PLMGH: What Matters in PLM-GNN Hybrids for Code Classification and Vulnerability Detection
Researchers have explored the effectiveness of combining pretrained language models (PLMs) with graph neural networks (GNNs) for code classification and vulnerability detection. Their study, titled "PLMGH," systematically paired various code-specialized PLMs and GNN architectures. The findings indicate that these hybrid approaches generally outperform GNN-only methods and can enhance the performance of frozen PLMs. Notably, the choice of PLM and its feature source proved more critical than the GNN backbone for task performance and robustness. AI
IMPACT Provides practical guidelines for designing PLM-GNN hybrids, potentially improving code analysis tools.