Re$^{3}$: Retrieval, Reranking, and Reasoning for Agentic Knowledge Graph Question Answering

ICLR 2026 Conference Submission17096 Authors

Keywords: Knowledge Graph, Question Answering, Large Language Models

TL;DR: Retrieval, Reranking, and Reasoning for Agentic Knowledge Graph Question Answering

Abstract: Recently, the integration of Large Language Models (LLMs) and knowledge graphs has emerged as a promising approach for knowledge graph question answering by enhancing the reasoning capability in knowledge-intensive applications. However, existing methods face a key trade-off: they either introduce high computational costs when LLMs reason directly on graphs, or suffer from poor reasoning quality due to over-reliance on retrieval methods. To mitigate these issues, we introduce a computationally efficient framework based on Retrieval, Reranking, and Reasoning (Re$^3$). Specifically, we first develop "cognitively-informed retrieval" that improves subgraph retrieval quality via Question-Entity (Q-E) discrepancy scoring and hierarchical information aggregation. Second, we propose path-aware reranking, which employs lightweight cross-encoders to evaluate and prune reasoning paths efficiently. Third, we apply "agentic reasoning" to perform autonomous reasoning on high-quality subgraphs while balancing reasoning quality and computational overhead. Extensive experimental results on WebQSP and CWQ demonstrate that Re$^3$ significantly outperforms existing methods.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 17096