MultiCube-RAG for Multi-hop Question Answering

TL;DR

This work introduces MultiCube-RAG, a training-free framework that uses ontology-guided, multi-dimensional cubes to model subjects, attributes, and relations for improved multi-hop QA. By decomposing complex queries into one-hop subqueries and routing each to specialized cube retrievers, it achieves robust reasoning with more efficient retrieval and inherent explainability. The approach demonstrates accuracy gains (average ~8.9 percentage points) over baselines across four datasets, along with favorable retrieval efficiency and provenance. Its cube-based design offers scalable, domain-adaptive retrieval with reduced noise compared to graph-based methods, making it practical for high-stakes contexts requiring transparent reasoning.

Abstract

Multi-hop question answering (QA) necessitates multi-step reasoning and retrieval across interconnected subjects, attributes, and relations. Existing retrieval-augmented generation (RAG) methods struggle to capture these structural semantics accurately, resulting in suboptimal performance. Graph-based RAGs structure such information in graphs, but the resulting graphs are often noisy and computationally expensive. Moreover, most methods rely on single-step retrieval, neglecting the need for multi-hop reasoning processes. Recent training-based approaches attempt to incentivize the large language models (LLMs) for iterative reasoning and retrieval, but their training processes are prone to unstable convergence and high computational overhead. To address these limitations, we devise an ontology-based cube structure with multiple and orthogonal dimensions to model structural subjects, attributes, and relations. Built on the cube structure, we propose MultiCube-RAG, a training-free method consisting of multiple cubes for multi-step reasoning and retrieval. Each cube specializes in modeling a class of subjects, so that MultiCube-RAG flexibly selects the most suitable cubes to acquire the relevant knowledge precisely. To enhance the query-based reasoning and retrieval, our method decomposes a complex multi-hop query into a set of simple subqueries along cube dimensions and conquers each of them sequentially. Experiments on four multi-hop QA datasets show that MultiCube-RAG improves response accuracy by 8.9% over the average performance of various baselines. Notably, we also demonstrate that our method performs with greater efficiency and inherent explainability.

Figures (5)

• Figure 1: An example of an ontology structure extracted by the LLM. $\mathcal{S}, \mathcal{A}, \mathcal{R}$ denote a set of subjects, attributes, and relations, respectively.
• Figure 2: An ontology-guided cube ("Culture Product") and document allocation into cube cell. We show only one of the attributes (i.e., release date) here; more attributes (e.g., genre, rating), and relations to locations (e.g., filmed at, release at) are included as other cube dimensions but are not shown up. Different dimensions are highlighted with distinct colors.
• Figure 3: MultiCube-RAG for multi-step reasoning and retrieval. It includes multiple ontology-guided cubes (A, B, C) and supports selecting the most suitable one for retrieval at each iteration. The dimensional values serve as search keys; the cube cells with red checkmarks are activated using the search keys in purple. While the "Location Cube" is not used in this example, it is useful for other queries. For abbreviation, SQ: subquery, SA: subquery answer, #: iteration number, and T: retrieved documents.
• Figure 4: MultiCube-RAG for Multi-turn Reasoning and Retrieval. [mycircled]a11 Prompting LLMs to reason/generate subqueries; [mycircled]a12 MultiCube-RAG for each subquery; [mycircled]a13 Aggregating intermediate subqueries and their answers; [mycircled]a14 Determining to output final answer or enter iterative reasoning and retrieval loop.
• Figure 5: Activated cubes and cube dimensions in MultiCube-RAG to represent the explainable retrieval process.