Exploring MLLMs Perception of Network Visualization Principles

Area of Science:

• Artificial Intelligence
• Computer Vision
• Network Science

Background:

• Assessing the quality of network layouts is crucial for understanding complex systems.
• Human perception of network properties, like stress, is well-documented but computationally challenging to replicate.
• Multimodal Large Language Models (MLLMs) offer new possibilities for analyzing visual data.

Purpose of the Study:

• To evaluate if MLLMs can match human performance in perceiving properties of network layouts.
• To compare MLLM performance against both human experts and non-experts.
• To investigate the perceptual mechanisms employed by MLLMs in network analysis.

Main Methods:

• Replication of a human-subject experiment on network layout quality perception.
• Utilizing advanced MLLMs such as GPT-4o, Gemini-2.5, and Qwen2.5.
• Controlled experimental setup providing identical information to MLLMs and human participants.

Main Results:

• MLLMs achieved performance comparable to trained human experts.
• MLLM performance surpassed that of untrained human non-experts.
• Prompt engineering modifications led to superior-than-human performance in specific scenarios.
• MLLMs appeared to use visual proxies, similar to humans, rather than direct stress value computation.

Conclusions:

• MLLMs demonstrate a significant capacity for perceiving network layout properties, mirroring human visual perception.
• The findings suggest MLLMs can serve as valuable tools for network analysis and quality assessment.
• Further research into MLLM perceptual mechanisms and prompt engineering is warranted for enhanced AI capabilities.