Hybrid-Policy Self-Editing for Composable Unstructured Knowledge Editing
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performance across natural language tasks, yet they are trained on static corpora and their knowledge quickly becomes outdated in a fast-changing world.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
arXiv에서 원문 보기