SPARE-RAG: Evidence Repair and Verification-Guided Sequential Neural Pruning for Multi-Hop RAG

Multi-hop RAG requires complementary facts from multiple documents to form a complete evidence chain. Relevance-based passage selection can miss bridging evidence, while adding more passages can introduce redundancy and distractors. Context compression reduces input length but does not, by itself, establish that the remaining evidence supports all required reasoning steps. We propose SPARE-RAG, which combines evidence repair with verification-guided sequential neural pruning. The key insight is to first repair missing support and then assess each proposed deletion by its effect on the remaining evidence set. Within a fixed candidate pool, the framework first supplements missing support, then proposes passage deletions and checks the remaining evidence before accepting or rejecting each proposed deletion. Experiments on HotpotQA, 2WikiMultiHopQA, and MuSiQue with frozen Qwen2.5-7B-Instruct show that SPARE-RAG consistently outperforms the compared methods in answer and evidence quality, achieving Answer F1 scores of 70.27, 52.49, and 35.68, respectively. These results demonstrate that combining evidence repair with verified sequential pruning improves both evidence quality and downstream answer generation.