albertoRodriguez97/MazeBench

---
license: mit
task_categories:
- visual-question-answering
- image-classification
language:
- en
tags:
- maze
- spatial-reasoning
- multimodal
- vision-language
- benchmark
size_categories:
- n<1K
---

# MazeBench

The evaluation set from the paper *From Pixels to BFS: High Maze Accuracy Does Not Imply Visual Planning*.

**Paper:** [arXiv:2603.26839](https://arxiv.org/abs/2603.26839)

**Code:** [github.com/alrod97/LLMs_mazes](https://github.com/alrod97/LLMs_mazes)

## Overview

110 procedurally generated maze images spanning 8 structural families and grid sizes from 5x5 to 20x20, with ground-truth shortest-path annotations. These are the mazes used to evaluate multimodal LLMs in the paper.

## Maze Families

- **sanity** — open corridors, minimal branching
- **bottleneck** — narrow chokepoints
- **dead_ends** — many dead-end branches
- **snake** — long winding corridors
- **symmetric** — mirror-symmetric layouts
- **trap** — trap tiles that must be avoided
- **near_goal_blocked** — goal is close but path is indirect
- **near_start_blocked** — start area has limited exits

## Contents

- `gen_maze_001.png` ... `gen_maze_110.png` — 1024x1024 maze images
- `maze_annotations.json` — ground-truth annotations (reachability, shortest path length, accepted shortest paths)

## Usage

```bash
pip install huggingface_hub
huggingface-cli download albertoRodriguez97/MazeBench --repo-type dataset --local-dir generated_mazes/
```

## Citation

```bibtex
@article{rodriguezsalgado2026mazebench,
title = {From Pixels to BFS: High Maze Accuracy Does Not Imply Visual Planning},
author = {Rodriguez Salgado, Alberto},
journal = {arXiv preprint arXiv:2603.26839},
year = {2026},
}
```