IC-996GPT-3.5 achieves 73.5% zero-shot accuracy on OGBN-ARXIV 40-class node classification and 73.56% on TAPe-ARXIV23, a dataset of papers published after its knowledge cutoff

Xiaoxin He, Xavier Bresson, Thomas Laurent, Adam Perold, Yann LeCun, Bryan Hooi

SourceHarnessing Explanations: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning

The paper prompts GPT-3.5 (gpt-3.5-turbo) with paper titles and abstracts in a zero-shot setting, asking it to predict the arXiv CS sub-category and provide an explanation. On OGBN-ARXIV (169,343 nodes, 40 classes) it reaches 73.5% accuracy, surpassing RevGAT with OGB shallow features (70.8%) but falling short of GLEM (76.6%). On TAPe-ARXIV23, a citation graph of 46,198 papers published in 2023 (beyond GPT-3.5's November 2022 knowledge cutoff), it still achieves 73.56%, indicating generalization to unseen papers.

Evidence
correlational
Key metric
73.5% on OGBN-ARXIV; 73.56% on TAPe-ARXIV23 (zero-shot, gpt-3.5-turbo)
Caveat
The paper acknowledges that GPT-3.5's training data may include some arXiv papers, making it infeasible to definitively rule out label leakage on OGBN-ARXIV; TAPe-ARXIV23 was constructed specifically to address this.
Model
GPT-3.5 / ChatGPT-3.5
Datasets
OGBN-ARXIV [eval]
Related findings
IC-997, IC-998
Extraction
automatic-extraction