LLM world models are mental: Output layer evidence of brittle world model use in LLM mechanical reasoning

arXiv 25.7 2025 42.8 theory

TLDR

LLMs use heuristics like pulley counting rather than full world models for mechanical reasoning, showing brittle world model use.

Reasoning

Strengths include systematic cognitive science-inspired experiments and clear evidence of heuristic use. Weaknesses are limited domain (pulley systems) and mixed results (chance-level performance in Study 3).

Read-first score

Read-first score 42.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 22.

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=result

Topical relevance 42%

31.4

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

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 334.

Keyword Scores

world model

world simulator

world dynamics prediction

generative world model

interactive world model

model-based reinforcement learning world model

video world model

Deep Analysis

Innovations

Adapting cognitive science methodologies from human mental models research to evaluate LLMs' world-modeling capacities
Using TikZ-rendered pulley system stimuli to test mechanical reasoning in LLMs
Three-study design probing different aspects of world model use: heuristic counting, spatial representation, and structural connectivity reasoning

Methodology

The authors adapt cognitive science methods from human mental models research to test LLMs on pulley system problems using TikZ-rendered stimuli. Three studies examine mechanical advantage estimation, differentiation of functional vs. fake systems, and comparison of functional vs. connected but non-force-transferring systems. Models are evaluated on accuracy, correlation with ground truth, and F1 scores.

Key Results

Study 1: State-of-the-art LLMs performed marginally but significantly above chance in estimating mechanical advantage, with estimates correlating with ground-truth MA and number of pulleys, suggesting a pulley counting heuristic. Study 2: Models identified a functional pulley system over a fake system with F1=0.8. Study 3: Models failed to distinguish functional systems from connected but non-force-transferring systems, achieving F1=0.46 (random guessing).

Limitations

LLMs may lack the facility to reason over nuanced structural connectivity, as shown in Study 3 where performance dropped to random guessing
The findings are compatible with the notion that LLMs use statistical associations (e.g., pulley count) rather than full internal simulation of mechanical systems
Generalization of results is uncertain; the paper states 'insofar as they may generalize'

Resources

Paper