매일 아침, 어제의 AI를 한 통으로 정리해 보내드립니다메일로 받아보기

METAL LAB

Hybrid-Policy Self-Editing for Composable Unstructured Knowledge Editing

arXiv:2608.116602026-08-13

arXiv:2608.11660v1 Announce Type: new Abstract: Large language models (LLMs) achieve remarkable performance across natural language tasks, yet they are trained on static corpora and their knowledge quickly becomes outdated in a fast-changing world. This motivates knowledge editing (KE), which updates specific knowledge in an LLM without changing unrelated others. Recent works move from structured knowledge triples toward unstructured KE (UKE), where the edit is a free-form passage that may state

저자 · Tianci Liu, Zihan Dong, Tianchun Li, Yi-Chung Chen, Qiming Cao, Xingchen Wang, Shiyang Wang, Zichen Miao, Linjun Zhang

arXiv에서 원문 보기