본문 바로가기
  • Home

RAG-Enhanced small Large Language Models: Enhancing Battlefield Analysis through Knowledge Distillation of Large Language Models

  • Journal of The Korea Society of Computer and Information
  • Abbr : JKSCI
  • 2025, 30(3), pp.43~57
  • Publisher : The Korean Society Of Computer And Information
  • Research Area : Engineering > Computer Science
  • Received : February 10, 2025
  • Accepted : March 12, 2025
  • Published : March 31, 2025

Wonjun Cho 1 Jaesung Yoo 1 Sang-Min Kim 1 Jaeeun Jang 1

1한화시스템

Accredited

ABSTRACT

The increasing complexity of modern battlefields and the importance of real-time data processing have heightened the need for effective battlefield situation analysis systems. This study proposes a battlefield analysis system utilizing Large Language Models (LLMs), specifically introducing an advanced approach that combines Retrieval-Augmented Generation (RAG) with Supervised Fine-Tuning (SFT). To address the hallucination problems and lost-in-the-middle phenomenon inherent in existing RAG systems, we introduce a triple-structured learning approach that incorporates reference documents in the SFT process. Based on synthetic battlefield datasets developed in collaboration with military experts, our experimental results demonstrate exceptional performance in source extraction accuracy and response quality evaluation. Notably, when applying triple-structured SFT to an 8B parameter model, we achieved comparable performance to a 405B parameter model, proving its practicality in actual battlefield environments. Furthermore, our lightweight model enhanced with specialized training strategies showed minimal performance degradation compared to larger models, suggesting its viability for deployment in resource-constrained environments. This research demonstrates the effective application of LLMs in battlefield situation analysis and presents a novel direction for military domains requiring real-time data processing and high reliability.

Citation status

* References for papers published after 2023 are currently being built.