GLOW pairs a pre-trained GNN with an LLM to answer questions over incomplete knowledge graphs and ships GLOW-BENCH, a 1,000-question evaluation.