RASER: Recoverability-Aware Routing Cuts Retrieval Costs in Multi-Hop QA by Selectively Escalating

arXiv·medium signal

Proposes RASER, a router that decides which questions actually need expensive multi-hop retrieval and which can be answered directly, based on recoverability estimation. Multi-hop QA systems typically run costly retrieval on every question; RASER reduces unnecessary LLM calls by routing simple questions to direct answering. Practical for anyone building RAG pipelines: add a lightweight router to avoid burning tokens on questions your model already knows.