only when performance degrades or environments shift, not through continuous logging.
- If reinforcement learning or optimization-based methods replace accuracy-only models, models can be tuned directly to quality-of-service and quality-of-experience targets, reducing the need for labeled datasets.
- Pursued jointly, these directions yield the hybrid solutions the paper concludes are needed for 6G: modular adaptive algorithms combined with data governance and continuous monitoring.
Reading between the lines
- The paper's emphasis implies that the research community's focus on novel architectures may be misaligned with what operators actually need; a survey of operator priorities across 3GPP meetings would test this.
- If cross-vendor interoperability becomes the binding constraint, standardization may shift toward specifying split-model interfaces (such as early-exit layer boundaries) rather than model internals, analogous to codec bitstream standards.
- The KPI-alignment argument suggests that task-oriented metrics—say, video-streaming beam quality instead of top-beam hit rate—could replace generic accuracy benchmarks in future standardization, changing how models are evaluated.
- The paper's hybrid conclusion could be sharpened into a concrete research program: a live-network prototype combining an adaptive RL policy with a self-supervised channel representation and opportunistic labeling; the paper does not describe such a system.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This position paper reviews the current 3GPP Release 18/19 discussions on AI/ML for the air interface, organizes the associated challenges into data governance, lifecycle management, and interoperability, and then proposes six research directions (multi-task learning, conditional architectures, root cause analysis, opportunistic data collection, reinforcement learning, and efficient label use). The central conclusion is that future work should prioritize hybrid solutions that combine modular adaptive algorithms with robust data governance and continuous model monitoring, rather than focusing on predictive accuracy alone. The manuscript is a qualitative position/review piece: it contains no quantitative experiments or derivations, and its contribution lies in synthesizing and framing the standardization landscape.
Significance. If the position is accepted by the community, it could usefully shift research attention toward lifecycle, governance, and interoperability issues that are often underplayed in AI/ML wireless papers. The paper provides a clear taxonomy of challenges and a compact table mapping research directions to those challenges, which can serve as a discussion basis for 6G standardization. It draws on real 3GPP TR 38.843 discussions and cites relevant prior work, including the authors' own contributions, which are used as illustrative examples. The manuscript is timely and readable, and it makes no falsifiable quantitative claims, so its soundness should be judged on the quality of its argumentation rather than on empirical evidence.
major comments (3)
- [Section 4, Conclusion and Section 2] The central priority claim—that data governance, lifecycle management, and interoperability are the main obstacles to air-interface AI/ML—is asserted rather than established. The paper states in Section 2 that a clean pipeline is 'equally (or more) important' than model performance and concludes that future research 'should focus' on hybrid solutions in these areas, but it does not provide evidence from 3GPP TR 38.843 that these challenges are the binding constraints. No comparison is made against alternative bottlenecks such as prediction accuracy thresholds, computational complexity, signaling overhead, or evaluation methodology. Table 1 assigns checkmarks to research directions without derivation. To make the recommendation actionable, the authors should either weaken the prioritization language or substantiate it, for example by categorizing the open issues in TR 38.843 by theme or by arguing explicitly why other bottlenecks are less limiting.
- [Section 3.5 and Table 1] The claim that reinforcement learning 'eases the burden on data collection for model training, testing, and monitoring' and the corresponding 'simplified testing' checkmark in Table 1 are not justified and are likely misleading. RL requires careful reward design, environment interaction, and off-policy evaluation, which are generally more involved than supervised evaluation; no testing protocol for learned policies is described. The paper should either remove the simplified-testing claim or replace it with a concrete discussion of how RL policies would be validated in a wireless deployment (e.g., via simulation, safe exploration, or online monitoring). This is load-bearing because the RL direction is one of the six proposed research avenues.
- [Section 2.2.1] The statement that 'larger models exhibit better performance and generalization capabilities' is supported only by a single theoretical study of the XOR problem (Brutzkus & Globerson, 2019). That result does not establish a general scaling relationship, and the subsequent inference that fulfilling Release 19 requirements will necessarily lead to larger models and increased energy consumption is too strong. The authors should either cite broader scaling-law literature or qualify the claim as one possible trend. This matters because the model-complexity challenge is part of the lifecycle-management argument that underpins the paper's overall position.
minor comments (4)
- [Table 1] The checkmarks in Table 1 are presented without explanation. Please add a sentence describing how these assignments were made (e.g., based on the authors' reading of the cited challenges) so that readers can interpret the table's authority.
- [Section 4, Conclusion] The phrase 'cost-deficient label-based approaches' is awkward and likely a typo; consider 'costly' or 'cost-inefficient'.
