RM2: Retrieve-More-Then-Read-More for Multi-Hop Question Answering
Year: 2025
Type: conference-paper
Abstract: Multi-hop question answering (MHQA) poses significant challenges for conventional retrieve-then-read approaches, primarily due to obstacles of comprehending interweaving documents and detecting relevant noises. Existing Retrieval-Augmented Generation (RAG) solutions tend to decompose a multi-hop question using Large Language Models (LLMs) with narrow use of external knowledge, vulnerably eliciting bias towards wrong answer. To more intensively harness external knowledge, we introduce Retrieve-More-then-Read-More (RM2), a novel two-stage pipeline. The first stage, retrieve-more, fine-tunes a retriever and re-ranker to dynamically explore intrinsic retrieval paths through expanding and pruning, to comprehensively gather related documents meanwhile constrain noises. In the subsequent read-more stage, we propose the Chain of Ground (CoG), which leverages LLMs to dynamically decompose sub-questions through glancing all retrieved documents, and then generate answer through accurately grounding with supporting documents when available. Furthermore, to reduce long token consumption during inference, we fine-tune LLMs in a self-provoking manner to execute the entire CoG process within a single call. Experimental results demonstrate that RM2 significantly outperforms state-of-the-art RAG methods on three MHQA datasets with lower inference costs.
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7127040788