Beyond the Golden Teacher: Enhancing Graph Learning through LLM-GNN Co-teaching

In three linesLLM-GNN Co-Teaching introduces a bidirectional co-teaching framework for few-shot learning on text-attributed graphs. Instead of designating one model as teacher, GNN and LLM exchange their most confident pseudo-labels and update mutually. RPL-PO mines DPO preference pairs from convergence trajectories. Achieves 7.86% 3-shot gains on Cora and 7.73% on ogbn-arxiv.

Summary generated by Claude — human-verified