DualView: Adaptive Local-Global Fusion for Multi-Hop Document Reranking
This work introduces an efficient dual-view cascaded reranking framework for multi-hop question answering. It combines a fine-grained Local Scorer with a Global Scorer that models document dependencies. By dynamically fusing these two views using an Adaptive Gate conditioned on query semantics, the model achieves state-of-the-art recall on MuSiQue while significantly reducing latency compared to larger cross-encoders.