Query Decomposition — References

Multi-hop and compositional question answering, and the interleaving of retrieval with reasoning that it requires.

Multi-hop question answering

• HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering — Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov and Christopher D. Manning, EMNLP, 2018. The dataset that made multi-hop retrieval measurable, including supporting-fact annotations that separate retrieval failure from reasoning failure. Cited by name.
• Constructing Datasets for Multi-hop Reading Comprehension Across Documents — Johannes Welbl, Pontus Stenetorp and Sebastian Riedel, Transactions of the Association for Computational Linguistics, 2018. Questions whose evidence is deliberately spread across documents, and the argument that no single passage suffices. Cited by name.
• MuSiQue: Multihop Questions via Single-hop Question Composition — Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot and Ashish Sabharwal, Transactions of the Association for Computational Linguistics, 2022. Composed questions built to resist shortcut answering, which is the failure mode a weak decomposition exhibits. Cited by name.

Producing the sub-questions

• Least-to-Most Prompting Enables Complex Reasoning in Large Language Models (Zhou et al., 2022), arXiv:2205.10625 — explicitly producing sub-problems before solving them, the prompting basis for a generated retrieval plan.
• Measuring and Narrowing the Compositionality Gap in Language Models — Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith and Mike Lewis, Findings of EMNLP, 2023. The self-ask formulation and the measurement that models answer each hop correctly while failing the composition. Cited by name.

Interleaving retrieval with reasoning

• Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions — Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot and Ashish Sabharwal, ACL,
  2023. Retrieving between reasoning steps rather than once at the start, which is the iterative plan in its published form. Cited by name.
• Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (Wei et al., 2022), arXiv:2201.11903 — the baseline that decomposition is measured against, and the reason intermediate steps are available to retrieve on at all.

No verified reference is recorded for fan-out and hop budgets, for constraint propagation into sub-queries, or for representing an unevidenced branch in a synthesised answer. The literature measures whether decomposition answers the question, not what it costs to run in a product.