multi-hop reasoning
**Multi-Hop Reasoning** is a complex inference paradigm where answering a question requires combining information from multiple distinct evidence sources or performing multiple sequential reasoning steps, each building on conclusions drawn from previous steps. Unlike single-hop QA (where the answer exists in a single passage), multi-hop reasoning demands that the model identify, retrieve, and logically chain multiple pieces of evidence to arrive at the final answer.
**Why Multi-Hop Reasoning Matters in AI/ML:**
Multi-hop reasoning is a **critical capability gap** in current AI systems, requiring compositional generalization, evidence tracking, and logical chaining that pushes beyond the pattern-matching capabilities of standard retrieval and QA approaches.
• **Bridge entities** — Multi-hop questions require identifying intermediate entities that connect the question to the answer: "Where was the director of Inception born?" requires first identifying the director (Christopher Nolan) then finding his birthplace (London)—the director is the bridge entity connecting two facts
• **Compositional reasoning** — Answers require composing multiple atomic facts through logical operations: comparison ("Which is taller, the Eiffel Tower or Big Ben?"), intersection ("Which actor appeared in both Film A and Film B?"), or sequential deduction across evidence chains
• **Evidence chain construction** — The model must identify and order 2-4 supporting passages that form a logical chain: Passage 1 → intermediate conclusion → Passage 2 → intermediate conclusion → final answer, with each step depending on previous conclusions
• **Reasoning shortcuts** — Models often exploit lexical overlap and entity co-occurrence to guess correct answers without genuine multi-hop reasoning (shortcut reasoning); adversarial evaluation and reasoning chain verification are needed to detect this
• **Benchmark datasets** — HotpotQA, MuSiQue, 2WikiMultiHopQA, and StrategyQA provide standardized multi-hop evaluation with annotated supporting facts and reasoning chains for training and evaluation
| Dataset | Hops | Task Type | Evidence | Reasoning Skills |
|---------|------|-----------|----------|-----------------|
| HotpotQA | 2 | Extractive QA | 2 Wikipedia passages | Bridge, comparison |
| MuSiQue | 2-4 | Extractive QA | 2-4 passages | Composition, intersection |
| 2WikiMultiHopQA | 2-5 | Extractive QA | Wikipedia | Bridge, comparison, inference |
| StrategyQA | 2-5 | Yes/No | Implicit decomposition | Strategy, world knowledge |
| FEVER | 1-3 | Verification | Wikipedia | Entailment, multi-evidence |
**Multi-hop reasoning represents one of the most challenging frontiers in AI question answering, requiring models to perform genuine compositional inference across multiple evidence sources and reasoning steps rather than relying on statistical shortcuts, making it a critical benchmark for measuring progress toward human-level language understanding.**