PaperlayerView exact revision on arXiv

Source-linked AI summary

Intent-Based Network for RAN Management with Large Language Models

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

arXiv:2507.14230v2cs.NIcs.AI

TL;DR

RAN management faces growing complexity, and accurate translation of high-level intents into network instructions remains important for autonomous operation. The paper uses an agentic LLM and structured prompts to translate intents and adjust RAN parameters in a closed loop, reporting improved energy efficiency and convergence toward different target values.

  • Problem

    RAN management is increasingly complex, and intent-based operation depends on accurately translating high-level goals into RAN instructions because misinterpretation can impair resource use, QoS, or network stability.

    Source: §I · p. 1

  • Method

    The paper integrates LLMs into an agentic RAN management system that interprets formalized intents, develops configuration strategies, and uses closed-loop feedback to adjust parameters.

    Source: §I · p. 2, §II · p. 2

  • Results

    The closed-loop evaluation reports increasing energy efficiency as transmit power falls, and successful convergence across attempts with varying target values.

    Source: A. Single-Run Evaluation · p. 4, B. Closed-Loop Precision Evaluation · p. 4

  • Takeaways & Limitations

    The evaluation demonstrates that an LLM-orchestrated closed loop can dynamically optimize RAN parameters for an energy-efficiency use case.

    Source: §I · p. 2, A. Single-Run Evaluation · p. 4

    Aggregated PM values can misrepresent the network’s true state because results depend on aggregation algorithms and configurable time windows.

    Source: B. The Impact of PM Data Inaccuracy · p. 5

Abstract

from arXiv · show

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.

I. INTRODUCTION

RAN management is becoming more complex, while intent-based networking aims to replace detailed configuration with high-level goals. The paper proposes an LLM-based agentic approach to translate those goals and optimize RAN parameters through closed-loop operation.

I. INTRODUCTION

  • RAN service diversity and dynamic resource demands increase operational complexity, while Open RAN can add configuration complexity and misconfiguration risk to manual low-level procedures.The paper notes that manual configuration makes agility and scalability difficult.Source: §I · p. 1
  • An intent expresses high-level network goals rather than prescribed methods, with expectations that can include requirements, goals, conditions, and constraints.IBN implements and manages these intents, abstracting underlying network complexity,.Source: §I · p. 1
  • LLMs are proposed as intermediaries that interpret formalized intents and generate granular RAN configurations using their natural-language processing capabilities.The paper motivates this role by pointing to LLMs’ semantic understanding of context and service requirements,.Source: §I · p. 1
  • Accurate, adaptable translation of natural-language intents into RAN instructions is crucial because misinterpretation can cause inefficient resource use, poor QoS, or network instability.Intent translation is identified as the most critical and complex component of the five-part IBN closed loop.Source: §I · p. 1
  • Existing approaches leave a translation gap: catalog-based methods lack LLM semantic understanding, while the Generative AI approach in lacks closed-loop self-adaptation.The paper presents translation as the step that turns high-level goals and requirements into actionable network policies.Source: §I · p. 2
  • The paper proposes integrating LLMs into an intent-based wireless network to enhance intent translation and dynamically optimize RAN parameters for energy efficiency through a closed loop.Its approach uses a structured prompt technique and presents an energy-efficiency use case.Source: §I · p. 2

II. PROPOSED ARCHITECTURE

The proposed RAN management system uses an agentic LLM to translate formalized intents into configuration strategies. Agents can retrieve network metrics and execute configuration through the O-RAN O1 interface.

II. PROPOSED ARCHITECTURE

  • The system uses cell configurations and measurements from an O-RAN experimental platform, and its RAN connection interface adheres to O-RAN specifications.The paper describes the LLM-driven initial configuration and optimization as setting the stage for efficient operation, while closed-loop monitoring may use deterministic approaches.Source: §II · p. 2
  • A Strategist Agent uses an LLM to translate the consumer’s formalized intent, while an Orchestrator Agent executes configuration strategies through the O-RAN O1 interface.The Strategist may use successful past attempts stored by the History Analyzer Agent when formulating a strategy.Source: §II · p. 2
  • The agent analyzes target objects, metrics, conditions, and values, then decomposes the objective into actionable strategies and plans steps through a LangGraph flow.Planning may include retrieving cell details and performance metrics from the RAN simulator via O1.Source: §II · p. 2

A. Formalized Intent Structure

The owner’s intent is formalized as a JSON object and parsed into a structured context containing target KPIs and the network objects to which they apply.

