1
LLM Features Can Hurt GNNs: Concatenation Interference on Homophilous Graph Benchmarks
研究发现LLM特征直接拼接会干扰GNN性能，颠覆了“特征越多越准”的普遍认知，值得关注。
arXiv:2606.17579v1 Announce Type: cross Abstract: Adding LLM-generated node features to graph neural networks (GNNs) is widely reported to improve acc…