Title: DTKG: A Dual-Track Knowledge Graph-Verified Reasoning Framework for Multi-Hop QA

Original: arXiv:2510.16302v2 Announce Type: replace

Abstract: In the context of retrieval-augmented generation (RAG) for contemporary large language models (LLMs), multi-hop reasoning is essential for question answering (QA). By extracting the relational structures of entities from knowledge graphs (KGs), precise answers can be derived. Based on their inherent reasoning patterns and relation dependencies, multi-hop reasoning generally falls into two distinct types: i) parallel fact-verification questions, which necessitate the concurrent validation of several independent sub-questions; and ii) chained multi-hop questions, which require sequential, multi-step inference where intermediate conclusions act as necessary premises for later steps.

Currently, existing multi-hop reasoning methods typically rely on just one of two techniques: fact verification based on LLM responses or chain construction via KG paths. However, this single-approach strategy has notable drawbacks. The LLM response-based method performs well in parallel fact-verification but struggles with chained reasoning tasks. Conversely, the KG path-based method is effective for chained multi-hop reasoning but encounters issues with redundant path retrieval when dealing with parallel fact-verification. These shortcomings negatively impact both the efficiency and accuracy of multi-hop QA tasks.

To overcome these challenges, we introduce DTKG, a novel dual-track KG verification and reasoning framework inspired by the Dual Process Theory from cognitive science. DTKG operates through two primary phases: the Classification Stage and the Branch Processing Stage.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC