Leveraging Multi-Agent System (MAS) and Fine-Tuned Small Language Models (SLMs) for Automated Telecom Network Troubleshooting

Jul 16, 2026

Upload by Patrick

Telecom networks expanding in scale and heterogeneity make troubleshooting difficult, slow, and dependent on subject matter experts who manually correlate performance metrics, configurations, alarms, and logs. Existing AI methods often need extensive labeled data and generalize poorly across deployments. This work proposes a Multi-Agent System where an orchestrator coordinates specialized agents for near real-time diagnosis. A fine-tuned small language model generates executable, domain-grounded remediation plans from internal troubleshooting documentation, improving automation and efficiency in both RAN and core networks.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages6

Reading Time15 min

Introduction

Motivation and Challenges

SLMs and Fine-Tuning Rationale

Proposed MAS Architecture

Related Work

Traditional Troubleshooting Approaches

Data-Driven and GenAI-Enabled Methods

Leveraging Multi-Agent System (MAS) and Fine-Tuned Small Language Models (SLMs) for Automated Telecom Network Troubleshooting Chenhua Shi Bhavika Jalli Gregor Macdonald John Zou Wanlu Lei Mridul Jain Joji Philip Ericsson arXiv :2511 .0065 1v2 [ cs .AI] 9 Jul 2026 Abstract—Telecom networks are rapidly increasing in scale and complexity, making management, operation, and optimization increasingly challenging. Although Artificial Intelligence (AI) has been applied to various telecom tasks, existing approaches are often limited in scope, require extensive labeled data, and struggle to generalize across heterogeneous deployments. Consequently, network troubleshooting still relies heavily on Subject Matter Experts (SMEs) to manually correlate multiple data sources and determine root causes and corrective actions. In this paper, we propose a Multi-Agent System (MAS) that leverages an agentic workflow in which Large Language Models (LLMs) coordinate specialized tools to enable automated network troubleshooting. Upon fault detection by AI/MLbased monitoring systems, the framework dynamically activates multiple agents—including an orchestrator, solution planner, executor, data retriever, and root-cause analyzer—to diagnose issues and recommend remediation strategies in near real time. A key component is the solution planner, which generates executable remediation plans grounded in internal operational documentation. To enable this capability, we fine-tune a Small Language Model (SLM) using proprietary troubleshooting documents to produce domain-specific solution plans. Experimental results demonstrate that the proposed framework significantly improves troubleshooting automation and efficiency across both Radio Access Network (RAN) and Core network domains. Index Terms—Large Language Models (LLMs), Multi-Agent System (MAS), Small Language Model (SLM), Agentic Workflow, Fine-tuning, Network Automation, Network Troubleshooting, Radio Access Network (RAN), Core Network I. INTRODUCTION Telecommunication networks are evolving into highly dynamic and heterogeneous environments spanning multiple standards, vendors, and deployment scenarios [1] . This growing complexity makes troubleshooting particularly challenging, as it requires correlating diverse data sources such as performance metrics, configurations, alarms, and logs. Although AI/ML algorithms have been integrated into fault detection and troubleshooting, much of the diagnostic and remediation process is still performed manually by SMEs. This results in slow, resource-intensive, and difficult-to-scale operations. Consequently, autonomous monitoring and troubleshooting is becoming essential to reduce reliance on human expertise and improve operational efficiency. Recent advances in Generative AI [2] and Foundation Models [3]—particularly LLMs—have opened new opportunities for intelligent network automation. LLM-driven agentic systems, often combined with Retrieval-Augmented Generation (RAG) [4] [5], have demonstrated strong reasoning and orchestration capabilities across multiple domains. Applied to telecom, such systems enable dynamic workflows in which specialized agents (e.g., solution planners, data retrievers, and root-cause analyzers) collaborate under the coordination of an LLM to perform complex troubleshooting tasks in a continuous REACT-style loop (Reasoning, Execution, and Action) [6] . However, practical deployment of these systems faces several challenges: (i) high operational costs associated with external LLM providers, (ii) data privacy risks when handling sensitive network information, and (iii) substantial capital expenditures (CapEx) required to host and deploy large models within operator environments. To address these limitations, there is growing interest in SLMs as lightweight, domain-specialized alternatives to LLMs. SLMs are the future of agentic AI [7] since they can offer sufficient reasoning capacity for many agentic workflows while being mo

FAQs

为什么传统的网络故障排查仍高度依赖人工专家？

因为排查需要关联多源数据（性能指标、配置、告警和日志），而现有AI方法在范围、标注数据需求和跨部署泛化能力上存在限制，导致大量诊断与处置仍由SMEs手动完成。

文中提出的MAS如何在故障发生后实现自动化排查？

监控系统先进行故障检测，随后框架动态激活多个代理：编排器、方案规划器、执行器、数据检索器和根因分析器，并在近实时流程中完成诊断与修复建议。

方案规划器（solution planner）为何采用微调的小语言模型？

方案规划器用专有的故障排查文档对小语言模型进行微调，使其生成贴合领域的、可执行的补救方案，从而降低成本并提高可部署性。

该框架的效果提升体现在哪些网络域？

实验结果表明，该方法能够显著提升故障排查自动化与效率，并在无线接入网（RAN）与核心网（Core network）两个领域都取得改进。

Jul 16, 2026

Upload by Patrick