Charts-of-Thought: Enhancing LLM Visualization Literacy (PDF, 2026) (www3.cs.stonybrook.edu)

A recent study has introduced an innovative prompting technique named "Charts-of-Thought" aimed at enhancing the visualization literacy of Large Language Models (LLMs). Researchers evaluated three state-of-the-art LLMs—Claude-3.7-sonnet, GPT-4.5-preview, and Gemini-2.0-pro—using a standardized test called the Visualization Literacy Assessment Test (VLAT). The Charts-of-Thought method guides these models through a structured process of data extraction, verification, and analysis before responding to visualization-related inquiries. Results showed significant performance improvements, with Claude-3.7 scoring 50.17, far surpassing the human baseline of 28.82, while also showing substantial increases for the other models. This advancement holds critical significance for the AI/ML community as it indicates that, with proper analytical frameworks, LLMs can exceed human performance in tasks involving visual interpretation. The structured prompting approach not only improved accuracy but also highlights the necessity for systematic methodologies when training LLMs for complex analytical tasks, pushing the boundaries of their application in data visualization contexts. Furthermore, the study suggests that such techniques could foster greater accessibility to visual information, benefiting individuals with visual impairments or lower visualization literacy, thereby broadening the impact of AI technologies in diverse domains.