A. Formalized Intent Structure

  • The intent context separates Target KPIs, which specify metrics and desired conditions or values, from Object Target, which identifies the network elements or scope.The declaration is formalized as a JSON object referencing 3GPP TS 28.312.Source: A. Formalized Intent Structure · p. 2
  • An example intent targets energy efficiency for Cell 7 in SubNetwork 1, with a desired value greater than 800,000 bits per joule.The target condition and value are represented alongside the object target in the formalized intent.Source: A. Formalized Intent Structure · p. 2
  • The structured intent representation serves as the initial input to the agentic LLM.Source: A. Formalized Intent Structure · p. 2

B. Prompt Techniques

The system uses one-shot prompting with a five-section structure to provide context and directives for consistent, predictable LLM-driven intent translation.

B. Prompt Techniques

  • One-shot prompting supplies an example of the expected output, and the five-section prompt is intended to standardize input and guide consistent LLM behavior.The prompt is framed as supporting reliable automated operation of the RAN intent-based network.Source: B. Prompt Techniques · p. 2
  • The prompt’s Instruction states the task, while Intent supplies the system-level objective or desired outcome.Source: B. Prompt Techniques · p. 3
  • Current Observation provides relevant operational data about network state, and Configuration Constraints specify which parameters and strategies are available for each RAN gNB.Source: B. Prompt Techniques · p. 3
  • Output Format specifies the response structure so that it can be parsed for further computation.Source: B. Prompt Techniques · p. 3

C. Agentic Framework

The framework uses two LLM agents to combine network measurements and prior strategy outcomes in configuration decisions. A structured prompt and LangGraph-based prompt processor support the agentic system.

C. Agentic Framework

  • The History Analyzer reviews real-time measurements and past strategy attempts, then recommends strategy outcomes to the Strategist Agent.The Strategist uses this processed data and relevant prior attempts for decision-making.Source: C. Agentic Framework · p. 3
  • A five-section user prompt standardizes input for LLM-driven intent translation.The figure identifies the user prompt structure for the LLM agent.Source: Figure 2 · p. 3
  • The agents’ underlying intelligence is configured in a Prompt Processor that uses LangGraph.Table I is titled “LLM Inference Parameters.”Source: C. Agentic Framework · p. 3, Table I · p. 3

D. O1 Configuration & Performance Management

Configuration Management and Performance Management over O-RAN O1 provide the system’s closed-loop network control. The experimental evaluation uses the VIAVI Solutions AI RSG RAN Simulator with parameters listed in Table II.

D. O1 Configuration & Performance Management

  • The agents retrieve real-time PM data and current configuration, then use NETCONF <edit-config> operations to apply optimized RAN settings.They retrieve current configuration with <get-config> and use intent and historical performance context to guide changes.Source: D. O1 Configuration & Performance… · p. 4
  • The intent-based wireless network management system is shown in Fig. 1.The passage identifies the system overview without describing its components.Source: D. O1 Configuration & Performance… · p. 4

III. EXPERIMENTAL RESULT

  • The experimental analysis uses the VIAVI Solutions AI RSG RAN Simulator, configured with parameters listed in Table II.Table II is titled “Scenario Simulation.”Source: §III · p. 4, Table II · p. 4

A. Single-Run Evaluation

In the single-run test, TxPower decreases as PEE.EnergyEfficiency increases. The precision evaluation shows convergence across several targets, but the system stops optimizing when it reaches a TxPower plateau.

