Retrieval--Reasoning Processes for Multi-hop Question Answering: A Four-Axis Design Framework and Empirical Trends

TL;DR

This work argues that multi-hop QA must be analyzed through an explicit retrieval–reasoning procedure. It introduces a four-axis framework—overall execution plan, index structure, next-step control, and stop/continue criteria—to describe and compare retrieval–reasoning systems across model families. Through a survey of 104 detailed-method papers, it documents empirical trends: explicit multi-step execution, graph-based indices, and verifier-guided stopping often yield better accuracy and faithfulness, especially under long-horizon or noisy retrieval settings. The paper highlights open challenges in aligning plans with indices, scalable structured retrieval, generalizable control policies, and robust stopping under distribution shift, aiming to standardize evaluation and design in retrieval–reasoning agents for QA. The framework provides a principled basis for organizing existing methods and informing future research directions with an emphasis on procedure-level analysis and comparability.

Abstract

Multi-hop question answering (QA) requires systems to iteratively retrieve evidence and reason across multiple hops. While recent RAG and agentic methods report strong results, the underlying retrieval--reasoning \emph{process} is often left implicit, making procedural choices hard to compare across model families. This survey takes the execution procedure as the unit of analysis and introduces a four-axis framework covering (A) overall execution plan, (B) index structure, (C) next-step control (strategies and triggers), and (D) stop/continue criteria. Using this schema, we map representative multi-hop QA systems and synthesize reported ablations and tendencies on standard benchmarks (e.g., HotpotQA, 2WikiMultiHopQA, MuSiQue), highlighting recurring trade-offs among effectiveness, efficiency, and evidence faithfulness. We conclude with open challenges for retrieval--reasoning agents, including structure-aware planning, transferable control policies, and robust stopping under distribution shift.

Figures (4)

• Figure 1: Overview of the four design axes we use to describe retrieval--reasoning processes in multi-hop QA: (A) overall execution plan, (B) index structure, (C) next step execution plan, and (D) stop/continue criteria. The framework abstracts over concrete model architectures and datasets.
• Figure 2: overall execution plans for retrieval--reasoning in multi-hop QA (Axis A of our framework). (a) Retrieve--then--Read: a single retrieval pass feeds the reader model. (b) Interleaved: the model alternates reasoning with additional retrieval calls, updating a partial answer/state. (c) Plan--then--Execute: the model first decomposes the question into sub-hops, then retrieves and answers for each hop following the plan. (d) Test-time search scaling: the model explores multiple candidate reasoning trajectories over the question/sub-questions and selects the best-scoring path for retrieval and answering.
• Figure 3: Contrasts flat passage lists, hierarchical / tree-structured indices, graph-based indices, and long-context storage.
• Figure 4: The large topics and how these topics develop from 2023 to 2025.