Answers are grounded in this paper and cite their sources.
Large Language Models (LLMs) have demonstrated impressive capabilities in various applications, motivating visualization researchers to explore the usage of LLMs for visualization tasks such as automated visualization recommendation, code generation and misleading visualization detection. However, it remains unclear how well will LLMs perform for graph layout, a classic and fundamental research question in visualization. To fill this gap, this paper presents a systematic evaluation of three state-of-the-art LLMs (i.e., GPT-4o, Gemini 2.0, DeepSeek-V3) on three key dimensions of graph layout: graph data understanding, layout generation, and layout evaluation. Our experiments cover five representative types of graphs, two graph scales, and two widely used graph representation formats. Our results provide insightful findings on the capabilities of LLMs for each key dimension of graph layout tasks. First, LLMs exhibit strong performance in fundamental graph understanding tasks when code generation is permitted, but their structural reasoning ability declines significantly in pure text-based scenarios. Second, LLMs have the potential to produce promising layouts, though they occasionally generate poor results. Third, visual input generally enhances their ability to evaluate layout quality, while text-only prompts may result in unreliable assessments of graph layout quality. These findings provide valuable insights for advancing future research on leveraging LLMs for graph layout. • We propose a systematic evaluation framework for examining LLMs’ capabilities in graph layout tasks, which covers graph data understanding, layout generation, and layout evaluation. • We conduct a large-scale comparative study across different graph types, sizes, formats, prompting modes, and layout constraints, using three mainstream LLMs. • We provide an in-depth analysis of LLMs’ strengths and limitations for graph layout tasks, and report the major findings in terms of their potential and capability boundaries for graph layouts.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: