LLM Agents Compared for Scientific Visualization Tasks

A recent investigation published on arXiv (2604.27996) assesses eight large language model agents through three distinct interaction paradigms aimed at scientific visualization tasks. These paradigms encompass domain-specific agents utilizing structured tools, computer-use agents, and general-purpose coding agents. The study involved testing these agents on 15 benchmark tasks, focusing on aspects such as visualization quality, efficiency, robustness, and computational expenses. Additionally, the research examined various interaction methods, including code scripts, model context protocol (MCP), API calls, command-line interfaces (CLI), and graphical user interfaces (GUI), along with the influence of persistent memory. Findings indicate notable tradeoffs among the paradigms and modalities, with general-purpose coding agents excelling in specific metrics.