Title: Search-on-Graph: Iterative Informed Navigation for Large Language Model Reasoning on Knowledge Graphs

Abstract:

Integrating knowledge graphs (KGs) with large language models (LLMs) presents a powerful strategy for tackling knowledge-intensive reasoning tasks. The core of this methodology lies in identifying suitable reasoning paths within the KG. However, current approaches suffer from a significant drawback: path selection is typically handled by isolated modules that utilize criteria loosely aligned with actual reasoning needs. This disconnect frequently leads to the selection of erroneous relations or the premature elimination of pertinent paths.

To address this, we introduce Search-on-Graph (SoG), a novel approach that tightly couples path selection with the reasoning process. In SoG, the LLM itself determines which relations to pursue, guided by both the existing KG structure and the full context of the reasoning history. SoG operates on an observe-think-navigate framework. At every step, the LLM first observes the relational links accessible from the current entity, then evaluates which trajectory most effectively moves toward resolving the query, and finally navigates down that path. This context-sensitive navigation leverages the LLM’s inherent reasoning strengths, eliminating the need for independent selection modules that rely on proxy metrics. Evaluations across six knowledge graph question answering (KGQA) benchmarks reveal that SoG surpasses state-of-the-art techniques. Notably, it achieves this performance without requiring task-specific fine-tuning and demonstrates strong generalization across diverse KG schemas.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC