REEF: Relation-Aware Graph Foundation Model
🎉 Accepted at NeurIPS 2026.
REEF is a graph foundation model (GFM) that treats relations as the fundamental transferable unit of graphs. Each relation type (a citation link, a molecular bond, a knowledge-graph predicate, …) is described in natural language, encoded by a sentence encoder, and turned by three hypernetworks into:
- a relation-specific aggregator for message passing,
- a relation-specific classifier for downstream prediction, and
- a dataset-specific projector + feature bias that adapts to different feature distributions.
A single pretrained model therefore transfers across datasets, domains, and tasks (node classification, link prediction) in zero-shot and few-shot settings.
Files in this repository
| File | Size | Description |
|---|---|---|
reef_pretrained.pth | ~677 MB | Pretrained state_dict of the GFM model (PyTorch pickle). |
edge_texts2id.json | ~30 KB | Relation-text → id mapping (254 relations) used to look up relation tokens. |
Training data
Pretrained on graphs spanning four domains:
- Knowledge graphs: FB15K237, WN18RR
- Citation networks: Citeseer, Pubmed
- Web pages (heterophilous): Texas, Wisconsin
- Co-purchase: Photo
Relation descriptions are encoded with Sentence-BERT (all-MiniLM-L6-v2). The relation vocabulary contains 254 relations (237 from FB15K237, 11 from WN18RR, and 6 domain-level descriptions), each with an LLM-generated textual description.
Citation
@article{yu2025relation,
title={Relation-Aware Graph Foundation Model},
author={Yu, Jianxiang and Zhu, Jiapeng and Qian, Hao and Liu, Ziqi and Zhang, Zhiqiang and Li, Xiang},
journal={arXiv preprint arXiv:2505.12027},
year={2025}
}