Intent-Based Network for RAN Management with Large Language Models

Chun-Kai Lai; Edwin K. P. Chong; Fransiscus Asisi Bimo; Maria Amparo Canaveras Galdon; Ray-Guang Cheng

REVIEW 4 major objections 5 minor 2 cited by

Intent-Based Network for RAN Management with Large Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read An LLM-orchestrated agent loop translates a formalized network intent into transmit-power settings that move simulated RAN energy efficiency toward the requested target, converging in two to four iterations and correctly declaring…

desk verdict A plausible proof-of-concept for LLM-driven RAN intent management, but the validation is a single un-replicated simulator run and the authors' own PM-accuracy admission undermines the central claim as stated. read the letter →

arxiv 2507.14230 v2 pith:JSWCTHM7 submitted 2025-07-17 cs.NI cs.AI

Fransiscus Asisi Bimo , Maria Amparo Canaveras Galdon , Chun-Kai Lai , Ray-Guang Cheng , Edwin K. P. Chong This is my paper · ORCID

classification cs.NIcs.AI

keywords intent-basednetworkinglargelanguagemodelsradioaccessnetworkclosed-loopcontrolenergyefficiencyO-RANagenticAItransmitpoweroptimization

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 tries to establish that a large language model can act as the translation and decision layer of an intent-based radio access network: a high-level goal such as raising energy efficiency above 800,000 bits per joule for a specific cell is converted into concrete configuration actions, and a closed loop keeps adjusting until the goal is met. The demonstration case is energy efficiency on a simulated two-cell RAN, where the LLM-orchestrated loop lowered transmit power from about $30$ dBm toward $11$ dBm and raised PEE.EnergyEfficiency from roughly $7.7\times 10^5$ to about $8\times 10^5$ bits per joule. Target values set near this baseline were reached in two to four iterations, while a target beyond the reachable power range was correctly declared a failure when the adjustment plateau was reached. If the claim is right, LLMs can close the intent-translation gap that has limited intent-based networking, turning formalized objectives into continuously re-tuned RAN configurations.

What carries the argument

The load-bearing mechanism is the structured prompt that forces the LLM to treat intent translation as a constrained decision: five named sections (Instruction, Intent, Current Observation, Configuration Constraints, Output Format) turn a natural-language goal plus current network measurements into one parseable configuration action per iteration. Around that prompt, a two-agent orchestration flow uses past strategy attempts as context and closes the loop by reading performance-management data after each NETCONF configuration edit, so the next prompt sees the effect of the previous action. The important property is that the prompt converts an open-ended language model into an incremental controller that lowers or raises transmit power by small steps, which is what makes the loop converge instead of drift.

What would settle it

Re-run the seven-attempt precision test while injecting a small offset (for example, +3 percent) into the energy-efficiency values delivered to the agent, leaving the simulator's true state unchanged. If the loop still converges to the true target, the demonstration is robust to measurement error; if it oscillates, diverges, or marks attempts as failures prematurely, the convergence claim depends on trustworthy measurements.

Watch

Extended reading notes

Core claim

The central claim is that the network can automatically improve its energy efficiency by dynamically optimizing critical RAN parameters through a closed-loop mechanism driven by LLM-orchestrated agents. In the reported experiment, a Strategist Agent parses a formalized JSON intent modeled on 3GPP TS 28.312, reads the current configuration and measurement data over an O-RAN O1 interface, and issues an incremental transmit-power edit through NETCONF; the resulting performance measurement is then fed back into the next decision. Across attempts with target values of $800$, $803$, and $805$ thousand bits per joule, the loop converged in $2$--$4$ iterations, and a seventh attempt stopped with a "Fail" outcome once the allowed transmit-power range was exhausted. The discovery is that an LLM, constrained by a structured prompt and supplied with history, can behave as a stable incremental optimizer for RAN configuration rather than a one-shot generator of settings.

Load-bearing premise

The loop assumes the measurement values returned by the simulator are accurate enough to guide control; the authors report that aggregated measurement values often do not reflect the network's true state and that inaccurate values can create a feedback loop of misinformation.

Editorial extensions

If this is right

If the central claim holds, catalog-based intent translation is not required: an LLM can interpret a formalized intent and emit actionable RAN configuration changes directly.
A convergence time of two to four iterations means an intent can be fulfilled and verified within a few performance-measurement cycles, making closed-loop assurance practical.
When the transmit-power range is exhausted, the system stops and reports the intent as unfulfillable, so an autonomous controller can avoid endless or destabilizing retries.
For targets already satisfied by the current state, the system performs monitoring and verification only, functioning as an intent-assurance check as well as an optimizer.

