본문 바로가기
  • Home

Asymmetric Soft Prompt–RAG Interactions in Liquid vs. Transformer Lightweight LLMs: An Empirical Study on SBA Policy QA

  • Journal of The Korea Society of Computer and Information
  • Abbr : JKSCI
  • 2026, 31(3), pp.117~125
  • Publisher : The Korean Society Of Computer And Information
  • Research Area : Engineering > Computer Science
  • Received : January 5, 2026
  • Accepted : February 20, 2026
  • Published : March 31, 2026

Jun Oh Cheong 1 Hyunchul Ahn 1

1국민대학교

Accredited

ABSTRACT

This study examines how retrieval-augmented generation (RAG) and soft prompt tuning jointly influence performance and efficiency in lightweight large language models. Using an SBA policy QA benchmark, we compare a Liquid-based model and a Transformer-based model across 18 experimental configurations covering precision settings, RAG modes, and soft prompting. Results reveal architecture-dependent effects: soft prompting improves both answer quality and latency stability in the Liquid model, whereas it introduces a quality?efficiency trade-off in the Transformer model. Category-level analysis identifies numeric and criteria-based questions as persistent bottlenecks, highlighting the need for normalization-oriented improvements in lightweight LLM deployment.

Citation status

* References for papers published after 2024 are currently being built.