Pith. sign in

REVIEW 2 major objections 6 minor 21 references

Enabling Real-Time AI in O-RAN: Deploying and Measuring AI Inside a Near-RT RIC xApp

T0 review · 2 major / 6 minor · reviewed 2026-07-12 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Lightweight AI compiled into a Near-RT RIC xApp runs in microseconds and stays under the 10 ms control budget on a live testbed.

desk verdict Solid systems measurement paper: C-compiled µs inference inside a live FlexRIC xApp, with honest feasibility scoping and a useful commodity testbed tool. read the letter →

arxiv 2607.01583 v2 pith:DXBV5VHC submitted 2026-07-02 cs.NI

Lawrence Obiuwevwi , Krzysztof J. Rechowicz , Sampath Jayarathna , Safdar Hussain Bouk , Fahmida Afrin , C. Nicolas Barati , Neda Moghim , Valentina Nanou
show 2 more authors
Muhammad Enayetur Rahman Sachin Shetty
This is my paper · ORCID
classification cs.NI
keywords O-RANNear-RTRICxAppAIinferencenetworkstateclassificationlogisticregressionMLPOpenAirInterface
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether simple AI can be put inside the Near-Real-Time RAN Intelligent Controller software loop without breaking the 10 ms to 1 s control timing that O-RAN requires. The authors build a network-state classification xApp on a live OpenAirInterface and FlexRIC stack, train logistic regression and a shallow multilayer perceptron offline on a structured synthetic proxy dataset of cross-layer latency and load features, export the trained weights as pure C arrays, and compile them into the xApp binary so there is no external machine-learning runtime. On the live testbed they measure inference at 1–5 µs for logistic regression and 10–25 µs for the MLP, with typical end-to-end service latency under 4 ms; both models meet the 10 ms Near-RT budget for more than 95 percent of projected loop executions. The result they establish is feasibility of deterministic embedded inference, not production-grade accuracy on real traffic. They also release a single-binary orchestration dashboard so others can reproduce the stack on ordinary hardware.

What carries the argument

The offline-to-online C-export pipeline: trained logistic-regression and shallow-MLP parameters are written as static const float arrays in C headers, compiled into the xApp binary, and run inline on five-dimensional service-model feature vectors (MAC, RLC, PDCP, GTP latencies and UE count) with monotonic-clock instrumentation of service, inference, and end-to-end latency.

What would settle it

Replace the synthetic proxy with live multi-cell E2 indications under realistic load, re-export the same C models (or a modestly larger one), and re-measure the end-to-end latency CDF: if more than a few percent of loops exceed 10 ms or inference ceases to be negligible relative to E2 transport, the central feasibility claim fails for that workload.

Watch

Extended reading notes

Core claim

Lightweight supervised models, exported as deterministic C inference modules and compiled into a Near-RT RIC xApp, can classify cross-layer network states on a live OAI/FlexRIC testbed with measured inference of 1–5 µs (logistic regression) or 10–25 µs (shallow MLP) and typical end-to-end service latency below 4 ms, satisfying the 10 ms Near-RT control budget for more than 95 percent of projected executions and showing that AI computation is not the bottleneck under this setup.

Load-bearing premise

The claim that lightweight AI can operate inside Near-RT timing rests on a rule-structured five-feature synthetic proxy task and a small RF-simulation testbed; if real traffic or larger models change the latency mix, the feasibility result may not transfer.

Editorial extensions

If this is right

  • Near-RT xApp designers can treat simple compiled classifiers as a negligible-cost inline step rather than an external service.
  • Model choice for this class of state classifiers can prioritize auditability and C-deployability over marginal accuracy gains.
  • Reproducible commodity-hardware stacks become a practical path for measuring AI inside the RIC loop without dedicated testbeds.
  • Future closed-loop control can start from the measured classification signal once E2SM-RC actuation is added.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same C-export pattern would likely keep inference negligible for other compact tabular models (small trees, linear SVMs) until feature dimension or depth forces dynamic allocation or larger matrix work.
  • If E2 transport remains the dominant cost, research effort on this path should shift from model compression toward message-path and scheduling optimizations.
  • A natural next measurement is multi-xApp co-residency: whether concurrent compiled inferences still leave headroom under the same 10 ms budget.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The paper demonstrates that lightweight supervised models (logistic regression and a shallow two-hidden-layer MLP) can be exported as deterministic C inference modules, compiled into a FlexRIC xApp binary, and executed inline on a live OAI/FlexRIC Near-RT RIC path. On a structured five-feature synthetic proxy for network-state classification (MAC/RLC/PDCP/GTP latencies + UE count), offline supervised accuracy clusters in 0.88–0.90; measured inference is 1–5 µs (LR) and 10–25 µs (MLP), typical end-to-end service latency stays below 4 ms, and both models are reported to meet the 10 ms Near-RT budget for >95% of projected loop executions. The contribution is scoped as embedding/execution feasibility and pipeline reproducibility (including the released RIC Workbench), not production generalization or closed-loop control.

