Addressing information bottlenecks in graph augmented large language models via graph neural summarization

This study investigates the problem of information bottlenecks in graph-level prompting, where compressing all node embeddings into a single vector leads to significant structural information loss. We clarify and systematically analyze this challenge, and propose the Graph Neural Summarizer (GNS), a continuous prompting framework that generates multiple query-aware prompt vectors to better preserve graph structure and improve context relevance. Experiments on ExplaGraphs, SceneGraphs, and WebQSP show that GNS consistently improves performance over strong graph-level prompting baselines. These findings emphasize the importance of addressing information bottlenecks when integrating graph-structured data with large language models. Implementation details and source code are publicly available at https://github.com/timothy-coshin/GraphNeuralSummarizer.