Structural Knowledge Induced Logical Learning for LLM Reasoning over Knowledge Graphs

Abstract

Knowledge Graphs (KGs) are fundamental for grounding Large Language Models (LLMs) in factually consistent reasoning across diverse data-mining tasks. However, when reasoning over complex graph topologies, LLMs can be brittle in maintaining structural consistency and may rely on spurious correlations. We propose SKILL (Structural Knowledge Induced Logical Learning), a novel paradigm that transforms relational rules from passive symbolic patterns into explicit, generation-time symbolic guidance for LLM-based reasoning. Unlike existing methods that isolate symbolic mining from neural inference, SKILL introduces a semantically aligned rule synthesis framework that filters spurious patterns and enables LLMs to leverage relational regularities during inference. By grounding generation in pruned, query-specific subgraph structures together with rule–path evidence, SKILL improves inductive generalization while preserving generative flexibility. Extensive experiments on three inductive benchmarks demonstrate that SKILL achieves state-of-the-art performance, with up to a 5% absolute improvement in Hits@1, highlighting a new and effective way to integrate symbolic structure into LLM reasoning over knowledge graphs. Source code is available at https://github.com/luilui97/SKILL.