DualView Reranking Hits 99.4% Recall at 4ms, Beats 600M Models

Researchers from the arXiv team (Zhang, Li, Zhao) have unveiled DualView, a cascaded reranking framework designed specifically for multi-hop document retrieval. Multi-hop question answering demands aggregating information from multiple documents, a task where traditional rerankers struggle with both recall and speed. DualView tackles this by combining two views: a Local Scorer that uses stacked cross-attention to measure query-document relevance at a fine-grained level, and a Global Scorer that models inter-document dependencies via a Transformer-based context aggregator. These views are dynamically blended using an Adaptive Gate conditioned on query semantics, ensuring the system adapts to the complexity of each query.

Under a fixed candidate set with offline cached embeddings, DualView achieves stellar results on the MuSiQue dataset: 99.4% Top-4 Recall and 97.8% Full Hit accuracy while maintaining a latency of just 4.0 ms (249 queries per second). This substantially outperforms larger cross-encoders like BGE-Large (92.0% Recall) and Jina-v3 (90.1% Recall) while being 5 to 6 times faster. Ablation studies confirm that both the local and global views contribute meaningfully to the multi-hop performance. For professionals building retrieval-augmented generation (RAG) systems, DualView offers a practical, high-speed alternative without sacrificing accuracy.