Significance. The identified gap—measured, in-process, deterministic C inference with three-component latency instrumentation on a live OAI/FlexRIC stack—is real and under-served relative to simulation or external-inference xApp work. The offline-to-online export path, monotonic-clock instrumentation, noise ablation, and six-model offline comparison are concrete systems evidence that AI compute is negligible versus E2 transport under the reported conditions. Releasing RIC Workbench as a single-binary commodity orchestration tool is a genuine reproducibility contribution. If the measurements hold as stated, the paper supplies a useful template and lower bound for Near-RT RIC AI embedding rather than a new learning algorithm.

major comments (2)
  1. §IV-E and Fig. 8 (left): the claim that both models satisfy the 10 ms Near-RT budget for >95% of loop executions is based on a “projected” end-to-end loop that includes action dispatch, while §VI and the design text state that closed-loop control is not implemented and the xApp only classifies. The primary, fully measured result is E2 arrival + inference (and the reported typical service latency <4 ms). The manuscript should (i) report the measured E2+inference CDF as the main result, (ii) state explicitly and quantitatively what latency was assumed for the projected action-dispatch term, and (iii) keep “projected closed-loop” clearly labeled so the central feasibility claim does not rest on an unmeasured component.
  2. Abstract / §I / §VII vs. §III-C and §VI: the abstract and conclusion state that the results “demonstrate that lightweight AI can operate within Near-RT RIC timing constraints,” while the body correctly scopes the work as embedding feasibility on a rule-structured synthetic proxy and a 2-gNB/2-UE RFsim + Docker commodity stack. The transfer claim should be tightened in the abstract and conclusion to match the body (e.g., “under the reported OAI/FlexRIC path and proxy task”), so the headline does not outrun the experimental conditions that the limitations section already acknowledges.
minor comments (6)
  1. Table IV “C Budget” column mixes qualitative deployability (✓ / Partial / No) with latency ranges; define the criteria for Partial/No in the caption or text so the six-model comparison is reproducible.
  2. Fig. 8 caption and §IV-E: clarify whether the rare >100 ms service spikes are included in the plotted CDFs and in the “>95% under 10 ms” statistic, or whether that statistic is conditioned on non-preempted samples.
  3. §III-C: state the exact random seed and generation script version (or release them with RIC Workbench) so the 32k-sample synthetic set is bit-reproducible by others.
  4. Table VI peak inference latencies (<2 ms / <3 ms) are orders of magnitude above the typical 1–25 µs ranges; briefly explain what produces those peaks (e.g., first-call, cache, or measurement artifact) so they are not read as model compute.
  5. Related work (§II-C): a short explicit contrast table (Python/external vs. compiled C; presence/absence of per-indication monotonic timing) would sharpen the claimed gap versus ColO-RAN and other cited xApp demos.
  6. Minor polish: consistent hyphenation of Near-RT / Near-Real-Time; “µs” vs “us” in axis labels; ensure Fig. 3 console screenshot remains legible in print.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: measured latencies and scoped feasibility claims are independent of the synthetic proxy labels.

full rationale

The paper's central claim is measured embedding feasibility of deterministic C inference (LR 1–5 µs, MLP 10–25 µs; typical E2E service latency <4 ms; >95% of projected loops under the 10 ms Near-RT budget) on a live OAI/FlexRIC path, not production generalization. Inference and service latencies are obtained from monotonic-clock instrumentation inside the xApp callback on the live stack; they are not derived from fitted constants or from the synthetic label rules. Classification accuracy (0.88–0.90) is evaluated only on the authors' rule-structured five-feature synthetic proxy and is explicitly scoped as pipeline/embedding evidence rather than a first-principles prediction (Abstract; §III-C; §VI). The six-model comparison, noise ablation, and confusion matrices are likewise internal to that proxy and do not force the timing results. There is no self-definitional loop, no fitted parameter renamed as prediction of a related quantity, no load-bearing uniqueness theorem imported from the authors, and no ansatz smuggled via self-citation. Self-contained construction of synthetic labels is disclosed and does not make the measured microsecond timings tautological. Score 0 is therefore the correct outcome.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The load-bearing claim is empirical systems feasibility, not a derivation from first principles. It rests on O-RAN timing budgets and FlexRIC/OAI behavior as domain facts, on a hand-designed synthetic label model, and on several architecture/noise choices that define the proxy problem. No new physical entities; RIC Workbench and the four operational states are engineering constructs. Free parameters mainly shape the synthetic dataset and model capacity, not the latency stopwatch results.

