BELLE: A Bi-Level Multi-Agent Reasoning Framework for Multi-Hop Question Answering

TL;DR

BELLE addresses the challenge of multi-hop QA by aligning question type with an operator-based reasoning workflow and solving plan via a bi-level multi-agent debate. The framework comprises a question type classifier, a first-level debate to select operator plans, a second-level memory-informed debate to evaluate those plans, and an executor that applies the chosen operators to generate sub-answers and a final response. Empirical results across four open-domain datasets show BELLE achieving state-of-the-art performance and favorable token-efficiency, with notable gains on harder 2- to 4-hop questions, especially MuSiQue. The work also analyzes how operator combinations vary by question type and demonstrates that second-level reflective debate substantially improves plan quality while reducing unnecessary retrieval. Overall, BELLE offers a model-agnostic, cost-aware approach to dynamic operator composition in complex QA tasks, with promising implications for real-world deployment and adaptation to new question types.

Abstract

Multi-hop question answering (QA) involves finding multiple relevant passages and performing step-by-step reasoning to answer complex questions. Previous works on multi-hop QA employ specific methods from different modeling perspectives based on large language models (LLMs), regardless of the question types. In this paper, we first conduct an in-depth analysis of public multi-hop QA benchmarks, dividing the questions into four types and evaluating five types of cutting-edge methods for multi-hop QA: Chain-of-Thought (CoT), Single-step, Iterative-step, Sub-step, and Adaptive-step. We find that different types of multi-hop questions have varying degrees of sensitivity to different types of methods. Thus, we propose a Bi-levEL muLti-agEnt reasoning (BELLE) framework to address multi-hop QA by specifically focusing on the correspondence between question types and methods, where each type of method is regarded as an ''operator'' by prompting LLMs differently. The first level of BELLE includes multiple agents that debate to obtain an executive plan of combined ''operators'' to address the multi-hop QA task comprehensively. During the debate, in addition to the basic roles of affirmative debater, negative debater, and judge, at the second level, we further leverage fast and slow debaters to monitor whether changes in viewpoints are reasonable. Extensive experiments demonstrate that BELLE significantly outperforms strong baselines in various datasets. Additionally, the model consumption of BELLE is higher cost-effectiveness than that of single models in more complex multi-hop QA scenarios.

Figures (10)

• Figure 1: Comparison between our approach and existing methods for multi-hop QA. (1) Closed-book reasoning does not consider the requirement for external knowledge. (2) Retrieval-augmented reasoning leverages an end-to-end fixed solution to solve all multi-hop questions. (3) Our agent-based reasoning framework provides an execution plan to dynamically combine appropriate multi-hop operators with respect to multi-hop question types.
• Figure 2: Comparison of single and combined operators in different multi-hop question types. The red and purple bars represent the combined operators of sub-step + single-step and sub-step + iterative-step, respectively.
• Figure 3: Model overview of BELLE. The left part is the existing MAD system containing three basic roles (i.e., an affirmative side, a negative side and a judge). The right part is the details of our bi-level MAD system including first-level and second-level debaters.
• Figure 4: Results of different question types in terms of F1 (%).
• Figure 5: Changes in the selection of combined operators. $\mathrm{MAD}_i^j$ denotes the debate stage at $i$-level and $j$-th debate round. (Best viewed in color.)