REL-RAG: Relation-Aware Retrieval-Augmented Generation for Generalizable Knowledge Graph Question Answering

Large Language Models (LLMs) augmented with knowledge graphs (KGs) have been widely studied for knowledge graph question answering (KGQA). Graph-based retrievers exhibit strong empirical performance, but their generalization ability remains limited. In this work, we show that applying a line graph transformation to the KG provably enhances the generalizability of GNN-based retrievers. By elevating relations to first-class objects, line graphs encode relation transitions explicitly, and the resulting inductive bias aligns naturally with relational reasoning in KGs. This alignment makes multi-hop reasoning substantially easier to learn and improves generalizability across different types of distribution shifts. Building upon this representation, we propose REL-RAG, a framework that emphasizes relational reasoning for graph retrievers and is equipped with two complementary training objectives for flexible integration with LLMs. Path-based learning achieves higher precision with fewer tokens, making it especially suitable for smaller LLMs with limited context capacity. Triple-based learning encourages richer evidence diversity, which stronger LLMs can exploit more effectively with larger token budgets. Empirically, REL-RAG surpasses prior GNN-based retrieval method for KGQA by up to 18.10%.