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

C. Nicolas Barati; Fahmida Afrin; Krzysztof J. Rechowicz; Lawrence Obiuwevwi; Muhammad Enayetur Rahman; Neda Moghim; Sachin Shetty; Safdar Hussain Bouk; Sampath Jayarathna; Valentina Nanou

Lightweight AI models run inference inside Near-RT RIC xApps in 1-25 microseconds while meeting the 10 ms timing budget.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.3

2026-07-03 04:42 UTC pith:DXBV5VHC

load-bearing objection The paper measures that LR and a small MLP compiled to C inside a Near-RT RIC xApp hit 1-25 μs inference and stay under 4 ms end-to-end on the OAI/FlexRIC testbed.

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

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

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

classification cs.NI

keywords O-RANNear-RT RICxAppAI inferencelatency measurementOpenAirInterfaceFlexRICreal-time control

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper demonstrates that simple supervised models can be compiled directly into an O-RAN xApp and execute on a live Near-RT RIC testbed without external runtimes. Logistic regression and a shallow multilayer perceptron classify emulated cross-layer network states from MAC, RLC, PDCP, GTP, and UE-count features. Measured inference takes 1-5 microseconds for logistic regression and 10-25 microseconds for the MLP, keeping end-to-end loop latency below 4 ms and satisfying the 10 ms constraint in over 95 percent of cases. Both models reach 0.88-0.90 accuracy on the synthetic data, and the work releases an orchestration dashboard for reproducing the setup on commodity hardware. This establishes feasibility of embedding deterministic AI inside the RIC software loop rather than proving broad generalization.

Core claim

An AI-enabled network-state classification xApp was built on an OpenAirInterface and FlexRIC testbed. Logistic regression and a shallow multilayer perceptron were exported as deterministic C inference modules and compiled into the xApp binary, removing external machine-learning dependencies. On a structured synthetic dataset emulating RAN states, inference latency measured 1-5 microseconds for logistic regression and 10-25 microseconds for the MLP, with end-to-end service latency below 4 ms. Both models satisfied the 10 ms Near-RT RIC budget for more than 95 percent of projected executions while preserving deterministic behavior.

What carries the argument

Deterministic C inference modules for logistic regression and multilayer perceptron compiled directly into the xApp binary to eliminate external runtime dependencies.

Load-bearing premise

The structured synthetic dataset that emulates cross-layer RAN states using MAC, RLC, PDCP, GTP, and UE-count features is representative enough for the timing and feasibility results to hold outside the testbed.

What would settle it

Deploy the identical xApp on a production O-RAN setup with live user equipment and real traffic, then check whether inference latency exceeds 10 ms or accuracy drops below 0.85 in more than 5 percent of loop executions.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

xApps can incorporate supervised AI without introducing external machine-learning runtime dependencies.
End-to-end service latency remains below 4 ms while inference stays in the microsecond range.
Both logistic regression and the shallow MLP satisfy the 10 ms Near-RT budget for over 95 percent of executions.
A six-model comparison shows supervised models reach 0.88-0.90 accuracy, with similarity tied to the proxy problem structure.
The released RIC Workbench enables reproduction of the testbed on commodity hardware.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same compilation approach could extend to other xApp functions or to Non-RT RIC rApps if similar timing headroom exists.
Validation on live RAN traffic rather than synthetic data would be required before claiming production readiness.
The microsecond inference times leave margin for modestly deeper models if the accuracy gap justifies added compute.
Releasing the orchestration dashboard lowers the barrier for other groups to test embedded AI loops in O-RAN environments.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

The paper measures that LR and a small MLP compiled to C inside a Near-RT RIC xApp hit 1-25 μs inference and stay under 4 ms end-to-end on the OAI/FlexRIC testbed.

read the letter

The main thing to know is that the authors embedded logistic regression and a shallow MLP as C code inside the xApp on their OAI/FlexRIC testbed and measured inference at 1-5 μs and 10-25 μs respectively, with end-to-end service under 4 ms. That puts them comfortably inside the 10 ms Near-RT window on this setup.

