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

arXiv: https://arxiv.org/abs/2604.10690 · HTML: https://mlsanity.com/paper/2604.10690/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in World models for language agents, one of 11 directions in World models.
- Builds on 17 of its 25 references that are in the map.
- First posted on arXiv 12 Apr 2026.

## Placed in

- World models / [World models for language agents](https://mlsanity.com/wm/world-models-for-language-agents/)
- LLM reasoning / [Spatial and embodied reasoning](https://mlsanity.com/lr/spatial-and-embodied-reasoning/)

## Builds on (most cited first, 17 of 17)

- [Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks](https://mlsanity.com/paper/1908.10084/) (2019)
- [A Survey on LLM-as-a-Judge](https://mlsanity.com/paper/2411.15594/) (2024)
- [Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task](https://mlsanity.com/paper/2210.13382/) (2022)
- [ROSCOE: A Suite of Metrics for Scoring Step-by-Step Reasoning](https://mlsanity.com/paper/2212.07919/) (2022)
- [Beyond A*: Better Planning with Transformers via Search Dynamics Bootstrapping](https://mlsanity.com/paper/2402.14083/) (2024)
- [GraphArena: Evaluating and Exploring Large Language Models on Graph Computation](https://mlsanity.com/paper/2407.00379/) (2024)
- [From task structures to world models: What do LLMs know?](https://mlsanity.com/paper/2310.04276/) (2023)
- [Can Graph Learning Improve Planning in LLM-based Agents?](https://mlsanity.com/paper/2405.19119/) (2024)
- [ACPBench: Reasoning about Action, Change, and Planning](https://mlsanity.com/paper/2410.05669/) (2024)
- [AlphaMaze: Enhancing Large Language Models' Spatial Intelligence via GRPO](https://mlsanity.com/paper/2502.14669/) (2025)
- [Evaluating the Robustness of Analogical Reasoning in Large Language Models](https://mlsanity.com/paper/2411.14215/) (2024)
- [Structured World Representations in Maze-Solving Transformers](https://mlsanity.com/paper/2312.02566/) (2023)
- [Transformers Use Causal World Models in Maze-Solving Tasks](https://mlsanity.com/paper/2412.11867/) (2024)
- [Transformers Can Navigate Mazes With Multi-Step Prediction](https://mlsanity.com/paper/2412.05117/) (2024)
- [From Text to Space: Mapping Abstract Spatial Models in LLMs during a Grid-World Navigation Task](https://mlsanity.com/paper/2502.16690/) (2025)
- [Do AI Models Perform Human-like Abstract Reasoning Across Modalities?](https://mlsanity.com/paper/2510.02125/) (2025)
- [LogiPlan: A Structured Benchmark for Logical Planning and Relational Reasoning in LLMs](https://mlsanity.com/paper/2506.10527/) (2025)

## Built on by (most cited first, 2 of 2)

- [Tool-Call Dependency Structure is Linearly Decodable in LLM Agent Residual Streams](https://mlsanity.com/paper/2605.25310/) (2026)
- [Do LLMs Build World Models From Text? A Multilingual Diagnostic of Spatial Reasoning](https://mlsanity.com/paper/2605.28277/) (2026)

## Abstract (arXiv)

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.