# Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models

**Authors:** Linhao Luo, Zicheng Zhao, Chen Gong, Gholamreza Haffari, Shirui Pan  
**arXiv:** [2410.13080](https://arxiv.org/abs/2410.13080)  
**Date:** 2024-10-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.13080v1/

## Key points

- Authors propose a novel framework called graph-constrained reasoning (GCR) to address challenges faced by large language models (LLMs) in achieving faithful reasoning.
- GCR integrates structured knowledge from knowledge graphs (KGs) into LLMs through KG-Trie, a trie-based index encoding KG reasoning paths for accurate KG-grounded reasoning.
- The framework employs a lightweight KG-specialized LLM alongside a general LLM for inductive reasoning over multiple paths, ensuring accurate reasoning without hallucinations.
- Extensive experiments on Knowledge Graph Question Answering benchmarks show that GCR achieves state-of-the-art performance and strong zero-shot generalizability to unseen KGs without additional training.

## Summary

, , , , 

In their paper titled "Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models," authors Linhao Luo, Zicheng Zhao, Chen Gong, Gholamreza Haffari, and Shirui Pan propose a novel framework called graph-constrained reasoning (GCR) to address the challenges faced by large language models (LLMs) in achieving faithful reasoning. LLMs often struggle with knowledge gaps and hallucinations, making it difficult to generate accurate reasoning paths. To overcome these limitations, the authors integrate structured knowledge from knowledge graphs (KGs) into LLMs through GCR. The key innovation of GCR is its use of KG-Trie, a trie-based index that encodes KG reasoning paths for faithful KG-grounded reasoning. This allows for direct reasoning on graphs and generates accurate paths grounded in KGs. Additionally, GCR employs a lightweight KG-specialized LLM alongside a powerful general LLM for inductive reasoning over multiple paths. This dual approach ensures accurate reasoning without any hallucinations. Through extensive experiments on various Knowledge Graph Question Answering (KGQA) benchmarks, the authors demonstrate that GCR achieves state-of-the-art performance and exhibits strong zero-shot generalizability to unseen KGs without requiring additional training. The framework not only addresses the limitations of existing KG-enhanced methods but also showcases the potential of integrating structured knowledge from KGs with large language models for improved reasoning capabilities.