Key points are not available for this paper at this time.
Large language models (LLMs) offer transformative potential for automated cartography, yet their effectiveness depends on their integration with specialized cartographic knowledge. Although strategies like retrieval-augmented generation (RAG), fine-tuning, and prompt engineering are the prevailing approaches for knowledge augmentation, their comparative effectiveness in enhancing LLM performance in cartography remains underexplored. Based on a specialized cartographic corpus, we compared these strategies, conceptualized as referencing, internalizing, and activating knowledge. We evaluated each strategy on a medium-parameter-scale LLM using a dual framework of automated metrics and expert assessment, validating the resulting trends on two additional models. Our findings show a clear performance hierarchy that shifts with cognitive complexity. For low-complexity recall tasks, RAG demonstrates superior performance, and fine-tuning and prompt engineering also prove effective. As tasks demand multistep reasoning, RAG and fine-tuning perform comparably, whereas prompt engineering falters. For the most demanding tasks involving creative synthesis, fine-tuning emerges as the optimal strategy, with prompt engineering’s performance regressing toward the baseline. We conclude that no single strategy is universally superior. Instead, the optimal approach depends on the task’s demand for factual fidelity, applied reasoning, or creative synthesis. This study provides foundational guidance for selecting appropriate knowledge augmentation strategies for LLM-driven cartography systems.
Wang et al. (Thu,) studied this question.