Evaluating intent-driven color generation: a framework for assessing LLM-assisted cartographic design

Large language models (LLMs) offer new possibilities for adaptive cartographic design, yet how they interpret user intent remains unclear. This study evaluates how stated analytical goals influence LLM-generated color schemes for thematic maps, assessing three perceptual dimensions: whether colors form a clear sequence matching the data hierarchy (order), whether categories are visually distinct from one another (differentiation), and whether related values share similar or thematically cohesive colors (association). Using a synthetic dataset of 20 lakes classified by pollution level (Low, Medium, High), we tested three LLMs across 12 prompt conditions varying in task framing and analytical emphasis. GPT-5 demonstrated systematic intent sensitivity: order prompts produced cohesive color schemes, differentiation prompts generated multi-hue palettes with greater separation, and lightness encoding reversed depending on whether users emphasized high or low pollution. DeepSeekV3.2 produced perceptually valid schemes but showed no adaptation to user intent, generating fixed red-yellow-green palettes regardless of prompting. Llama3.2 did not produce complete outputs. These findings suggest intent-sensitive color generation requires reasoning capabilities not uniformly present across models. We propose an evaluation framework linking user intent to measurable color properties and discuss implications for exploratory mapping systems where analytical goals shift dynamically.