- [Section 2.1] The sentence 'the fundamental issue of which data can be collected in a way that respects user's security and privacy aspects plays central role' has grammatical issues; consider revising to 'plays a central role'.
- [Section 3.5] The phrase 'eases the burden on data collection' should be substantiated with an example of how reward measurement would be obtained in an air-interface setting without extensive labeled data, since reward calculation itself may require ground-truth information.
Circularity Check
No circularity: a position/review paper whose claims are asserted and supported by external citations, with no derivation, fitting, or prediction that reduces to its own inputs.
full rationale
This is a position and survey paper, not an empirical or theoretical derivation. It summarizes 3GPP Release 18/19 discussions, organizes challenges into a taxonomy (data governance, lifecycle management, interoperability), and proposes six research directions. There is no fitted parameter that is later called a prediction, no equation whose output is defined in terms of its input, and no uniqueness theorem imported from the authors' prior work. The authors cite their own prior papers (Alawieh & Kontes 2023 in the Introduction; Ott et al. 2024 in Section 3.1; Manjunath et al. 2025 in Section 3.3) but only as illustrative examples of existing methods, not as load-bearing justification for the paper's central recommendation. The conclusion that future research should focus on hybrid solutions combining modular adaptive algorithms with robust data governance and continuous model monitoring is an opinionated research-prioritization claim, not a derived result; its plausibility depends on whether the taxonomy and selection of directions are representative, which is a scope/selection concern rather than a circularity concern. The paper is therefore self-contained as a position statement, and the appropriate circularity score is 0.
Assumptions & free parameters
assumptions (4)
- domain assumption AI/ML will be essential for 5G/6G air interface to handle directional beams, channel variation, feedback overhead, and localization.
- domain assumption The authors' summary of 3GPP Release 18/19 AI/ML discussions and the challenge list in Sections 2.1 to 2.3 is complete and representative.
- ad hoc to paper The six research directions in Table 1 are the relevant levers for 6G AI/ML adoption.
- domain assumption Larger AI/ML models have better generalization and therefore increase energy consumption in a problematic way.
Cite this review
Pith. "Pith review of Position Paper: Rethinking AI/ML for Air Interface in Wireless Networks." pith.science (2026). https://pith.science/paper/X7CD2Q62
@misc{pith2026250611466,
author = {Pith},
title = {Pith review of: Position Paper: Rethinking AI/ML for Air Interface in Wireless Networks},
year = {2026},
howpublished = {\url{https://pith.science/paper/X7CD2Q62}},
note = {Machine review of arXiv:2506.11466}
}
read the original abstract
AI/ML research has predominantly been driven by domains such as computer vision, natural language processing, and video analysis. In contrast, the application of AI/ML to wireless networks, particularly at the air interface, remains in its early stages. Although there are emerging efforts to explore this intersection, fully realizing the potential of AI/ML in wireless communications requires a deep interdisciplinary understanding of both fields. We provide an overview of AI/ML-related discussions in 3GPP standardization, highlighting key use cases, architectural considerations, and technical requirements. We outline open research challenges and opportunities where academic and industrial communities can contribute to shaping the future of AI-enabled wireless systems.
Reference graph
Works this paper leans on
-
[1]
write newline
" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
-
[2]
3GPP TR38.843 . Technical Report Group Radio Access Network; Study on Artificial Intelligence (AI)/Machine Learning (ML) for NR air interface ( V18.0.0) . Technical report, The 3rd Generation Partnership Project, Sophia Antipolis, France, 2024
work page 2024
-
[3]
Alawieh, M. and Kontes, G. 5G positioning advancements with AI/ML . arXiv preprint arXiv:2401.02427, 2023
arXiv 2023
-
[4]
Assuring the machine learning lifecycle: Desiderata, methods, and challenges
Ashmore, R., Calinescu, R., and Paterson, C. Assuring the machine learning lifecycle: Desiderata, methods, and challenges. ACM Computing Surveys (CSUR), 54 0 (5): 0 1--39, 2021
work page 2021
-
[5]
