Do LLMs Build Spatial World Models? Evidence from Grid-World Maze Tasks

TL;DR

This paper evaluates whether large language models can build internal spatial world models using maze tasks, revealing their limitations in spatial reasoning and the influence of representation formats.

Contribution

It provides systematic evidence that LLMs' spatial reasoning is representation-dependent and not indicative of true internal spatial world models.

Findings

Gemini-2.5-Flash achieves 80-86% accuracy on small mazes with tokenized adjacency.

Performance drops to 16-34% with visual grid formats, showing format-dependent reasoning.

Models fail to leverage semantic understanding for consistent spatial computations.

Abstract

Foundation models have shown remarkable performance across diverse tasks, yet their ability to construct internal spatial world models for reasoning and planning remains unclear. We systematically evaluate the spatial understanding of large language models through maze tasks, a controlled testing context requiring multi-step planning and spatial abstraction. Across comprehensive experiments with Gemini-2.5-Flash, GPT-5-mini, Claude-Haiku-4.5, and DeepSeek-Chat, we uncover significant discrepancies in spatial reasoning that challenge assumptions about LLM planning capabilities. Using chain-of-thought prompting, Gemini achieves 80-86% accuracy on smaller mazes (5x5 to 7x7 grids) with tokenized adjacency representations, but performance collapses to 16-34% with visual grid formats, which is a 2-5x difference, suggesting representation-dependent rather than format-invariant spatial reasoning. We further probe spatial understanding through sequential proximity questions and compositional distance comparisons. Despite achieving 96-99% semantic coverage in reasoning traces, models fail to leverage this understanding for consistent spatial computations, indicating that they treat each question independently rather than building cumulative spatial knowledge. Our findings based on the maze-solving tasks suggest that LLMs do not develop robust spatial world models, but rather exhibit representation-specific and prompting-dependent reasoning that succeeds only under narrow conditions. These results have critical implications for deploying foundation models in applications requiring spatial abstraction.