GD4LLM: How Layout Quality and Prompting Influence LLM Understanding of Graph Drawings

Area of Science:

• Artificial Intelligence
• Computer Vision
• Graph Theory

Background:

• Large Language Models (LLMs) are increasingly used for complex tasks.
• Graph analysis often requires interpreting visual representations.
• The effectiveness of LLMs with visual graph data is an emerging research area.

Purpose of the Study:

• To investigate how visual input quality affects LLM performance on graph-related tasks.
• To determine the impact of graph layout, drawing aesthetics, and prompting techniques.
• To identify key factors for optimizing LLM performance in visual graph analysis.

Main Methods:

• Experimental analysis of LLM performance using visual graph inputs.
• Varying graph layout paradigms and drawing aesthetics.
• Testing different prompting techniques for graph-related queries.

Main Results:

• Graph layout significantly impacts LLM performance.
• Optimizing drawing readability enhances model accuracy.
• Prompting technique selection is critical for optimal results.

Conclusions:

• Visual input quality is paramount for LLM graph analysis.
• Human-centric design of graph visualizations improves AI performance.
• Effective prompting strategies are essential for maximizing LLM utility in visual graph tasks.