Psychological Imagination Networks Show Cross-Population Centrality and Clustering Alignment in Humans That Large Language Models Fail to Replicate

What Happened

A new preprint from arXiv (2510.04391v5) applies psychological network analysis to compare how humans and large language models structure mental imagery. The researchers mapped "psychological imagination networks"—relational graphs of how different imagined scenarios connect in terms of vividness, emotional valence, and semantic similarity—across human populations and several LLMs. Their key finding: human imagination networks exhibit stable cross-population centrality (certain scenarios consistently anchor the network) and clustering alignment (scenarios group into coherent thematic clusters). LLMs, despite generating vivid text, fail to replicate these structural properties. The models produce networks where centrality is less stable and clustering patterns diverge from human baselines, suggesting a fundamental mismatch in how synthetic and biological minds organize imagined content.

Why It Matters

This study moves beyond typical benchmarks of LLM performance—factual accuracy, fluency, or reasoning—into the domain of cognitive architecture. Imagination is not merely generating plausible scenarios; it involves a structured relational system where some ideas serve as cognitive hubs, and others cluster by emotional or thematic ties. That LLMs fail to mirror this structure has several implications:

• Limits of surface mimicry: LLMs can produce vivid descriptions of imagined scenes, but the underlying network organization is different. This suggests that current training objectives (next-token prediction, RLHF) do not capture the relational coherence of human imagination.
• Cross-population validity: Human imagination networks are robust across individuals, indicating a shared cognitive grammar. LLMs, trained on diverse internet text, should theoretically absorb this grammar, yet they do not—implying a gap in how statistical patterns in language map to cognitive structures.
• New evaluation dimension: Standard NLP metrics (BLEU, perplexity, human preference) miss this structural dimension. The paper introduces a novel way to assess whether models "think" like humans in open-ended generative tasks.

Implications for AI Practitioners

• Architecture design: If imagination networks are a core cognitive feature, future models may need explicit relational memory or graph-based reasoning modules to capture hub-and-cluster dynamics, rather than relying solely on transformer attention.
• Safety and alignment: Models that organize imagination differently may generate unexpected associations or fail to maintain coherent narrative structures. For applications in therapy, creative writing, or simulation, this misalignment could produce outputs that feel "off" to human users.
• Evaluation pipelines: Practitioners should consider adding network analysis to their testing suites for generative models, especially for tasks requiring sustained imaginative coherence (e.g., story generation, world-building, scenario planning).
• Data and training: The failure suggests that language-only training is insufficient for replicating human imagination structure. Multimodal data (visual, emotional, embodied) or structured cognitive priors may be necessary.

Key Takeaways

• Human imagination networks show stable centrality and clustering that LLMs fail to replicate, revealing a structural gap in how synthetic minds organize imagined content.
• This finding introduces a new evaluation dimension beyond surface-level fluency, focusing on relational cognitive architecture.
• AI practitioners should consider network-based metrics for assessing generative models, especially in creative or therapeutic applications.
• The study implies that current LLM training paradigms may need supplementation with structured relational or cognitive priors to better align with human imagination.