Reading between the lines

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

The paper leaves implicit that caching the history agent's successful strategy trajectories could warm-start later attempts at the same intent, likely reducing iteration counts below the reported two to four.
A testable extension suggested by the authors' measurement-accuracy concern is to feed the LLM raw, unaggregated measurement streams; the expected effect is fewer misinformation loops but larger prompt sizes.
Because the experiment uses static user equipment and two cells, an unstated next question is whether the same five-section prompt remains stable when mobility and inter-cell interference corrupt the measurement-to-configuration mapping.

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.

Desk Editor's Note

read the letter

The paper is a proof-of-concept for using a two-agent LLM loop over O-RAN O1 to translate intents into RAN parameter changes, with energy efficiency as the test case. What's genuinely new is the concrete integration: a Strategist agent and a History Analyzer agent, a five-section prompt, and NETCONF for configuration with PM feedback. That combination is not in the cited prior work. The authors also deserve credit for Section IV, where they openly admit the PM aggregation issue and the risk of a 'feedback loop of misinformation.'

The soft spot is the validation, which is very thin. The central demonstration is a single simulator run—no repeated trials, no error bars, no baseline comparison, no released code or exact prompts. The unit inconsistency is telling: Fig. 3 uses bits/joule, Table III uses kWh. That makes it hard to know what quantity was actually optimized. And the authors' own admission in Section IV-B undermines the load-bearing assumption: if the aggregated PM data don't reflect true network state, the reported convergence could be convergence to a distorted metric. They offer streaming PM as future work, which confirms the current results used the questionable aggregated path. There's also no quantitative measure of intent translation accuracy; the LLM is essentially given the target in the prompt and judged by whether it reaches that target, which is circular in a practical sense.

That said, the central claim isn't refuted. The architecture is plausible and the direction is timely. But as it stands, the evidence doesn't support the abstract's assertion that the network 'automatically improves its energy efficiency' in a robust sense. This is a workshop-grade prototype paper, not a rigorous journal result.

Who should read it: anyone working on intent-based networking, LLM agents for O-RAN, or RAN automation. It's a useful reference for the architectural sketch and the honest limitations. I'd send it to peer review rather than desk reject—the topic is important and a serious referee could push the authors to fill the gaps—but the expectation should be major revision, with replication, baselines, and a resolution of the PM accuracy question. For my own work, I wouldn't cite it as evidence yet; it's a pointer to a direction, not a verified result.

Referee Report

4 major / 5 minor

Summary. The paper proposes an intent-based network (IBN) architecture for RAN management in which large language models (LLMs), embedded in a Strategist Agent and a History Analyzer Agent orchestrated with LangGraph, translate formalized intents (JSON per 3GPP TS 28.312) into TxPower configurations via the O-RAN O1 interface. The authors introduce a five-section prompt structure and report a single simulator experiment (VIA VI AI RSG) in which TxPower decreases from ~30 dBm to ~11 dBm while a PEE.EnergyEfficiency metric rises to ~8×10^5 bits/joule, plus a closed-loop precision test with seven attempts and a boundary test in Table III. The paper claims that the network can automatically improve energy efficiency through closed-loop LLM-orchestrated adaptation, but it also acknowledges in Section IV that aggregated PM data may be inaccurate and can cause a feedback loop of misinformation.

Significance. If the claimed results hold, the work would provide a concrete, standard-interface-based demonstration of LLM-driven intent translation and closed-loop RAN optimization, with strengths including a clear agentic decomposition, a structured prompt design, use of the O-RAN O1 interface for CM/PM, and an explicit discussion of contextual and data-accuracy limitations. The paper does not provide code, data, or a reproducible experimental pipeline, and the validation evidence is currently too limited to support the general claim of automatic energy-efficiency improvement. The authors honestly flag a critical vulnerability in their own data source, but they do not show that the reported runs were unaffected by that vulnerability.

major comments (4)