They trained on synthetic cross-layer features, exported the models to avoid runtime dependencies, and ran noise ablation plus latency characterization. The six-model accuracy check is useful to show the results aren't model-specific. Releasing RIC Workbench is a practical plus for reproduction on commodity hardware.

The measurements are direct and the code paths are fixed-size, so the timing claims don't depend on the training data contents. The paper scopes itself to feasibility on the testbed rather than real-world transfer.

The soft spot is the synthetic dataset. Without real traffic validation, it's hard to know if the models would classify actual states the same way or if the timing would hold under variable loads. The similar accuracies (0.88-0.90) across models also hint that the classification task may be straightforward.

This work is for engineers and researchers focused on O-RAN xApp development who need concrete timing numbers. It has enough testbed grounding to merit peer review, even if the generalization step needs more data in a revision.

Referee Report

0 major / 2 minor

Summary. The paper presents an AI-enabled network-state classification xApp implemented on an OpenAirInterface (OAI) and FlexRIC testbed for the Near-RT RIC. Logistic regression and a shallow MLP are trained on a structured synthetic dataset emulating cross-layer RAN states (MAC, RLC, PDCP, GTP, UE-count features), exported as deterministic C inference modules, and compiled into the xApp binary. Direct measurements show inference latencies of 1-5 μs (LR) and 10-25 μs (MLP), end-to-end service latency below 4 ms, and compliance with the 10 ms Near-RT budget for more than 95% of projected loop executions. A six-model comparison yields accuracies of 0.88-0.90. The work releases the RIC Workbench for testbed reproduction and scopes results to embedding feasibility on this testbed rather than production generalization.

Significance. If the reported measurements hold, the paper provides concrete testbed evidence that lightweight supervised models can be embedded directly in the Near-RT RIC xApp loop using deterministic C implementations that eliminate external ML runtimes and satisfy timing constraints. The direct latency measurements (independent of training data contents) and public release of the RIC Workbench for reproduction on commodity hardware are clear strengths that enhance the contribution's utility for the O-RAN community.

minor comments (2)

[Abstract] Abstract: the six-model comparison reports accuracy 0.88-0.90 but does not name the other four models; adding their identities would clarify why LR and MLP similarity is attributed to problem structure.
The CDF-based latency characterization and 95% compliance claim would be strengthened by an explicit equation or pseudocode showing how 'projected loop executions' are computed from the measured service latencies.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the thorough summary and positive recommendation to accept the manuscript. The report accurately captures the scope, contributions, and limitations of the work.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central claims rest on direct testbed measurements of inference latency (1-5 μs LR, 10-25 μs MLP), end-to-end service latency (<4 ms), and accuracy (0.88-0.90) for C-compiled deterministic modules. These timings derive from fixed-size code paths and compilation, independent of training data contents or any fitted parameters. No equations, self-citations, or ansatzes are invoked as load-bearing steps for the feasibility conclusions. The synthetic dataset affects only model weights, not the reported execution metrics. This is a standard non-circular empirical measurement paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that the synthetic dataset captures enough structure of real RAN states for timing feasibility to be meaningful; no free parameters or invented entities are introduced.

axioms (1)

domain assumption The synthetic dataset emulating cross-layer RAN states is representative for feasibility demonstration
Training and evaluation are performed exclusively on this dataset; the timing conclusions are presented as evidence for real deployments.

pith-pipeline@v0.9.1-grok · 5899 in / 1281 out tokens · 28067 ms · 2026-07-03T04:42:21.804326+00:00 · methodology

0 comments

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 C. Nicolas Barati, Fahmida Afrin, Krzysztof J. Rechowicz, Lawrence Obiuwevwi, Muhammad Enayetur Rahman, Neda Moghim, Sachin Shetty, Safdar Hussain Bouk, Sampath Jayarathna, Valentina Nanou.