SIRLM Enhances LLM Reasoning for Knowledge Graphs.

Knowledge Graph Reasoning (KGR) aims to uncover hidden facts by leveraging the intricate structural evidence within knowledge graphs (KGs). While large language models (LLMs) have shown promise in KGR through in-context learning, they often struggle with a "reasoning evidence perception drift," where their parametric knowledge doesn't fully align with the structural constraints of KGs, leading to less faithful and effective reasoning. To tackle this, researchers propose the Structure-Internalized Rule Language Model (SIRLM). SIRLM centers on generating structural rules to tightly couple the LLM's parametric learning with the faithfulness of its reasoning logic, ensuring it remains anchored to KG-grounded evidence. This is achieved through a Structure-Internalized Rule Generator (SIRG), which combines an in-context learning block with a structural relation memory. SIRG is further enhanced with a KG tokenizer based on structural invariance learning and a neuro-symbolic reasoner that provides rule-constrained message propagation and faithful rule-execution feedback. SIRLM can be integrated into standard LLM training paradigms and has demonstrated significant superiority over 17 state-of-the-art KGR methods across 36 datasets, proving its effectiveness in improving structural semantic understanding and reasoning faithfulness.