Multi-hop reasoning connects facts through intermediate steps rather than finding a complete answer in one passage. Standard RAG can fail when retrieval must identify one fact before using it to retrieve or infer the next. section 4.5.5
Chaining evidence across sources
For “What was the population of Thomas Jefferson’s birthplace?”, the system must identify Shadwell and then determine Shadwell’s population. A more complex chain identifies the signer of the Civil Rights Act of 1964 as Lyndon B. Johnson, finds that he was born in Texas, and correlates that state with the relevant twentieth-century president, Dwight D. Eisenhower. section 4.5.5 section 5.3
Fusion-in-Decoder handles this pattern by encoding many retrieved passages separately and letting the decoder attend jointly to them. Unlike an extractive span selector, it can combine evidence from documents about the event, person, birthplace, and candidate answer while filtering distractors. Increasing retrieved passages from 10 to 100 improved Exact Match by 6 points on TriviaQA and 3.5 on Natural Questions, with gains still rising at 100. section 5.3
Treat multi-hop reasoning as a pipeline failure rather than assuming retrieval alone is responsible.
- Check whether the required evidence is missing from the top retrieved results. section 4.5.5
- Compare one-pass retrieval with multi-hop RAG, chain-of-thought reasoning, or iterative retrieval when questions require chained facts. section 4.5.5
