MazeBench

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

Paper: arXiv:2603.26839

Code: 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

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

Citation

@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},
}