Graph Visualizations in Prompts for Knowledge Graph Question Answering
Summary
- Researchers studied how multimodal input improves knowledge graph question answering (KGQA) in large vision-language models using 1,500 WebQSP and CWQ questions.
- Adding a graph visualization to text prompts yielded an average gain of 1.85 F1 points, with a +4.65 F1 precision improvement on low-complexity questions.
- Multimodal gains did not depend strongly on model scale, but were highly sensitive to graph layout and design choices.
- Hierarchical Graphviz layouts and decomposing large subgraphs into separated paths yielded the largest performance improvements.
- The authors recommend adaptive visual prompting policies to selectively append diagrams to small, visually sparse subgraphs with structured layouts.
Why it matters
Understanding how visual prompt design influences vision-language model performance helps AI practitioners optimize multimodal pipelines, balancing accuracy and token overhead for knowledge graph question answering.