본문 바로가기 주메뉴 바로가기
국회도서관 홈으로 정보검색 소장정보 검색

결과 내 검색

동의어 포함

초록보기

거대 언어 모델의 발전에 따라 문맥 내 학습은 언어 모델의 대표적인 활용법으로 주목받으며, 이에 다양한 프롬프트 기법이 연구되고 있다. 특히, 사고 과정의 명시를 유도한 few-shot CoT(Chain-of-Thought)는 소수의 예시 제공만으로 추론 성능을 극대화한 방법으로 알려져 있으나, 예시 구성에 따라 성능 편차가 발생한다는 한계가 존재한다. 기존 연구는 다양성이나 불확실성등 일부 기준을 중심으로 예시를 구성할 질문을 선정해 왔으나, 난이도 기반 질문 선정에 관한연구는 상대적으로 미진한 실정이다. 이에 본 연구는 언어 모델을 활용한 쌍대 난이도 비교와 스위스 토너먼트 구조를 결합하여, 고난도 질문을 체계적으로 선별하고 이를 기반으로 few-shot CoT 예시를 구축하는 새로운 방법론을 제안한다. 제안 방법론의 성능 평가를 위해 수학 서술형 벤치마크인 GSM8K 데이터셋 1,319개 문항을 대상으로 실험을 수행한 결과, 제안 방법론이 무작위 선정, 불확실성 기반, 그리고 난이도 직접 평가 방식 대비 정확도 측면에서 각각 2.12%p, 1.36%p, 그리고 10.16%p의 성능 향상을 보임을 확인하였다.

With the advancement of large language models(LLMs), in-context learning has become a key approach, driving research on prompting techniques. Among them, few-shot Chain-of-Thought(CoT) prompting, which induces explicit reasoning, shows strong performance but depends on example composition. While prior work focused on diversity or uncertainty in example selection, difficulty-based approaches remain underexplored. This study proposes a method to identify high-difficulty questions by combining pairwise difficulty comparisons conducted by an LLM with a Swiss tournament structure, constructing few-shot CoT exemplars with human reasoning annotations. Experiments on 1,319 GSM8K problems show that the proposed method outperforms random, uncertainty-based, and direct difficulty evaluation approaches by 2.12%p, 1.36%p, and 10.16%p, respectively.

권호기사

권호기사 목록 테이블로 기사명, 저자명, 페이지, 원문, 기사목차 순으로 되어있습니다.
기사명 저자명 페이지 원문 목차
(A) GNN-based bitcoin mixing detection method via subgraph generation of mixing transactions Hyeon-Woo Lee, Eun-Young Park, Sooncheol Kim, Jiyeon Kim p. 139-151
(The) effects of proactive AI agent intervention methods on user experience in social conversation contexts Seong-Yeon Kim, Ju-Hye Ha, Chang-Hoon Oh p. 1-16
From avoidance to approach : LLM agent effects on appraisals and approach motivation Hae-Ryung Lee, Ju-Hye Ha, Chang-Hoon Oh p. 17-29
(An) LLM-based automatic selection of high-difficulty questions using the Swiss tournament format Jooeun Lee, Minseob Song, Namgyu Kim p. 31-42
(A) comparative study on the performance of GPT-4o-mini, Claude 4 Sonnet, and Gemini 2.5 flash models using the prompt runner framework Misun Lee p. 43-50
RAR-agent : a rationale-augmented retrieval framework for legal question answering Gyuhyeong Kim, Yunhyeok Do, Joonhyeon Song, Ziyang Liu p. 51-64
ProbCert : probabilistic certification framework for black-box LLM outputs Jong Wook Kim p. 65-73
(A) novel two-stage attacks on Korean language models : single-token triggers search and morphology-preserving minimal edits Areum Im, Taehwa Lee, Soojin Lee p. 75-85
(A) machine learning–based analysis of short-term surges and listing premiums for newly listed altcoins on Korean exchanges using offshore prices and social signals Eun Hong Park, Yeong-In Lee, Ha Young Kim p. 87-96
Perception gaps between IT managers and developers on generative AI coding : a co-orientation approach Miyoung Bang, Sungkap Park, Hanjin Lee, Dongsu Lee p. 97-110
Empirical analysis of the factors influencing project success in Korean reward-based crowdfunding Changhyun Kim, Sungkap Park, Hanjin Lee p. 111-121
Stable SDF-based collision handling for cloth-solid interaction in position-based dynamics Jong-Hyun Kim p. 123-130
Adaptive digital watermarking based on multi-rendering signals : perceptual quality enhancement through JND model integration Ji-Hoon Kim, Dong-Heon Cho, Seung-Kwan Choi p. 131-138
Development of a UWB-based indoor positioning RTLS system Hoeteak Jung, Seungmin Oh p. 153-159
Design and implementation of a personalized recipe recommendation system based on Jaccard similarity Yeong-Woong Jo, Han-Jun Jung, Yong-Won Jo, Wu-Jin Jang, Kyeong-Sook Han p. 161-171
ERC-7891 : a new protocol design for hierarchical NFTs with splitting and merging Nitin Bhagat, JongWook Bae, Su-Hyun Lee p. 173-181
Task scheduling scheme for maintaining freshness in digital twin systems Taehun Yang, Cheonyong Kim p. 183-189
Analysis of air circulation characteristics in a space-approximate cultivation chamber according to CFD-based intake-exhaust arrangement Hyeon-O Choe p. 191-199
Effects of communication skills, nunchi skills, major satisfaction, and metacognition on clinical practice stress among health-related college students Mi-Ju Yang, Tae-Hee Jang, Jun-Gyo Lee, Min-Hyeok Lee, Yeon-Ji Lee, Jung-Hee Park p. 201-209
(The) effects of dysfunctional beliefs on happiness in nursing students : focusing on the moderating effect of emotional expressiveness Youn-Kyoung Kwag p. 211-218
Intellectual property technology trends in the hanbok Jeogori Su-Joung Cha p. 219-230
(A) comparative analysis of BERTopic and LDA for topic modeling of Korean sleep health discourse on social media JongHwi Song p. 231-239
Do political news affect financial markets? Evidence from exchange rate and investor trading behavior in Korea Jieun Kim, Zoonky Lee, Dojoon Park p. 241-248
Analysis of errors and misconceptions in scratch programming among pre-service elementary teachers Junghee Jo p. 249-257
(A) generative AI-based system for business plan development and investment simulation in startup education Young-Cheon Kim p. 259-267
(The) effects of history classes utilizing physical computing on students’ interest in history and SW Minkuk Cho, Kwihoon Kim p. 269-280