free parameters (4)
  • Gaussian noise sigma for synthetic features = 15–30 µs (ablation 5–75 µs)
    σ = 15–30 µs (paper experiments use σ = 30 µs) is chosen to match observed FlexRIC indication jitter and to create class overlap; it directly sets classification difficulty.
  • MLP hidden architecture = 2 × 32 units
    Two hidden layers of 32 ReLU units is a hand-chosen capacity that fixes ~2300 ops/inference and the 10–25 µs latency band.
  • Synthetic dataset size and balance = 32000 samples, 4×8000
    32,000 samples, 8,000 per class, 80/20 split with fixed seed—chosen by authors to make accuracy a meaningful aggregate.
  • iperf3 load and indication rate = 10 Mbps DL/UE; ~1k ind/s
    10 Mbps DL per UE and ~1000 indications/s define the runtime measurement regime on the testbed.
assumptions (5)
  • domain assumption Near-RT RIC control loops must complete within a 10 ms–1 s budget (O-RAN Alliance timing model).
    Used throughout as the success criterion for CDFs and “not the bottleneck” discussion (§I, §II, §IV-E).
  • ad hoc to paper MAC/RLC/PDCP/GTP latencies plus UE count form a sufficient five-dimensional feature vector for the proxy state task.
    Feature choice and fusion are design decisions of §III-C/G; not derived from a standard theorem.
  • ad hoc to paper Four rule-structured labels (Healthy, Congestion Forming, User-Plane Stress, Control-Plane Stress) are valid operationally motivated proxy classes for embedding tests.
    Explicitly synthetic and conceptual (§III-C); accuracy only validates pipeline correctness on this construction.
  • domain assumption FlexRIC E2 indication path and OAI RFsim on commodity Linux are an acceptable software-path proxy for Near-RT timing characterization (not production OTA latency).
    Stated measurement scope in §III-E and §IV-A; spikes attributed to OS/Docker rather than model.
  • standard math Standard supervised learning (softmax LR, ReLU MLP, train/test split metrics) applies without further theoretical justification.
    Offline training via scikit-learn/NumPy and usual accuracy/F1 reporting (§III-G, §IV-B).
invented entities (2)
  • RIC Workbench
    purpose: Single-binary orchestration dashboard to run OAI CN5G, gNBs, FlexRIC, UEs, xApp IDE, and KPM panels on commodity hardware.
    Engineering artifact introduced to make the testbed reproducible; independent evidence would be a public binary/repo used by others (asserted release, no URL in text).
  • Four synthetic RAN operational states (labels 0–3)
    purpose: Provide a controlled multi-class proxy so inline C inference can be trained and scored without live ground truth.
    Defined by author rules over latency profiles and UE density; not observed live labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enabling Real-Time AI in O-RAN: Deploying and Measuring AI Inside a Near-RT RIC xApp." pith.science (2026). https://pith.science/paper/DXBV5VHC

@misc{pith2026260701583,
  author       = {Pith},
  title        = {Pith review of: Enabling Real-Time AI in O-RAN: Deploying and Measuring AI Inside a Near-RT RIC xApp},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DXBV5VHC}},
  note         = {Machine review of arXiv:2607.01583}
}
read the original abstract