A. Single-Run Evaluation

  • ~30 dBm to 11 dBm TxPower by iteration 4 accompanied an increase from ~7.7 × 10^5 to near 8 × 10^5 bits/joule PEE.EnergyEfficiency.Efficiency stabilized near 8 × 10^5 bits/joule from iteration 2 onward; the authors describe the trend as improved overall energy efficiency.Source: Figure 3 · p. 4, A. Single-Run Evaluation · p. 4

B. Closed-Loop Precision Evaluation

  • A seven-attempt closed-loop precision evaluation used targets 800 for attempts 1–5, 803 on attempt 6, and 805 on attempt 7; it reported successful convergence with iteration counts 3, 3, 2, 2, 2, and 4.Source: B. Closed-Loop Precision Evaluation · p. 4
  • The seventh attempt failed after 7 iterations when the maximum allowable TxPower adjustment was reached and the system identified a plateau.The system then abandoned further optimization because it could not fulfill the intent.Source: B. Closed-Loop Precision Evaluation · p. 4

C. Intent Boundary Testing

Intent Boundary Testing compares PEE.EnergyEfficiency targets with a baseline established from initial simulation data. Below-baseline targets are treated as already satisfied, while above-baseline targets prompt optimization until a plateau is reached.

C. Intent Boundary Testing

  • Table III reports Intent Boundary Testing for PEE.EnergyEfficiency against a baseline established from initial simulation data.The table’s caption identifies its scope as boundary testing for this metric.Source: C. Intent Boundary Testing · p. 4, Table III · p. 5
  • 750 kWh is processed through monitoring and verification only because the system considers the network already in a preferred state.This target is below the baseline.Source: C. Intent Boundary Testing · p. 5
  • At targets starting at 803 kWh above baseline, the system attempts optimization until a plateau, then abandons the process.The passage states that no further optimization is possible beyond the plateau.Source: C. Intent Boundary Testing · p. 5

IV. LESSON LEARNED AND FUTURE WORKS

The paper presents structured prompt engineering and collaborative LLM agents for iterative RAN configuration, while identifying contextual understanding as a reliability challenge. It argues that domain-specific ground truth can improve the accuracy of generated configurations.

IV. LESSON LEARNED AND FUTURE WORKS

  • Two LLM agents collaboratively refine RAN configurations through iterative optimization, adjusting parameters gradually based on real-time feedback to achieve the intent and prevent disruption.The prompt format translates high-level intents into structured JSON containing intent, network topology, KPIs, and configuration parameters.Source: §IV · p. 5

A. Contextual Understanding and Domain Specificity

  • Comprehensive PM data sometimes produces hallucinations or inconsistent configuration strategies during preliminary evaluation.The authors identify domain-specific ground truth as a way to improve LLM response accuracy and reliability in RAN configuration.Source: A. Contextual Understanding and Domain… · p. 5
  • Future work proposes representing network entities and parameter-to-metric relationships explicitly, using formalisms such as RDF or ontologies, before supplying domain knowledge to the LLM.The stated aim is to support accurate, concise, scientifically sound justification texts for translation.Source: A. Contextual Understanding and Domain… · p. 5

B. The Impact of PM Data Inaccuracy

The authors identify inaccurate aggregated PM data as a threat to reliable closed-loop RAN management. They propose exploring unaggregated streaming metrics and reinforcement learning as future directions.

B. The Impact of PM Data Inaccuracy

  • Aggregated PM values may not reflect the RAN’s true state, despite simulator metrics being transmitted at fixed intervals.The authors attribute this to aggregation algorithms and configurable time windows.Source: B. The Impact of PM Data Inaccuracy · p. 5
  • Inaccurate PM data can trigger a feedback loop in which flawed LLM decisions generate further unreliable PM data, potentially destabilizing the network.Source: B. The Impact of PM Data Inaccuracy · p. 5
  • Future work proposes streaming PM metrics without gNB aggregation and using reinforcement learning to adapt to noisy or outlier data through feedback and reward-based optimization.Source: B. The Impact of PM Data Inaccuracy · p. 5

Related papers

Loading 2507.14230v2…