We conduct a controlled study on how multimodal input improves knowledge graph question answering (KGQA) in large vision‑language models. Using 1,500 WebQSP and CWQ questions, six VLMs, and a fixed KGQA pipeline, we compare text‑only prompting with text‑plus‑graph visualization per question, examining the effects of reasoning complexity, model size, and image layout. Adding a graph image yields an average gain of $1.85$ F1, driven by a strong precision boost on low‑complexity questions (+4.65 F1). Mixed‑effects and regression analyses show that these gains are weakly tied to model scale but highly sensitive to layout: image‑only settings degrade performance, hierarchical Graphviz layouts outperform chaotic renderings, and breaking large subgraphs into separate paths provides the largest improvements. Overall, visual prompting mainly aids KGQA by enhancing precision in low‑complexity contexts, provided the visual design respects cognitive and perceptual constraints. To balance accuracy and inference efficiency, we advocate adaptive visual prompting policies that attach diagrams only to small, visually sparse subgraphs with well‑structured layouts, avoiding unnecessary visual clutter and token overhead on complex queries. We conclude by linking these insights to a visual notation design theory, suggesting that principled diagram encoding can guide future multimodal VLM prompting.
Review