본문 바로가기 주메뉴 바로가기
22대 대한민국 국회 국회정보길라잡이 국회의원검색
회원가입 로그인
HOME

정보검색

소장정보 검색

국회도서관 홈으로 정보검색 소장정보 검색

결과 내 검색

동의어 포함

고급검색

인명/단체명

저자정보 상세정보
인명/단체명을 입력하세요.

키워드

  • 대표어

  • 외국어

-

초록보기

본 연구는 경량 대규모 언어모델 환경에서 검색 증강 생성(RAG)과 소프트 프롬프트 튜닝이 성능과 효율성에 미치는 결합 효과를 분석한다. SBA 정책 질의응답 벤치마크를 활용하여, 정밀도설정, RAG 방식, 소프트 프롬프트 적용 여부를 포함한 총 18개의 실험 구성 하에서 Liquid 기반모델과 Transformer 기반 모델을 비교하였다. 분석 결과, 소프트 프롬프트의 효과는 모델 아키텍처에 따라 상이하게 나타났다. Liquid 모델에서는 응답 품질과 지연시간 안정성이 동시에 개선된 반면, Transformer 모델에서는 품질과 효율성 간의 뚜렷한 상충관계가 확인되었다. 또한 범주별 분석을 통해 수치 및 기준 기반 질문 유형이 지속적인 성능 병목 요인으로 나타났으며, 이는 경량LLM 실무 적용을 위한 정규화 중심 개선의 필요성을 시사한다.

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.

권호기사

권호기사 목록 테이블로 기사명, 저자명, 페이지, 원문, 기사목차 순으로 되어있습니다.
기사명 저자명 페이지 원문 목차
(An) efficient simultaneous thermal and electrical state estimation for preventive maintenance of PV-ESS in low-temperature environments : a case study of LSTM/transformer Yong-Je Ko, Ho-Young Kwak p. 1-8
Adversarial noise-based speaker de-identification for cloud speech services Haram Kang, Sangwoon Yun, Jemin Ahn, Kyungtae Kang p. 9-19
Semantic tabular-to-image conversion and contrastive learning for lightweight intrusion detection Jun Yeong Park, Kunwoo Kang, Hoin Lee, Seungeun Lee, Yu Rang Park p. 21-35
Transformer based arrival time prediction for a target bus stop using single stop information Ga-On Cheon, Young-Sup Hwang p. 37-48
Training-free feature-level interpolation with sparse-backbone scheduling for YOLO detection in low-FPS video Min-Ho Kim, Kyu-Cheol Cho p. 49-59
Band-attention BFE-Net for subject-independent EEG emotion recognition Yujin Ji, Jungpyo Hong p. 61-69
CNN-based thread defect detection system for manufacturing process automation Yeong-Seo Lee, Seongbae Eun, Dong-beom Shin, Jae-Heum Lee p. 71-78
(A) comparative study on the explainability and efficiency of foundation models in prompt-based sentiment analysis Misun Lee p. 79-87
Time-period-aware difficulty assessment and modeling framework for improved building energy consumption prediction Yong-Je Ko, Ho-Young Kwak p. 89-96
Android malware classification using an edge-centric graph isomorphism network Yeeun Lee, Hyeona Jang, Hanseul Jung, Eunjung Choi p. 97-106
Edge-cloud pet behavior monitoring system based on on-device encoding anonymization Hyuksoon Choi, JinHwan Yang, Nammee Moon p. 107-116
Asymmetric soft prompt–RAG interactions in liquid vs. transformer lightweight LLMs : an empirical study on SBA policy QA Jun Oh Cheong, Hyunchul Ahn p. 117-125
(An) empirical study on adoption and purchase intentions for imported electric vehicles based on innovation diffusion theory Jihoon Choi, Junghwan Kim, Sangkue Park, Sungkap Park, Dongsu Lee p. 127-137
(A) Korean movie genre prediction model using multi-representation learning Jong-Hyun Kim p. 139-147
(The) impact of background music in dramas on viewers’ emotions and immersion, mood Sang-Sik Choi, Jae-Im Yun p. 149-160
Privacy concerns in online media usage : a comparative study across gender and age groups Sang Dol Kim p. 161-168
Parameter importance analysis of a bio-signal-based crisis recognition model using explainable artificial intelligence Joon-Yong Kim p. 169-177
(A) study on object-oriented standard architecture for export-oriented IFF display software in naval combat systems Sung-Uk Jo, Su-Seok Park, Young-San Kim p. 179-189
User satisfaction with electronic record systems in long-term care facilities Se Young Kim, Haejin Kim, Jungpyo Hong, Mi Kyung Kim p. 191-197
Research on KS standard-based outlier classification and intelligent cleaning framework for smart greenhouse sensor data Kyeong-il Ko, Byeongbeom Kang, Hyun Yoe p. 199-208
Comparison of perioperative hemoglobin changes by anesthesia type and measurement method in hip surgery Je Bog Yoo p. 209-217
(A) study on digital transformation and AI advancement strategy in customs administration : comparative analysis of KCS, CBP, EUCA and GACC Beom-Soo Shin p. 219-225
Factors influencing online broadcast paid sponsorship intentions based on gift-giving and customer value models Hyun Jin Yeo p. 227-232
(A) study on the potential use of generative AI “Nano Banana 2” in the production of disability-focused brand webtoons Sang-Hwa Lee, Byong-Kwon Lee, Na-ra Jung p. 233-240

추천서가 (다양한 추천 자료를 만나보세요)

권호기사보기

권호기사 목록 테이블로 기사명, 저자명, 페이지, 원문, 기사목차 순으로 되어있습니다.
기사명 저자명 페이지 원문 기사목차

권호

기사명 원문보기 기사목차
닫기