Internal World Models as Imagination Networks in Cognitive Agents

arXiv 25.10 2025 36.8 theory

TLDR

Compares human and LLM imagination networks via vividness ratings, finding humans have consistent structure while LLMs lack clustering.

Reasoning

Strengths include large-scale human data across populations and multiple LLM variants, with a novel framework for evaluating offline world models. Weaknesses are reliance on self-reported vividness ratings and limited behavioral validation, plus LLMs may not be directly comparable to biological imagination.

Read-first score

Read-first score 36.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 12.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=analysis,metric

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Topical relevance 42%

17.1

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 225.

Keyword Scores

world model

generative world model

world simulator

interactive world model

video world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Using psychological network analysis to compare internal world models (IWMs) between humans and large language models (LLMs) via imagination vividness ratings
Distinguishing offline world models (persistent memory structures) from online models (task-specific representations) in cognitive agents
Providing quantitative metrics for evaluating offline world models in cognitive agents

Methodology

The study employs psychological network analysis on imagination vividness ratings collected from 2,743 humans across three populations and six LLM variants. It compares network structures (centrality, clustering) between human and LLM networks to assess alignment of internal world models.

Key Results

Human imagination networks exhibit robust structural consistency with high centrality correlations and aligned clustering, while LLMs show minimal clustering and weak correlations with human networks, even when conversational memory is included, across environmental and sensory contexts.

Resources

Paper