[Section IV-B vs. Sections II-D and III] The authors state in Section IV-B that aggregated PM values "often do not accurately reflect the network's true state" and can trigger a "feedback loop of misinformation," while Section II-D describes retrieving PM via the O1 interface with fixed-interval transmission and aggregation at the gNB, and Section III gives no indication that streaming PM was used in the reported runs. Since the convergence shown in Figs. 3 and 4 is judged on exactly these PM values, the central demonstration is valid only if the PM data in those runs were accurate. Please report the actual aggregation settings and time windows used, validate the PM values against the simulator's ground-truth state, or repeat the experiments with streaming PM to rule out the self-identified failure mode.
[Section III-A, Fig. 3] Figure 3 reports a single trajectory from one un-replicated simulator run, with no error bars, no multiple seeds, and no baseline comparison (e.g., a rule-based policy, a simple hill-climbing controller, or a fixed-power scheme). One run cannot support the abstract's claim that "the network can automatically improve its energy efficiency" through the proposed method. The authors should provide multiple independent runs with variance, and compare the LLM-driven loop against at least one non-LLM baseline to demonstrate that the agentic architecture is responsible for the improvement.
[Table III and Figs. 3-4] The energy-efficiency metric is inconsistent across the experimental sections: Fig. 3 and the intent example use bits/joule with a target around 800,000, and Section III-B refers to targets 800, 803, and 805 without units, while Table III labels PEE.EnergyEfficiency in "KWh" with values 765.4, 750, and 803. As written, it is unclear whether Table III and the Fig. 4 targets describe the same quantity on the same scale, and the "KWh" unit is not appropriate for an energy-efficiency metric. Please unify the units, rescale consistently, and restate the boundary-test results in the same metric used in the optimization.
[Sections II-B and III] The paper's primary contribution is described as enhancing intent translation through a structured prompt, but no quantitative metric for translation accuracy is reported (e.g., parse success rate, semantic correctness of generated strategies, or alignment between the LLM output and a ground-truth configuration). The presented loop evaluation only checks whether the metric reaches the target value that was already provided in the prompt, so it conflates control performance with translation quality. Please add a direct evaluation of translation quality and an ablation that isolates the effect of the proposed prompt structure.

minor comments (5)

[General] The manuscript contains several grammatical and typographical issues; for example, Section III-A says "for iteration(s)" and Section III-B says "finally is 805." A careful language revision is needed.
[Figure 1] Figure 1 appears to be a low-resolution screenshot with small text; it should be redrawn as a vector diagram with readable labels so that the agentic architecture and the O1 interface connections are clearly visible.
[Section II-B] The description of one-shot prompting would benefit from showing the actual one-shot example in the prompt, rather than only describing the five-section structure, so that the structured technique can be reproduced.
[Section III-B] The text says the first five attempts target 800, the sixth targets 803, and the seventh targets 805, but it does not state whether these values have the same units as the 800,000 bits/joule in the example intent; please clarify the scale and the precision of the comparison.
[Section III-C] The baseline value in Table III is described as "recorded initial simulation data," but the text does not state how many samples or over what time window; please specify the measurement procedure and the reported value's units.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical closed-loop demonstration with no fitted parameters or derivation that reduces to its inputs.

full rationale

The paper makes no mathematical derivation and fits no parameters to data; its contribution is an experimental demonstration that an LLM-orchestrated loop can adjust TxPower to move a reported KPI toward a requested target in a RAN simulator. The target KPI is supplied as an input to the controller and the loop is evaluated against that same target, which is standard feedback-control evaluation rather than a fitted input being renamed as a prediction. The self-citations (e.g., references [1] and [12]) are background material about Open RAN surveys and a research platform; they are not load-bearing for the central empirical claim, which rests on an external simulator (VIA VI AI RSG) and an external LLM inference API (NVIDIA NIM). The admitted PM-data inaccuracy in Section IV-B is a correctness and reliability limitation, not a circularity: it questions whether the reported convergence reflects the true network state, but it does not show that any result is equivalent to its input by construction. No quoted step exhibits the required specific reduction, so the appropriate finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper introduces no fitted constants, but it rests on domain assumptions about simulator fidelity, PM data accuracy, LLM output validity, and the monotonic effect of TxPower on energy efficiency. Section IV-B explicitly concedes that the PM-data assumption is fragile. No new physical or conceptual entities are postulated.

assumptions (4)

domain assumption PEE.EnergyEfficiency is a valid, aggregation-faithful measure of RAN energy efficiency.
Used throughout Section III; Section IV-B states aggregated PM data is often inaccurate, undermining this.
domain assumption The VIA VI AI RSG simulator behaves like a real RAN for closed-loop control.
All experiments use this simulator; no validation against a real RAN or standard model is provided.
domain assumption The LLM's parsed JSON configuration is always syntactically valid and safely applicable.
The pipeline directly issues edit-config based on LLM output; no validation or fallback is described.
domain assumption Lowering TxPower strictly improves energy efficiency in the test scenario.
Interpretation of Fig. 3 relies on this monotonic relationship; no analytic model is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Intent-Based Network for RAN Management with Large Language Models." pith.science (2026). https://pith.science/paper/JSWCTHM7

