Visualization artifacts have long served as anchors for collaboration and knowledge transfer in data analysis. While effective for human-human collaboration, little is known about their role in capturing and externalizing knowledge when working with large language models (LLMs). Despite the growing role of LLMs in analytics, their linear text-based workflows limit the ability to structure artifacts into useful and traceable representations of the analytical process. We argue that dynamic visual representations of evolving analysis-organizing artifacts and provenance into semantic structures, such as idea development and shifts in inquiry-are critical for effective human-LLM workflows. We demonstrate the current opportunities and limitations of using LLMs to track, structure, and visualize analytic processes, and propose a research agenda to leverage rapid advances in LLM capabilities. Our goal is to present a compelling argument for maximizing the role of visualization as a catalyst for more structured, transparent, and insightful human-LLM analytical interactions.
Elshehaly et al. (Mon,) studied this question.