TLDR: This research introduces a Multi-Agent System (MAS) that uses a Large Language Model (LLM) to coordinate specialized AI agents for automated telecom network troubleshooting. A key component is a Small Language Model (SLM), fine-tuned on internal documents, which generates precise remediation plans. This system significantly reduces troubleshooting time and improves accuracy in both Radio Access Network (RAN) and Core network domains, offering a cost-effective and privacy-preserving alternative to relying solely on large external LLMs.
Telecommunication networks are becoming increasingly vast and intricate, making their management, operation, and optimization a significant challenge. Traditionally, resolving network issues heavily relies on human Subject Matter Experts (SMEs) who manually sift through various data sources to pinpoint problems and devise solutions. This process is often slow, resource-intensive, and difficult to scale.
A new research paper introduces an innovative approach to tackle these challenges: a Multi-Agent System (MAS) that employs an agentic workflow, using Large Language Models (LLMs) to coordinate multiple specialized tools for fully automated network troubleshooting. This system aims to significantly reduce the reliance on human expertise and enhance operational efficiency.
How the Multi-Agent System Works
Once faults are detected by existing AI/ML-based monitors, the framework dynamically activates a suite of specialized agents. These include an orchestrator, a solution planner, an executor, a data retriever, and a root-cause analyzer. These agents work together to diagnose issues and recommend remediation strategies swiftly.
The orchestrator acts as the central brain, interpreting network faults or user requests and delegating tasks to the appropriate agents. For instance, if an “Input Power Failure” is detected in a Radio Access Network (RAN), the orchestrator initiates the solution planner to generate a step-by-step troubleshooting guide.
The data retriever agent is equipped with tools to access diverse databases, collecting performance metrics, alarms, and logs from various sources. The root-cause analyzer agent correlates this data with its knowledge base to identify the underlying cause of the problem and suggest remediation strategies. The executor agent refines the troubleshooting plans with specific details like node identifiers and time intervals, and, after human approval, implements these plans. Finally, a dashboard display agent visualizes the results in an interactive interface for SMEs to review.
The Role of Fine-Tuned Small Language Models (SLMs)
A key innovation in this system is the solution planner, which generates appropriate remediation plans based on internal documentation. To achieve this, the researchers fine-tuned a Small Language Model (SLM) on proprietary troubleshooting documents. This fine-tuning enables the SLM to produce domain-grounded solution plans, making it highly effective and tailored to telecom-specific issues.
The use of SLMs addresses several practical challenges associated with deploying large external LLMs, such as high operational costs, data privacy risks when handling sensitive network information, and the substantial capital expenditures required to host and deploy large models. SLMs offer sufficient reasoning capacity for many agentic workflows while being more economical, privacy-preserving, and adaptable after fine-tuning.
The fine-tuning process involved a two-stage strategy: Supervised Fine-Tuning (SFT) as a warm-up phase, followed by Reinforcement Fine-Tuning (RFT). The RFT stage, implemented using the Transformers Reinforcement Learning (TRL) framework, guides the model’s behavior through feedback and rewards, strengthening its reasoning ability to produce accurate, domain-grounded troubleshooting steps.
Also Read:
- Language Models Powering Smarter Multi-Agent Collaboration
- Intelligent AI System Enhances Medical Pre-Consultation with Proactive Inquiry
Experimental Results and Benefits
The proposed framework was evaluated through experiments across both RAN and Core network domains. The results are compelling: the autonomous network operations MAS demonstrated superior performance compared to human engineers, achieving a six-fold reduction in mean troubleshooting time per node and a 10% improvement in accuracy. This translates to faster issue resolution, lower operational costs, and an improved customer experience.
For example, in troubleshooting power system issues in RAN, the system could generate detailed steps like checking correlated alarms, analyzing affected units, and monitoring specific power-related counters. Similarly, for PDU session degradation in the Core network, it could suggest monitoring various session-related counters, performing call flow analysis, and checking logs and alarms.
The fine-tuned SLM solution planner, particularly the 8B parameter model, showed significant improvements in reward values and stability compared to its base version, generating solutions of comparable quality to more powerful models like GPT-4o-mini for specific troubleshooting scenarios. This highlights the potential of combining LLM-based orchestration with fine-tuned SLMs to achieve scalable, efficient, and privacy-preserving automation in telecom network operations.
For more in-depth technical details, you can read the full research paper here.
