Guo Y., Tang Q., Trang H., Nguyen C.D.
International Conference on Agents and Artificial Intelligence, vol. 3, pp. 2441-2448, 2026
Large language models (LLMs) have significantly advanced natural language processing across multiple domains with their remarkable capabilities in processing, understanding, and generating human-like text at scale. Different strategies have emerged for adapting these models to specialized domains, such as fine-tuning on domain-specific data and augmenting base models with external knowledge retrieval systems. In this paper, we focus on the telecom security domain and compare three strategies for leveraging LLMs: using unmodified base LLMs with prompt engineering, fine-tuning LLMs on domain-specific data, and enhancing base LLMs with Retrieval-Augmented Generation (RAG). Our experiments demonstrate that while fine-tuned LLMs (slightly) improves performance than base models, the RAG strategy outperforms fine-tuned LLMs while requiring significantly lower computational resources and offering greater flexibility. These findings suggest that for practical telecom security applications, RAG represents a more efficient and effective strategy than domain-specific fine-tuning.
