TAIKC: Tree-Organized Active Internal Knowledge Completion for Multi-Hop Question Answering

ACL ARR 2025 May Submission2903 Authors

Abstract: Iterative dynamic Retrieval-Augmented Generation (RAG) methods have demonstrated strong performance on Multi-Hop Question Answering (MHQA). However, they still suffer from high inference costs, redundant information processing, and retrieval decisions that depend heavily on internal states. To this end, we propose Tree-Organized Active Internal Knowledge Completion (TAIKC), a novel approach designed to address two significant challenges: efficient information aggregation and active retrieval decision-making. TAIKC hierarchically decomposes multi-hop questions into a tree of sub-questions. For each sub-question, the model either extracts confident internal knowledge based on its perception of knowledge boundaries or leverages external knowledge to fill the knowledge gap. This process incrementally constructs a knowledge tree that integrates both internal and external information, and knowledge chains are then induced from the knowledge tree to solve the complex question. Furthermore, we align the model with our framework via knowledge distillation and model bootstrapping. Extensive experiments on four MHQA datasets demonstrate the effectiveness of our method.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: multihop QA;open-domain QA;

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 2903