LLMs fail to capture how humans organize imagined experiences, new study shows

Researchers examining how humans and AI systems mentally organize imagined scenes have uncovered a significant gap: the network structures underlying human imagination remain stubbornly absent in large language models, even at massive scale.

The experiment#

The team analyzed vividness ratings from two established psychological questionnaires—the Vividness of Visual Imagery Questionnaire (VVIQ-2) and the Plymouth Sensory Imagery Questionnaire (PSIQ)—using data from geographically and linguistically diverse human samples totaling 2,743 participants from Florida, Poland, and London. They then compared these networks to imagination patterns generated by six LLMs across various sizes, including Gemma3-12B/27B, Llama3.3-70B, and Llama4-16x17B.

Using regularized partial correlation graphs, the researchers mapped the relationships between imagined scenarios and identified each node's centrality (its importance in the network) and clustering patterns.

Human networks show striking consistency#

Across all three geographic populations, human imagination networks displayed robust structural alignment. Measures like expected influence, strength, and closeness correlated strongly between groups (r = 0.31–0.93). Community detection algorithms successfully recovered clusters that matched both the scene contexts embedded in VVIQ-2 and the distinct sensory modalities measured by PSIQ, with alignment scores ranging from 0.27 to 1.0.

The one exception was betweenness centrality, which varied substantially across populations—a finding the authors attribute to its sensitivity to individual life experience and memory accumulation.

LLMs fall flat#

Despite their impressive language capabilities, the tested models failed to produce human-like imagination networks. LLM-generated centrality patterns showed weak correlations with human networks and were mostly statistically insignificant after correction for multiple comparisons. Most troublingly, nearly all LLM configurations collapsed into single-cluster topologies devoid of meaningful internal structure (median alignment index = 0).

This failure held across different architectures, parameter scales (from 12 billion to 272 billion parameters), and whether the models were given conversational context or not.

Why it matters#

The researchers propose that human imagination networks reflect —how sensory encounters and real-world interactions structure what and how we can vividly imagine. Language training alone, the authors argue, cannot build this kind of experiential scaffold, regardless of how many parameters a model contains or how it is prompted.