Open Radio Access Network (O-RAN) architectures introduce programmable Near-Real-Time RAN Intelligent Controllers (Near-RT RICs) that support closed-loop control through xApps at timescales from 10 ms to 1 s. Although AI has been widely studied for RAN optimization, fewer works demonstrate measured AI inference embedded directly within the Near-RT RIC software loop on a live testbed. This paper presents an AI-enabled network-state classification xApp implemented on an OpenAirInterface (OAI) and FlexRIC testbed. The xApp is trained and evaluated on a structured synthetic dataset that emulates cross-layer RAN states using MAC, RLC, PDCP, GTP, and UE-count features. The results validate embedding and execution feasibility rather than production-level generalization. Logistic regression and a shallow multilayer perceptron (MLP) are exported as deterministic C inference modules and compiled into the xApp binary, eliminating external machine-learning runtime dependencies. Measured inference latency is 1 to 5 microseconds for logistic regression and 10 to 25 microseconds for the MLP, while end-to-end service latency remains below 4 ms. A six-model comparison shows that supervised models achieve similar accuracy, ranging from 0.88 to 0.90, indicating that LR and MLP similarity reflects the proxy problem structure rather than limited model exploration. Noise ablation, confusion-matrix analysis, and CDF-based latency characterization show that both embedded models satisfy the 10 ms Near-RT budget for more than 95% of projected loop executions. These results demonstrate that lightweight AI can operate within Near-RT RIC timing constraints while preserving deterministic execution. We also release RIC Workbench, a lightweight orchestration dashboard for reproducing the testbed on commodity hardware.

Figures

Figures reproduced from arXiv: 2607.01583 by the authors.

Figure 1
Figure 1. End-to-end architecture of the AI-enabled network state classification xApp within the O-RAN Near-RT RIC. Offline training uses structured synthetic [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. RIC Workbench setup tab showing all O-RAN stack components (CN5G, gNB, FlexRIC Near-RT RIC) running on a local workstation with one [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. RIC Workbench xApp IDE tab during live operation of [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Kernel density distribution of MAC, RLC, PDCP, and GTP latencies across Healthy (0), Congestion Forming (1), User-Plane Stress (2), and Control [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Confusion matrices for Logistic Regression (left) and MLP (right) on [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Macro F1 versus Gaussian noise level σ for all model families. The vertical dashed line marks σ = 30 µs, the noise level used in paper experiments. ML models maintain near-parity across the full sweep, confirming that LR– MLP similarity is structural. The Rule-Based ba…
Figure 8
Figure 8. Figure 8: Closed-loop latency characterization for the AI-enabled xApp. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references

  1. [1]

    What will 5G be?

    J. G. Andrews, S. Buzzi, W. Choi, S. V . Hanly, A. Lozano, A. C. K. Soong, and J. C. Zhang, “What will 5G be?”IEEE Journal on Selected Areas in Communications, vol. 32, no. 6, pp. 1065–1082, 2014

    2014

  2. [2]

    5G wire- less network slicing for eMBB, URLLC, and mMTC: A communication- theoretic view,

    P. Popovski, K. F. Trillingsgaard, O. Simeone, and G. Durisi, “5G wire- less network slicing for eMBB, URLLC, and mMTC: A communication- theoretic view,”IEEE Access, vol. 6, pp. 55 765–55 779, 2018

    2018

  3. [3]

    Ultrareliable and low-latency wireless communication: Tail, risk, and scale,

    M. Bennis, M. Debbah, and H. V . Poor, “Ultrareliable and low-latency wireless communication: Tail, risk, and scale,”Proceedings of the IEEE, vol. 106, no. 10, pp. 1834–1853, 2018

    2018

  4. [4]

    O-RAN architecture description,

    O-RAN Alliance, “O-RAN architecture description,” O-RAN Alliance, Tech. Rep., 2020

    2020

  5. [5]

    Under- standing O-RAN: Architecture, interfaces, algorithms, security, and re- search challenges,

    M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Under- standing O-RAN: Architecture, interfaces, algorithms, security, and re- search challenges,”IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 1376–1411, 2023

    2023

  6. [6]

    Intelli- gence and learning in O-RAN for data-driven NextG cellular networks,

    L. Bonati, S. D’Oro, M. Polese, S. Basagni, and T. Melodia, “Intelli- gence and learning in O-RAN for data-driven NextG cellular networks,” IEEE Communications Magazine, vol. 59, no. 2, pp. 21–27, 2021

    2021

  7. [7]

    Open, programmable, and virtualized 5G networks: State-of-the-art and the road ahead,

    L. Bonati, M. Polese, S. D’Oro, S. Basagni, and T. Melodia, “Open, programmable, and virtualized 5G networks: State-of-the-art and the road ahead,”Computer Networks, vol. 182, p. 107516, 2020

    2020

  8. [8]

    RIC: A RAN intelligent controller platform for AI-enabled cellular networks,

    B. Balasubramanian, E. S. Daniels, M. Hiltunen, R. Jana, K. Joshi, R. Sivaraj, T. X. Tran, and C. Wang, “RIC: A RAN intelligent controller platform for AI-enabled cellular networks,”IEEE Internet Computing, vol. 25, no. 2, pp. 7–17, 2021

