SIRLM: New Structure-Internalized Rule Language Model Aims to Align LLMs with Knowledge Graph Constraints
A recent paper published on arXiv presents the Structure-Internalized Rule Language Model (SIRLM), designed to address the issue of reasoning evidence perception drift in large language models (LLMs) within Knowledge Graph Reasoning (KGR). KGR focuses on uncovering hidden facts through structural evidence found in knowledge graphs. The authors highlight that, while LLMs have made strides in KGR tasks, inconsistencies between the representation in knowledge graphs and the parametric knowledge of LLMs create challenges. This discrepancy results in 'reasoning evidence perception drift.' SIRLM aims to generate structural rules to synchronize parametric learning with reasoning mechanisms. The paper, identified as arXiv:2608.17443, seeks to improve the reliability and accuracy of automated reasoning, which is crucial for applications such as semantic search and question answering.
Key facts
- The paper is available on arXiv under the identifier 2608.17443.
- The paper's announcement type is 'new'.
- Knowledge Graph Reasoning (KGR) aims to discover latent facts using structural evidence in KGs.
- Large Language Models (LLMs) have shown progress on KGR tasks via in-context learning.
- The problem of representation inconsistency between KG structure and LLM parametric knowledge is identified.
- This inconsistency causes what the authors call 'reasoning evidence perception drift' in LLMs over KGs.
- SIRLM stands for Structure-Internalized Rule Language Model.
- SIRLM centers on structural rule generation to couple parametric learning of structural knowledge with reasoning.
Entities
Institutions
- arXiv