Saurabh Ranjan¹, Brian Odegaard¹; ¹University of Florida

Visual imagination is an aspect of consciousness that results in experiences of the sensory world without any external stimulation. How is the experience of imagination structured in human minds and machines? To explore this question, we probed how the vividness ratings of different imagined experiences were associated with one another in both humans and large language models (LLMs). We analyzed responses from the Vividness of Visual Imagery Questionnaire (VVIQ), which requires participants to imagine different visual aspects of a scene under given context, and report their vividness on a rating scale (1-5). To understand how the vividness of different experiences were associated with one another, we first constructed imagination networks using pairwise partial correlations (edge weights) of the items (nodes) from 1,776 human responses to the VVIQ. Next, we constructed the same networks based on VVIQ responses from LLMs to understand how their generative behavior differs from human experiences of visual imagination. We found that humans exhibited more positive edge associations between the vividness of different items, but LLM responses showed more negative edges. These differences in edges resulted in topological differences in the imagination networks as shown across multiple centrality measures like strength, closeness, betweenness, and expected influence of each node. Further, to understand how vividness of different experiences clustered in the networks, we investigated the community structure of nodes in networks from humans and LLMs. While all eight VVIQ contexts clustered in human imagination networks; clustering of items from LLM networks was extremely diffused. Overall, our study not only reveals how the vividness of different experiences are associated with each other in human visual imagination, but also shows how responses differ in the imagination of LLMs. Together, our results reflect differences in internal world-building across natural and artificial generative processes, resulting in different vividness responses to visual imagination.

Acknowledgements: No Funding Sources Available