    2021

Show all 21 references
  1. [9]

    Colo- ran: Developing machine learning-based xapps for open ran closed-loop control on programmable experimental platforms,

    M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Colo- ran: Developing machine learning-based xapps for open ran closed-loop control on programmable experimental platforms,”IEEE Transactions on Mobile Computing, vol. 22, no. 10, pp. 5787–5800, 2023

    2023

  2. [10]

    O-RAN working group overview,

    O-RAN Alliance, “O-RAN working group overview,” O-RAN Alliance, Tech. Rep., 2021

    2021

  3. [11]

    Artificial neural networks-based machine learning for wireless networks: A tutorial,

    M. Chen, U. Challita, W. Saad, C. Yin, and M. Debbah, “Artificial neural networks-based machine learning for wireless networks: A tutorial,” IEEE Communications Surveys & Tutorials, vol. 21, no. 4, pp. 3039– 3071, 2019

    2019

  4. [12]

    Deep learning for wireless communications: An emerging interdisciplinary paradigm,

    L. Dai, R. Jiao, F. Adachi, H. V . Poor, and L. Hanzo, “Deep learning for wireless communications: An emerging interdisciplinary paradigm,” IEEE Wireless Communications, vol. 27, no. 4, pp. 133–139, 2020

    2020

  5. [13]

    Applications of deep reinforcement learning in communications and networking: A survey,

    N. C. Luong, D. T. Hoang, S. Gong, D. Niyato, P. Wang, Y .-C. Liang, and D. I. Kim, “Applications of deep reinforcement learning in communications and networking: A survey,”IEEE Communications Surveys & Tutorials, vol. 21, no. 4, pp. 3133–3174, 2019

    2019

  6. [14]

    To- ward 6G networks: Use cases and technologies,

    M. Giordani, M. Polese, M. Mezzavilla, S. Rangan, and M. Zorzi, “To- ward 6G networks: Use cases and technologies,”IEEE Communications Magazine, vol. 58, no. 3, pp. 55–61, 2020

    2020

  7. [15]

    Machine learning paradigms for next-generation wireless networks,

    C. Jiang, H. Zhang, Y . Ren, Z. Han, K.-C. Chen, and L. Hanzo, “Machine learning paradigms for next-generation wireless networks,” IEEE Wireless Communications, vol. 24, no. 2, pp. 98–105, 2017

    2017

  8. [16]

    Edge intelligence: Paving the last mile of artificial intelligence with edge computing,

    Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, and J. Zhang, “Edge intelligence: Paving the last mile of artificial intelligence with edge computing,”Proceedings of the IEEE, vol. 107, no. 8, pp. 1738–1762, 2019

    2019

  9. [17]

    Team learning-based re- source allocation for open radio access network (O-RAN),

    H. Zhang, H. Zhou, and M. Erol-Kantarci, “Team learning-based re- source allocation for open radio access network (O-RAN),” inICC 2022 – IEEE International Conference on Communications, 2022, pp. 4938– 4943

    2022

  10. [18]

    Network- aided intelligent traffic steering in 6G O-RAN: A multi-layer optimiza- tion framework,

    V .-D. Nguyen, T. X. Vu, N. T. Nguyen, D. C. Nguyen, M. Juntti, N. C. Luong, D. T. Hoang, D. N. Nguyen, and S. Chatzinotas, “Network- aided intelligent traffic steering in 6G O-RAN: A multi-layer optimiza- tion framework,”IEEE Journal on Selected Areas in Communications, vol. ...

    2024

  11. [19]

    Demonstration of closed loop AI-driven RAN controllers using O-RAN SDR testbed,

    N. H. Stephenson, A. J. Chiejina, N. B. Kabigting, and V . K. Shah, “Demonstration of closed loop AI-driven RAN controllers using O-RAN SDR testbed,” inMILCOM 2023 – 2023 IEEE Military Communications Conference, 2023, pp. 241–242

    2023

  12. [20]

    FlexRIC: An SDK for next-generation SD-RANs,

    R. Schmidt, M. Irazabal, and N. Nikaein, “FlexRIC: An SDK for next-generation SD-RANs,” inProceedings of the 17th International Conference on Emerging Networking Experiments and Technologies (CoNEXT ’21). ACM, 2021, pp. 411–425

    2021

  13. [21]

    OpenAirInterface 5G software documentation,

    OpenAirInterface Software Alliance, “OpenAirInterface 5G software documentation,” 2023, https://www.openairinterface.org

    2023

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.