New set-level retrieval method boosts multi-hop QA by 15%

Multi-hop question answering requires retrieving multiple passages that together contain the full answer. Current retrievers score each passage independently or use sequential supervision, often selecting incomplete or conflicting sets. To fix this, Song and Lee propose a set-level compatibility learning framework that directly ranks evidence sets based on their joint usefulness.

They implement two complementary scorers: ParaSet, a lightweight late-interaction model that applies self-attention over precomputed bi-encoder embeddings for fast candidate-set exploration, and SetCE, a cross-encoder reranker trained with the same set-level objective. Experiments on multi-hop QA benchmarks (e.g., HotpotQA, 2WikiMultihop) show that set-level retrieval outperforms document-level retrieval and that combining outputs from both scorers yields stronger results than simply retrieving more passages from a single retriever. The approach is much more efficient than LLM-based set selection, making it practical for real-world retrieval-augmented generation.