LLM Features Can Degrade GNN Performance on Homophilous Graphs.

It is widely assumed that incorporating features generated by large language models (LLMs) into Graph Neural Networks (GNNs) generally improves accuracy on standard benchmarks. However, new research presents a contrasting observation: when LLM features are simply concatenated to the input (without joint training, distillation, or prompt conditioning), they can actually *reduce* GNN accuracy, particularly on homophilous graph benchmarks. The study found significant accuracy drops on datasets like PubMed (-17.0 pp) and Cora (-4.3 pp) when using SBERT-encoded GPT-4o-mini TAPE features with an MLP backbone. This "concatenation interference" effect diminishes with different GNN backbones or random data splits and can even reverse on medium-homophily datasets like WikiCS and ogbn-arxiv. To predict when this degradation occurs, the researchers propose a simple measure called Delta_sig, which quantifies the LLM's standalone discriminability. This measure correlates more strongly with the concatenation cost than graph homophily itself. The findings suggest that simply adding LLM features via concatenation is not always beneficial and can introduce interference, especially when the LLM features are not sufficiently discriminative on their own for the specific task.