loganbolton/sketchvlm-maze-navigation

SketchVLM: Maze Navigation

This dataset is associated with the paper: SketchVLM: Vision Language Models Can Annotate Images to Explain Thoughts and Guide Users.

SketchVLM is a training-free, model-agnostic framework that enables Vision-Language Models (VLMs) to produce non-destructive, editable SVG overlays on input images to visually explain their answers. The Maze Navigation dataset is one of the benchmarks introduced to evaluate a model's ability to trace a path from start to end through a maze.

• Project Page: https://sketchvlm.github.io/
• Repository: https://github.com/Brandon-Collins7/sketchvlm
• Interactive Demo: https://sketch-vlm-demo.vercel.app/

Sample Usage

Downloading the dataset

You can download the dataset locally using the huggingface-cli:

huggingface-cli download loganbolton/sketchvlm-maze-navigation --repo-type dataset --local-dir datasets/maze

Citation

@misc{collins2026sketchvlmvisionlanguagemodels,
      title={SketchVLM: Vision language models can annotate images to explain thoughts and guide users},
      author={Brandon Collins and Logan Bolton and Hung Huy Nguyen and Mohammad Reza Taesiri and Trung Bui and Anh Totti Nguyen},
      year={2026},
      eprint={2604.22875},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.22875},
}