README.md

metadata

task_categories:
  - graph-ml

LoReC: Rethinking Large Language Models for Graph Data Analysis

This repository contains data associated with the paper LoReC: Rethinking Large Language Models for Graph Data Analysis.

GitHub Repository

Introduction

LoReC (Look, Remember, Contrast) is a novel plug-and-play method for the GraphLLM paradigm. It enhances the understanding of graph data in Large Language Models through three stages:

1. Look: Redistributing attention to the graph.
2. Remember: Re-injecting graph information into the Feed-Forward Network (FFN).
3. Contrast: Rectifying the vanilla logits produced in the decoding process.

This method improves performance on graph-related tasks without requiring extra fine-tuning.

Usage

The graph data in this project is typically stored in PyTorch Geometric (PyG) format within .pt files. You can load the data as follows:

import torch as th

# Load the graph data
graph_data = th.load('all_graph_data.pt')

# Access specific dataset components (e.g., 'arxiv')
arxiv_graph = graph_data['arxiv']

Citation

@misc{zhan2026lorecrethinkinglargelanguage,
      title={LoReC: Rethinking Large Language Models for Graph Data Analysis}, 
      author={Hongyu Zhan and Qixin Wang and Yusen Tan and Haitao Yu and Jingbo Zhou and Shuai Chen and Jia Li and Xiao Tan and Jun Xia},
      year={2026},
      eprint={2604.17897},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2604.17897}, 
}