# GraphReader: Building Graph-based Agent to Enhance Long-Context Abilities of Large Language Models

**Authors:** Shilong Li, Yancheng He, Hangyu Guo, Xingyuan Bu, Ge Bai, Jie Liu, Jiaheng Liu, Xingwei Qu, Yangguang Li, Wanli Ouyang, Wenbo Su, Bo Zheng  
**arXiv:** [2406.14550](https://arxiv.org/abs/2406.14550)  
**Date:** 2024-06-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2406.14550v1/

## Key points

- Authors introduce GraphReader as a solution for challenges faced by large language models (LLMs) in processing long inputs effectively
- GraphReader utilizes a graph-based agent to structure long texts into a graph for exploration
- The agent conducts systematic analysis, devises strategic plan, and navigates through node content and neighbors in the graph when presented with a question
- Coarse-to-fine exploration strategy allows continuous gathering of insights and path adjustments until accurate answer is generated
- Experimental results on LV-Eval dataset show GraphReader outperforms GPT-4-128k significantly across context lengths from 16k to 256k using a 4k context window
- Superior performance demonstrated on challenging single-hop and multi-hop benchmarks
- Study spans 27 pages and highlights the potential of graph-based agents in enhancing long-context capabilities of LLMs

## Summary

In their paper titled "GraphReader: Building Graph-based Agent to Enhance Long-Context Abilities of Large Language Models," authors Shilong Li, Yancheng He, Hangyu Guo, Xingyuan Bu, Ge Bai, Jie Liu, Jiaheng Liu, Xingwei Qu, Yangguang Li, Wanli Ouyang, Wenbo Su and Bo Zheng introduce GraphReader as a solution to the challenges faced by large language models (LLMs) in processing long inputs effectively. The system utilizes a graph-based agent that autonomously structures long texts into a graph for exploration. When presented with a question, the agent conducts a systematic analysis and devises a strategic plan before employing predefined functions to navigate through node content and neighbors in the graph. This approach allows for a coarse-to-fine exploration strategy where the agent continuously gathers insights and adjusts its path based on current circumstances until it has gathered enough information to generate an accurate answer. Experimental results on the LV-Eval dataset demonstrate that GraphReader outperforms GPT-4-128k significantly across context lengths ranging from 16k to 256k when using a 4k context window. Furthermore, the system showcases superior performance on challenging single-hop and multi-hop benchmarks. The collaborative effort of the authors resulted in a comprehensive study spanning 27 pages that sheds light on the potential of graph-based agents in enhancing the long-context capabilities of LLMs.