Towards combining multiple knowledge graphs for large language models reasoning

Hong Yang

doi:10.1016/j.knosys.2026.116307

Towards combining multiple knowledge graphs for large language models reasoning

2026-06-01

1

PRE

AI

Shirui Pan (Shirui Pan)

DOI：10.1016/j.knosys.2026.116307

Original

Original request for help

Save

Abstract

En 中文

Large language models (LLMs) have achieved remarkable success in reasoning tasks. However, LLMs often produce hallucinations and incorrect reasoning results due to lack of up-to-date domain knowledge. Knowledge graphs build structured networks of entity relationships, providing LLMs with reliable background knowledge to reduce errors. However, existing LLM reasoning methods based on knowledge graphs assume that only a single knowledge graph is available, which often leads to missing key facts and biased information. In this paper, we propose to combine multiple knowledge graphs for LLM reasoning (MKG4LLM for short) to improve the accuracy and reliability of LLM reasoning. First, MKG4LLM sets a learnable weight for each knowledge graph and then maximizes the likelihood function of the LLM reasoning based on multiple weighted knowledge graphs. To learn both the weight parameters of the knowledge graphs and the fine-tuned LLM parameters for reasoning, a variational inference algorithm is used to approximately optimize these weight parameters. By retrieving weighted reasoning paths from the knowledge graphs, MKG4LLM extracts useful knowledge for LLM reasoning. The experimental results show that MKG4LLM outperforms existing methods in popular KGQA datasets.

Keywords:

Large language models

Multiple knowledge graphs

LLM reasoning

Towards combining multiple knowledge graphs for large language models reasoning

Abstract

Journal

Organization

Cited Papers