Optimizing LLM Prompts for Automation of Network Management: A User's Perspective
V. S. Sai Rama Krishna Komanduri, Sebastian Estropia, Scott Alessio, Gokhan Yerdelen, Tyler Ferreira, G Roldan, Ziqian Dong, Roberto Rojas‐Cessa · 2025
Although being linguistic assemblers and decomposers, large language models (LLMs) have found their use in analysis and generation of code, art, and design of electronic circuitry, industrial parts, and network design. However, the probabilistic nature of generative LLMs makes network design and implementation scenarios prone to errors. The complexity levels of design may exacerbate the number of inaccuracies included in an LLM's response. Therefore, it is necessary to identify the features that make prompts generate effective and error-free responses as users. To reduce the error rate of the responses, we test prompt specificity in text and schematic descriptions. As various degrees of specificity, we compare highly intuitive to highly specific prompting. The responses are expressed as network schematics and router configuration commands that are evaluated with our proposed scoring policy. Our tests include networks with three levels of complexity and multiple levels of specificity in text and graphic prompts. The results show the trade-offs on the text and graphic modes and the degrees of specificity.