Answer presence, not evidence quality, drives RAG rewriting gains

Arxiv CS.AI (top papers)

Study shows LLM-based rewrites in RAG pipelines improve F1 by injecting correct answers into context, not by improving evidence relevance. The finding challenges the common assumption that better evidence selection drives rewriting benefits.

rag

Arxiv CS.AI (top papers)·Yuejie LiYueying HuaKe YangLi ZhangYueping HeYueping HeRuiqi LiBolin ChenTao WangBowen LiChengjun Mao

6 days ago