question decomposition for multi-hop
**Question Decomposition for Multi-Hop** is a reasoning strategy that breaks complex multi-hop questions into a sequence of simpler sub-questions, each answerable with a single retrieval or reasoning step, and chains the sub-answers together to reach the final answer. This decompose-then-solve approach makes multi-hop reasoning more interpretable, accurate, and debuggable by explicitly structuring the reasoning process into verifiable intermediate steps.
**Why Question Decomposition Matters in AI/ML:**
Question decomposition provides **interpretable, modular reasoning** that reduces the difficulty of multi-hop questions by converting them into sequences of manageable single-hop queries, improving both accuracy and the ability to verify each reasoning step.
• **Sequential decomposition** — A complex question like "What is the population of the country where the Taj Mahal is located?" decomposes into: Q1: "Where is the Taj Mahal located?" → A1: "India" → Q2: "What is the population of India?" → A2: Final answer
• **Decomposition models** — Trained decomposition models (e.g., DecompRC, Break-It-Down) or prompted LLMs generate sub-questions; few-shot prompting with decomposition examples enables GPT-4 and similar models to decompose questions zero-shot
• **Iterative retrieval** — Each sub-question triggers a separate retrieval step, using the sub-answer to inform subsequent queries; this iterative retrieve-and-reason process avoids the single-retrieval bottleneck that causes standard systems to miss bridge entities
• **Answer composition** — Sub-answers are composed through operations (comparison, union, intersection, arithmetic, boolean) defined by the question structure, with each operation verified independently for correctness
• **Self-ask prompting** — The Self-Ask framework prompts LLMs to explicitly ask and answer follow-up questions, generating intermediate reasoning steps that mimic question decomposition: "Follow up: [sub-question]? Intermediate answer: [sub-answer]"
| Method | Decomposition Source | Retrieval Strategy | Reasoning Type |
|--------|---------------------|-------------------|----------------|
| DecompRC | Trained decomposer | Per sub-question | Extractive span |
| Self-Ask | LLM prompting | Search engine per step | Generative |
| IRCoT | Interleaved with CoT | Iterative retrieval | Chain-of-thought |
| Least-to-Most | LLM prompting | Per sub-question | Sequential buildup |
| IRCOT + Decomp | Combined approach | Multi-step retrieval | Hybrid |
**Question decomposition is the most effective strategy for multi-hop reasoning, converting intractable complex questions into manageable sequences of simple sub-questions that can be independently answered and verified, providing interpretable reasoning chains that improve both accuracy and trustworthiness of multi-step question-answering systems.**