Exploring the interplay between LLMs and graph-based knowledge, this study initiates Graph Chain-of-Thought (Graph-CoT). The framework aids in augmenting LLMs with graph reasoning abilities, leveraging knowledge encoded in interconnected texts to answer complex queries.

• GRBench dataset created, with real-world domain graphs
• Graph-CoT employs LLM reasoning, interaction, and graph execution iteratively
• Consistent outperformance over baseline models
• Framework supports multi-step reasoning tasks

The Graph-CoT methodology is a stride toward combining structured and unstructured data in AI reasoning. It’s a move towards models that can acclimate to rich data types beyond plain text, broadening the AI’s understanding of intricate information networks with applications in scientific research, marketing analysis, and more.