Question Decomposition

Question decomposition studies how to split complex questions or tasks into simpler subquestions, constraints, or executable plans, then combine their results. It matters because multi-hop question answering, database querying, and retrieval over long or multilingual documents often require finding several pieces of evidence while preserving their dependencies. Common approaches use prompts or learned policies to create atomic questions or dependency graphs, then retrieve, verify, and iteratively refine evidence before synthesizing an answer; some systems adapt decomposition to retrieved evidence or task complexity. Decomposition can improve multi-hop QA and text-to-SQL, but poor splits, lost entities, extra latency, and accumulated errors can outweigh the gains, so deciding when and how to decompose remains a key challenge.