@misc{pith2026250714230,
  author       = {Pith},
  title        = {Pith review of: Intent-Based Network for RAN Management with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JSWCTHM7}},
  note         = {Machine review of arXiv:2507.14230}
}

read the original abstract

Advanced intelligent automation becomes an important feature to deal with the increased complexity in managing wireless networks. This paper proposes a novel automation approach of intent-based network for Radio Access Networks (RANs) management by leveraging Large Language Models (LLMs). The proposed method enhances intent translation, autonomously interpreting high-level objectives, reasoning over complex network states, and generating precise configurations of the RAN by integrating LLMs within an agentic architecture. We propose a structured prompt engineering technique and demonstrate that the network can automatically improve its energy efficiency by dynamically optimizing critical RAN parameters through a closed-loop mechanism. It showcases the potential to enable robust resource management in RAN by adapting strategies based on real-time feedback via LLM-orchestrated agentic systems.

Figures

Figures reproduced from arXiv: 2507.14230 by the authors.

**Figure 2.** User prompt structure for LLM agent agents is configured within the Prompt Processor component, which utilizes the LangGraph library 1 . TABLE I LLM INFERENCE PARAMETERS Parameter Value Model llama-3.1-70b-instruct Temperature 0.2 Top P 0.7 Max Tokens 1024 Stream True 1https://github.com/langchain-ai/langgraph [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Correlation between configured TxPower and PEE.EnergyEfficiency [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Closed-loop precision evaluation for an IBN System [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Agentic-V2X: Small Language Model Agents for Deadline-Aware V2X Scheduling in 5G/6G Networks
cs.NI 2026-07 conditional novelty 5.5 of 10

A validated small-LLM rApp-style policy layer plus a 100 ms deterministic xApp produces executable deadline-aware V2X scheduler weights that are competitive at high density but not best overall.
Agentic World Modeling for 6G: Near-Real-Time Generative State-Space Reasoning
cs.NI 2025-11 conditional novelty 5.0 of 10

WM-MS3M adds a compact stochastic latent and dual decoders to a causal multi-scale SSM, improving KPI MAE by 1.69% over MS3M with 32% fewer parameters and enabling PRB what-if rollouts.

Reference graph

Works this paper leans on

14 extracted references · 13 canonical work pages · cited by 2 Pith papers

[1]

A survey on open radio access networks: Challenges, research directions, and open source approaches,

W. Azariah, F. A. Bimo, C.-W. Lin, R.-G. Cheng, N. Nikaein, and R. Jana, “A survey on open radio access networks: Challenges, research directions, and open source approaches,”Sensors, vol. 24, no. 3, p. 1038, Feb. 2024. [Online]. Available: http://dx.doi.org/10. 3390/s24031038

work page 2024
[2]

A survey on intent-based networking,

A. Leivadeas and M. Falkner, “A survey on intent-based networking,” IEEE Communications Surveys & Tutorials, vol. 25, no. 1, pp. 625–655, 2023

work page 2023
[3]

Intent driven management

G. W. SA5, “Intent driven management.” [Online]. Available: https://www.3gpp.org/technologies/intent

work page
[4]

Intent- Based Networking - Concepts and Definitions,

A. Clemm, L. Ciavaglia, L. Z. Granville, and J. Tantsura, “Intent- Based Networking - Concepts and Definitions,” RFC 9315, Oct. 2022. [Online]. Available: https://www.rfc-editor.org/info/rfc9315

work page 2022
[5]

Tm forum introductory guide autonomous networks tech- nical architecture;,

TMForum, “Tm forum introductory guide autonomous networks tech- nical architecture;,” TMForum, Tech. Rep. Version 1.1.1, 2023

work page 2023
[6]

Open RAN Security Report,

National Telecommunications and Information Administration (NTIA), “Open RAN Security Report,” National Telecommunications and Information Administration, U.S. Department of Commerce, Tech. Rep., 2023. [Online]. Available: https://www.ntia.gov/sites/default/files/ publications/open_ran_security_report_full_report_0.pdf

work page 2023
[7]

Llama: Open and efficient foundation language models,

H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample, “Llama: Open and efficient foundation language models,” 2023. [Online]. Available: https: //arxiv.org/abs/2302.13971

arXiv 2023
[8]

