GraphGPT: A Self Supervised-Learning for Intuitive, Logical, and Visual Education
Naile Wang · 2025
In this paper, we introduce GraphGPT, a novel machine-learning approach designed for intuitive, logical, and visual education. GraphGPT is inspired by the observation that scientific graphs help learners visualize concepts, while causal graphs enhance the logic of responses. By combining the masked prediction capabilities of GPT with transform prediction using causal graphs, we can improve explanation accuracy. GraphGPT generates logical steps alongside diagrams, significantly enhancing intuition and visual presentation compared to ChatGPT. Experimental results show that GraphGPT outperforms ChatGPT by 25% in accuracy for multiple-choice questions. Furthermore, a vast majority of participants in the subjective tests confirmed that GraphGPT surpasses ChatGPT in all four evaluation categories. Many highlighted the diagrams and graphical explanations as key factors, giving GraphGPT a notable intuitive advantage. Our work is available at https: //github.com/nicholasxwang/GraphGPT.