DualRAG: A Dual-Process Approach to Integrate Reasoning and Retrieval for Multi-Hop Question Answering (2504.18243v1)

Abstract: Multi-Hop Question Answering (MHQA) tasks permeate real-world applications, posing challenges in orchestrating multi-step reasoning across diverse knowledge domains. While existing approaches have been improved with iterative retrieval, they still struggle to identify and organize dynamic knowledge. To address this, we propose DualRAG, a synergistic dual-process framework that seamlessly integrates reasoning and retrieval. DualRAG operates through two tightly coupled processes: Reasoning-augmented Querying (RaQ) and progressive Knowledge Aggregation (pKA). They work in concert: as RaQ navigates the reasoning path and generates targeted queries, pKA ensures that newly acquired knowledge is systematically integrated to support coherent reasoning. This creates a virtuous cycle of knowledge enrichment and reasoning refinement. Through targeted fine-tuning, DualRAG preserves its sophisticated reasoning and retrieval capabilities even in smaller-scale models, demonstrating its versatility and core advantages across different scales. Extensive experiments demonstrate that this dual-process approach substantially improves answer accuracy and coherence, approaching, and in some cases surpassing, the performance achieved with oracle knowledge access. These results establish DualRAG as a robust and efficient solution for complex multi-hop reasoning tasks.

This paper introduces DualRAG, a novel framework designed to improve Retrieval-Augmented Generation (RAG) for Multi-Hop Question Answering (MHQA) tasks (Cheng et al., 25 Apr 2025 ). MHQA requires reasoning over multiple pieces of information, often retrieved from external knowledge sources. Existing iterative RAG methods struggle with dynamically identifying when more knowledge is needed, determining what specific information to retrieve next, and effectively organizing the accumulating retrieved knowledge, which can be noisy and fragmented.

DualRAG addresses these issues through a synergistic dual-process architecture:

1. Reasoning-augmented Querying (RaQ): This process actively drives the reasoning forward.
   • It uses a Reasoner module ($\MR$) to analyze the current question ($x$), the aggregated knowledge ($K^{t-1}$), and the reasoning history ($R^{t-1}$) to perform a reasoning step ($r^t$).
   • Crucially, the Reasoner determines if the current knowledge is sufficient or if a knowledge gap exists, setting a retrieval trigger flag ($f^t$).
   • If retrieval is needed ($f^t = \text{True}$), an Entity Identifier module ($\MEI$) analyzes the reasoning context to identify key entities ($E^t$) relevant to the knowledge gap.
   • For each key entity ($e \in E^t$), it generates specific, targeted queries ($Q^t(e)$) designed to fill the identified gap. These queries are then used to retrieve documents ($D^t$) using a standard retriever ($\Retrieve$) followed by reranking ($\Rerank$).
2. Progressive Knowledge Aggregation (pKA): This process manages the retrieved knowledge.
   • It takes the reranked documents ($D_e$) retrieved by RaQ for each entity.
   • A Knowledge Summarizer module ($\MKS$) filters and summarizes these documents ($D_e$) based on the specific knowledge demand (guided by $x, R^t, e, Q^t(e)$), generating concise knowledge fragments ($k_e$). This step reduces noise and redundancy.
   • These fragments are then integrated into a Progressive Knowledge Outline ($K^t$). This outline maintains structured, entity-centric knowledge that accumulates over iterations ($K^t(e) = K^{t-1}(e) \cup \{ k_e \}$).

These two processes work in a closed loop: RaQ identifies knowledge needs and generates queries, guiding pKA; pKA filters, summarizes, and organizes the retrieved knowledge into a structured outline ($K^t$), which then provides a refined knowledge base for the next reasoning step in RaQ. This iterative refinement continues until the Reasoner determines no further retrieval is needed and the final answer ($\hat{a}$) can be generated by an Answer Generator ($\MA$) using the complete knowledge outline ($K^T$) and reasoning history ($R^T$).

Implementation Details:

• The framework relies on LLMs to implement the Reasoner, Entity Identifier, Knowledge Summarizer, and Answer Generator components, guided by specific prompts (detailed in Appendix Figs \ref{fig:prompt-reasoner}-\ref{fig:prompt-ks}).
• Standard retrieval components are used: a dense retriever (e.g., bge-small-en-v1.5) and a reranker (e.g., bge-reranker-v2-m3).
• The maximum number of iterations is typically capped (e.g., 5 steps in the experiments).

Fine-Tuning for Compact Models:

To make DualRAG practical with smaller, less computationally expensive LLMs, the paper proposes a fine-tuning strategy:

• A large "teacher" LLM (Qwen2.5-72B-Instruct) is used with DualRAG to generate high-quality reasoning and retrieval trajectories on MHQA training datasets.
• A specialized dataset is derived from these trajectories, targeting three key capabilities where smaller models falter:
  • Reasoner: Learning when to trigger retrieval.
  • Entity Identifier: Generating more precise, less redundant queries and performing entity linking.
  • Knowledge Summarizer: Better identifying relevant information within retrieved documents, especially implicit connections.
• Fine-tuning a smaller "student" LLM (e.g., Qwen2.5-7B-Instruct) on this dataset (DualRAG-FT) significantly improves its performance within the DualRAG framework, preserving the core advantages while reducing computational cost.

Evaluation and Results:

• Experiments on HotpotQA, 2WikiMultihopQA, and MuSiQue show that DualRAG significantly outperforms baseline RAG methods (NativeRAG, IRCoT, MetaRAG, GenGround), approaching or exceeding oracle performance in some cases (Table \ref{tab:exp-main}).
• DualRAG-FT demonstrates substantial gains over the base small model, making it highly competitive (Table \ref{tab:exp-main}).
• Ablation studies confirm the necessity of each component (Reasoner's gap detection, Entity Identifier's targeted queries, pKA's knowledge outlining) and the effectiveness of the fine-tuning strategy (Tables \ref{tab:exp-abl-framework}, \ref{tab:exp-abl-sft}).
• The framework also shows strong performance on single-hop and long-form QA tasks (Tables \ref{tab:exp-single-hop}, \ref{tab:exp-long-format}).

In summary, DualRAG provides a structured, iterative approach to RAG for complex QA by tightly coupling reasoning-driven query generation (RaQ) with progressive, entity-centric knowledge organization (pKA). Its ability to dynamically identify and target knowledge gaps and manage retrieved information leads to improved accuracy and coherence. The proposed fine-tuning strategy further enhances its practical applicability by enabling effective deployment on smaller LLMs.