Multi-hop retrieval failures follow predictable structural patterns that vary by dataset; you can detect high-risk queries using simple features (query length, retrieval concentration) and a learned confidence score, enabling safe abstention without additional LLM calls.
This paper identifies why multi-hop retrieval systems fail predictably on certain queries and proposes a method to detect these failures without extra LLM calls. The authors prove that failure patterns cluster in specific subpopulations and that different query features predict failure in different scenarios.