# Large Language Models on Graphs: A Comprehensive Survey

**Authors:** Bowen Jin, Gang Liu, Chi Han, Meng Jiang, Heng Ji, Jiawei Han  
**arXiv:** [2312.02783](https://arxiv.org/abs/2312.02783)  
**Date:** 2023-12-05  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2312.02783v1/

## Key points

- Large language models (LLMs) like ChatGPT and LLaMA have made advancements in natural language processing
- LLMs are primarily designed for processing pure texts
- Graphs provide rich structural information in real-world scenarios such as academic networks and e-commerce networks
- Graph data can be paired with textual information, like molecules with descriptions
- The survey paper titled "Large Language Models on Graphs" provides a systematic review of adopting LLMs on graphs
- Potential scenarios are categorized into three categories: pure graphs, text-rich graphs, and text-paired graphs
- Techniques discussed include using LLMs as predictors, encoders, and aligners for graph data or textual information
- Advantages and disadvantages of different models within these techniques are compared
- Real-world applications and available open-source codes and benchmark datasets are mentioned
- The survey paper highlights potential future research directions in this field.

## Summary

Large language models (LLMs), such as ChatGPT and LLaMA, have made significant advancements in natural language processing by leveraging their strong text encoding/decoding ability and emergent reasoning capability. However, while LLMs are primarily designed for processing pure texts, there are numerous real-world scenarios where text data is associated with rich structural information in the form of graphs, such as academic networks and e-commerce networks. Additionally, there are scenarios where graph data is paired with textual information, like molecules with descriptions. In this comprehensive survey paper titled "Large Language Models on Graphs," authors Bowen Jin, Gang Liu, Chi Han, Meng Jiang, Heng Ji, and Jiawei Han provide a systematic review of various scenarios and techniques related to adopting LLMs on graphs. The authors categorize potential scenarios into three main categories: pure graphs, text-rich graphs, and text-paired graphs. The paper discusses detailed techniques for utilizing LLMs on graphs. These techniques include using LLMs as predictors to make predictions based on graph data or textual information associated with the graph; employing LLMs as encoders to encode graph structures or textual information into vector representations; and utilizing LLMs as aligners to align graph structures or textual information. The advantages and disadvantages of different models within these schools of thought are compared. The authors also mention real-world applications of these methods and provide a summary of open-source codes and benchmark datasets available in this field. Overall, this survey paper sheds light on the underexplored area of applying large language models to graph-based reasoning. It concludes by highlighting potential future research directions in this rapidly growing field.