← AI Terminology
Multi-Hop Reasoning
Multi-hop reasoning requires combining multiple pieces of evidence or intermediate inferences to reach an answer — not solvable from one span alone.
It stresses both models and retrieval systems.
It stresses both models and retrieval systems.
Why It Matters in AI
Real questions chain facts (“Who is the CEO of the company that acquired X?”). Benchmarks and RAG designs target multi-hop explicitly; single-chunk retrieval is not enough. Agents and GraphRAG often exist for this reason.
Key Points
| Aspect | Description |
|---|---|
| Eval | Support fact coverage, not only final EM |
| Failure | Retrieve only first hop; miss bridge entity |
| Methods | Iterative retrieval, chain-of-thought, graphs |
| Related | Agentic RAG, knowledge graphs |
| Benchmarks | HotpotQA, MuSiQue, 2WikiMultiHop |
| Definition | Answer needs ≥2 evidence hops or steps |
Simple Analogy
Solving a treasure hunt where clue one only names the library, clue two inside names the shelf — you must complete both hops.
Common Usage Examples
- HotpotQA-style systems
- Retrieve → extract bridge → retrieve again
- Graph traversal over entities
- Decompose questions into sub-questions
Summary
In short: Multi-hop reasoning combines several evidence steps to answer — the hard case that breaks one-shot retrieval and shallow prompts.