Brutzkus, A. and Globerson, A. Why do larger models generalize better? a theoretical perspective via the xor problem. In International conference on machine learning, pp.\ 822--830. PMLR, 2019
work page 2019
-
[6]
Meta-learning in neural networks: A survey
Hospedales, T., Antoniou, A., Micaelli, P., and Storkey, A. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44 0 (9): 0 5149--5169, 2021
work page 2021
-
[7]
Huang, Q. and Zhao, T. Data collection and labeling techniques for machine learning. arXiv preprint arXiv:2407.12793, 2024
arXiv 2024
-
[8]
AI for CSI Prediction in 5G-Advanced and Beyond
Jiang, C., Guo, J., Li, X., Jin, S., and Zhang, J. AI for CSI prediction in 5G -advanced and beyond. arXiv preprint arXiv:2504.12571, 2025
work page Pith review arXiv 2025
Show all 24 references
-
[9]
The bridge toward 6G : 5G -advanced evolution in 3GPP release 19
Lin, X. The bridge toward 6G : 5G -advanced evolution in 3GPP release 19. IEEE Communications Standards Magazine, 9 0 (1): 0 28--35, 2025
2025
-
[10]
Self-supervised learning: Generative or contrastive
Liu, X., Zhang, F., Hou, Z., Mian, L., Wang, Z., Zhang, J., and Tang, J. Self-supervised learning: Generative or contrastive. IEEE transactions on knowledge and data engineering, 35 0 (1): 0 857--876, 2021
2021
-
[11]
Multimodal-to-text prompt engineering in large language models using feature embeddings for GNSS interference characterization
Manjunath, H., Heublein, L., Feigl, T., and Ott, F. Multimodal-to-text prompt engineering in large language models using feature embeddings for GNSS interference characterization. arXiv preprint arXiv:2501.05079, 2025
2025 arXiv
-
[12]
I know what you trained last summer: A survey on stealing machine learning models and defences
Oliynyk, D., Mayer, R., and Rauber, A. I know what you trained last summer: A survey on stealing machine learning models and defences. ACM Computing Surveys, 55 0 (14s): 0 1--41, 2023
2023
-
[13]
Radio foundation models: Pre-training transformers for 5g-based indoor localization
Ott, J., Pirkl, J., Stahlke, M., Feigl, T., and Mutschler, C. Radio foundation models: Pre-training transformers for 5g-based indoor localization. In 2024 14th International Conference on Indoor Positioning and Indoor Navigation (IPIN), pp.\ 1--6. IEEE, 2024
2024
-
[14]
B., Chen, X., and Wang, X
Ren, P., Xiao, Y., Chang, X., Huang, P.-Y., Li, Z., Gupta, B. B., Chen, X., and Wang, X. A survey of deep active learning. ACM computing surveys (CSUR), 54 0 (9): 0 1--40, 2021
2021
-
[15]
Exploring LLM -based agents for root cause analysis
Roy, D., Zhang, X., Bhave, R., Bansal, C., Las-Casas, P., Fonseca, R., and Rajmohan, S. Exploring LLM -based agents for root cause analysis. In Companion Proceedings of the 32nd ACM International Conference on the Foundations of Software Engineering, pp.\ 208--219, 2024
2024
-
[16]
and Norvig, P
Russell, S. and Norvig, P. Artificial Intelligence: A Modern Approach (4th Edition) . Pearson, 2020. ISBN 9780134610993
2020
-
[17]
Channel charting: Locating users within the radio environment using channel state information
Studer, C., Medjkouh, S., Gonulta s , E., Goldstein, T., and Tirkkonen, O. Channel charting: Locating users within the radio environment using channel state information. IEEE Access, 6: 0 47682--47698, 2018
2018
-
[18]
and Barto, A
Sutton, R. and Barto, A. Reinforcement Learning: A n Introduction . Cambridge, MA, MIT Press, 2018
2018
-
[19]
Branchynet: Fast inference via early exiting from deep neural networks
Teerapittayanon, S., McDanel, B., and Kung, H.-T. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd international conference on pattern recognition (ICPR), pp.\ 2464--2469. IEEE, 2016
2016
-
[20]
K., and Ristenpart, T
Tram \`e r, F., Zhang, F., Juels, A., Reiter, M. K., and Ristenpart, T. Stealing machine learning models via prediction APIs . In 25th USENIX security symposium (USENIX Security 16), pp.\ 601--618, 2016
2016
-
[21]
Wolsey, L. A. and Nemhauser, G. L. Integer and combinatorial optimization. John Wiley & Sons, 1999
1999
-
[22]
AI/ML for beam management in 5G -advanced: a standardization perspective
Xue, Q., Guo, J., Zhou, B., Xu, Y., Li, Z., and Ma, S. AI/ML for beam management in 5G -advanced: a standardization perspective. IEEE Vehicular Technology Magazine, 2024
2024
-
[23]
Deep learning based recommender system: A survey and new perspectives
Zhang, S., Yao, L., Sun, A., and Tay, Y. Deep learning based recommender system: A survey and new perspectives. ACM computing surveys (CSUR), 52 0 (1): 0 1--38, 2019
2019
-
[24]
and Yang, Q
Zhang, Y. and Yang, Q. A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34 0 (12): 0 5586--5609, 2021
2021
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.