Contrastive Evidence Exploration for Multi-Hop RAG

This training-free framework improves multi-hop retrieval by using evidence-conditioned exploration and passage-specific contrastive refinement. It addresses the noise introduced by one-shot query expansion by characterizing passages relative to semantically similar neighbors during indexing.