Large language models for software engineering: Survey and open problems,

A. Fan, B. Gokkaya, M. Harman, M. Lyubarskiy, S. Sengupta, S. Yoo, and J. M. Zhang, “Large language models for software engineering: Survey and open problems,” in2023 IEEE/ACM International Confer- ence on Software Engineering: Future of Software Engineering (ICSE- FoSE), 2023, pp. 31–53

work page 2023

Show all 14 references

[9]

Applying routenet and lstm to achieve network automation: An intent-based networking approach,

T. Ahmed Khan, K. Abbas, J. J. Diaz Rivera, A. Muhammad, and W.-c. Song, “Applying routenet and lstm to achieve network automation: An intent-based networking approach,” in2021 22nd Asia-Pacific Network Operations and Management Symposium (APNOMS), 2021, pp. 254– 257

2021
[10]

Towards intent-based network management for the 6g system adopting multimodal generative ai,

D. Brodimas, K. Trantzas, B. Agko, G. C. Tziavas, C. Tranoris, S. Denazis, and A. Birbas, “Towards intent-based network management for the 6g system adopting multimodal generative ai,” in2024 Joint European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit...

2024
[11]

Intent-based networking: Current advances, open challenges, and future directions,

M. Gharbaoui, B. Martini, and P. Castoldi, “Intent-based networking: Current advances, open challenges, and future directions,” in2023 23rd International Conference on Transparent Optical Networks (ICTON), 2023, pp. 1–5

2023
[12]

Design and implementation of next-generation research platforms,

F. A. Bimo, R.-G. Cheng, C.-C. Tseng, C.-R. Chiang, C.-H. Huang, and X.-W. Lin, “Design and implementation of next-generation research platforms,” in2023 IEEE Globecom Workshops (GC Wkshps), 2023, pp. 1777–1782

2023
[13]

langgraph

LangChain, “langgraph.” [Online]. Available: https://langchain-ai. github.io/langgraph/
[14]

Lte; 5g; management and orchestration; intent driven man- agement services for mobile networks (3gpp ts 28.312 version 18.6.0 release 18),

3GPP, “Lte; 5g; management and orchestration; intent driven man- agement services for mobile networks (3gpp ts 28.312 version 18.6.0 release 18),” 3GPP, Tech. Rep. ETSI TS 128 312 V18.6.0 (2025-01), 2025

2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

A survey on open radio access networks: Challenges, research directions, and open source approaches,

W. Azariah, F. A. Bimo, C.-W. Lin, R.-G. Cheng, N. Nikaein, and R. Jana, “A survey on open radio access networks: Challenges, research directions, and open source approaches,”Sensors, vol. 24, no. 3, p. 1038, Feb. 2024. [Online]. Available: http://dx.doi.org/10. 3390/s24031038

work page 2024

[2] [2]

A survey on intent-based networking,

A. Leivadeas and M. Falkner, “A survey on intent-based networking,” IEEE Communications Surveys & Tutorials, vol. 25, no. 1, pp. 625–655, 2023

work page 2023

[3] [3]

Intent driven management

G. W. SA5, “Intent driven management.” [Online]. Available: https://www.3gpp.org/technologies/intent

work page

[4] [4]

Intent- Based Networking - Concepts and Definitions,

A. Clemm, L. Ciavaglia, L. Z. Granville, and J. Tantsura, “Intent- Based Networking - Concepts and Definitions,” RFC 9315, Oct. 2022. [Online]. Available: https://www.rfc-editor.org/info/rfc9315

work page 2022

[5] [5]

Tm forum introductory guide autonomous networks tech- nical architecture;,

TMForum, “Tm forum introductory guide autonomous networks tech- nical architecture;,” TMForum, Tech. Rep. Version 1.1.1, 2023

work page 2023

[6] [6]

Open RAN Security Report,

National Telecommunications and Information Administration (NTIA), “Open RAN Security Report,” National Telecommunications and Information Administration, U.S. Department of Commerce, Tech. Rep., 2023. [Online]. Available: https://www.ntia.gov/sites/default/files/ publications/open_ran_security_report_full_report_0.pdf

work page 2023

[7] [7]

Llama: Open and efficient foundation language models,

H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample, “Llama: Open and efficient foundation language models,” 2023. [Online]. Available: https: //arxiv.org/abs/2302.13971

arXiv 2023

[8] [8]

Large language models for software engineering: Survey and open problems,

A. Fan, B. Gokkaya, M. Harman, M. Lyubarskiy, S. Sengupta, S. Yoo, and J. M. Zhang, “Large language models for software engineering: Survey and open problems,” in2023 IEEE/ACM International Confer- ence on Software Engineering: Future of Software Engineering (ICSE- FoSE), 2023, pp. 31–53

work page 2023