Multi-hop retrieval-augmented generation (RAG) is a task where an agent must retrieve evidence for a sequence of dependent sub-questions. When the final answer is wrong, it often lacks a required passage. A new arXiv preprint, "Beyond Retrieval Failure: Understanding Why LLMs Miss Evidence in Multi-Hop RAG," investigates the causes of this missing evidence.
The abstract notes that one reason a required passage may be absent is that it is inaccessible to the retriever. It also signals that the work goes beyond retrieval failure to consider other explanations, though the full details are not included in the abstract. The paper is a preprint, so it has not yet undergone peer review.
Because the abstract is truncated, only part of the analysis is visible. Still, the framing makes clear that the authors aim to distinguish between cases where the retriever simply cannot find the passage and cases where the LLM fails to use evidence that was actually retrieved.