Scholarlink

<p>As a step toward intent-based networking (IBN), we investigate automating Virtual Network Function (VNF) deployment based on natural language input. In particular, we developed a system that leverages recent advances in Large Language Models (LLMs). LLMs demonstrate strong performance not only in natural language understanding but also in reasoning and code generation, to process natural language descriptions of VNF deployment. We proposed, implemented, and validated an LLM-based AI Agent that accepts natural language inputs describing the desired VNF deployment and generates a corresponding automated workflow with high success rates. The developed AI Agent employs a few-shot learning approach to generate VNF installation workflows. Before applying the generated workflow to a real network environment, it is first validated within a Kubernetes-based Network Digital Twin (NDT) to ensure correctness. In the event of errors, the system extracts logs from the NDT and supplies them to the LLM for iterative refinement of the workflow. To enhance this process, we integrated Retrieval-Augmented Generation (RAG) and Multi-Agent Debate (MAD) techniques, enabling the LLM to incorporate external knowledge when resolving issues. In particular, to enable effective integration of RAG into the AI Agent, we proposed and empirically validated input and output filtering mechanisms, thereby enhancing the self-healing capabilities of the AI agent. Experimental results demonstrate that even LLMs that were initially unable to achieve fully automated VNF deployment via prompt-only interaction were capable of generating functional workflows for the NDT when guided by the proposed AI Agent. These findings validate our approach as a meaningful contribution to the translation and activation phases of IBN.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLM-Based AI Agent for Virtual Network Function Deployment

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月3日

查看全文
  • Sukhyun Nam,
  • Nguyen Van Tu,
  • James Won-Ki Hong

摘要

As a step toward intent-based networking (IBN), we investigate automating Virtual Network Function (VNF) deployment based on natural language input. In particular, we developed a system that leverages recent advances in Large Language Models (LLMs). LLMs demonstrate strong performance not only in natural language understanding but also in reasoning and code generation, to process natural language descriptions of VNF deployment. We proposed, implemented, and validated an LLM-based AI Agent that accepts natural language inputs describing the desired VNF deployment and generates a corresponding automated workflow with high success rates. The developed AI Agent employs a few-shot learning approach to generate VNF installation workflows. Before applying the generated workflow to a real network environment, it is first validated within a Kubernetes-based Network Digital Twin (NDT) to ensure correctness. In the event of errors, the system extracts logs from the NDT and supplies them to the LLM for iterative refinement of the workflow. To enhance this process, we integrated Retrieval-Augmented Generation (RAG) and Multi-Agent Debate (MAD) techniques, enabling the LLM to incorporate external knowledge when resolving issues. In particular, to enable effective integration of RAG into the AI Agent, we proposed and empirically validated input and output filtering mechanisms, thereby enhancing the self-healing capabilities of the AI agent. Experimental results demonstrate that even LLMs that were initially unable to achieve fully automated VNF deployment via prompt-only interaction were capable of generating functional workflows for the NDT when guided by the proposed AI Agent. These findings validate our approach as a meaningful contribution to the translation and activation phases of IBN.