As large language models (LLMs) and Retrieval-Augmented Generation (RAG) systems become widely used, generated text is increasingly conditioned on retrieved or user-provided materials. While prior evaluations of AI-generated text (AIGT) de...

http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
https://www.riss.kr/link?id=T17568010
경산 : 영남대학교 대학원, 2026
2026
한국어
경상북도
; 26 cm
지도교수: 徐永昔
I804:47017-200001038937
0
상세조회0
다운로드다국어 초록 (Multilingual Abstract)
As large language models (LLMs) and Retrieval-Augmented Generation (RAG) systems become widely used, generated text is increasingly conditioned on retrieved or user-provided materials. While prior evaluations of AI-generated text (AIGT) de...
As large language models (LLMs) and Retrieval-Augmented Generation (RAG) systems become widely used, generated text is increasingly conditioned on retrieved or user-provided materials. While prior evaluations of AI-generated text (AIGT) detection have examined factors such as generator models, domains, languages, task types, and post-generation transformations, the composition of retrieval evidence available at generation time remains underexplored as an explicit evaluation factor. In this paper, we propose Retrieval-Grounded Text Generation (RGT) as an evaluation scenario for AIGT detection and compare the Generic and Task-specific baseline scenarios with RGT in-doc and RGT external based on the same source documents and human-written texts (HWT) across arXiv, PubMed, and Multi-News using four generation models and five score-based detectors. The overall mean AUROC is 0.952 for Generic, 0.666 for Task-specific, 0.630 for RGT in-doc, and 0.591 for RGT external, and detector rankings also change across generation scenarios. The HWT-AIGT difference in mean negative log-likelihood (Mean NLL) narrows in Task-specific and the two RGT scenarios and reverses under some conditions, whereas probability-structure-based scores exhibit a distinct pattern. The average TPR@1%FPR decreases from 0.563 in Generic to 0.066–0.071 in Task-specific and the two RGT scenarios, showing that AUROC alone does not adequately represent performance at operational points. These results establish the composition of retrieval evidence as an explicit evaluation axis for AIGT detection and demonstrate the need to revalidate detection performance and rankings in settings that reflect the generation task, input context, and retrieval evidence rather than directly generalizing results from existing non-retrieval evaluations to RGT settings.
목차 (Table of Contents)
[제18회 김옥길기념강좌] 인공지능, 감정, 휴머니즘(Human-Compatible Artificial Intelligence’)’
이화여자대학교 스튜어드 러셀누구나 할 수 있는 데이터 분석과 인공지능[Data Analysis and Artificial Intelligence for Everyone]
K-MOOC 인하공업전문대학 이세훈누구나 할 수 있는 데이터 분석과 인공지능[Data Analysis and Artificial Intelligence for Everyone]
K-MOOC 인하공업전문대학 이세훈생성형 AI의 이해와 활용
강원대학교 이재웅, 최정아, 장윤재, 이효준대형언어모델의 기초
영남대학교 박우길