Retrieving from Your Own Mistakes: On-Policy Training of Embedding Models for Multi-Hop Retrieval-Augmented Generation
On-policy state training, which trains the encoder on the states it reaches itself, which retrieves complete evidence more often than teacher-forced training and the systems the authors compare, and it transfers without further training to harder long-chain variants of HotpotQA and 2WikiMultiHopQA that are constructed.
Chen-Jie Qiu, Zi-Xuan Zeng, Shao-Bin Shi et al.
· 0 citations