Haifeng Zhang

LLMs can actively disentangle signal from noise in text-attributed graphs, leading to state-of-the-art performance in graph contrastive learning by shifting from random perturbation to semantic-aware disentanglement.