retrieval-augmented QA 相关 AI 资讯

10:31

官方一手arXiv: DeepSeek@Jingjie Ning, Xueqi Li

A study reveals that a retrieval-augmented QA system can produce different answers after index expansion, despite fixed model settings. The research introduces a new method, Snapshot Compatibility Audit, to estimate excess answer churn. The study finds significant churn in both Natural Questions and TriviaQA datasets, indicating the importance of auditing compatibility alongside utility in retrieval-augmented systems.

推荐理由：This paper offers a new perspective on the consistency of retrieval-augmented QA systems, highlighting the importance of the Snapshot Compatibility Audit method. It's a must-read for those interested in QA system reliability and the nuances of corpus-induced changes.

原文