loganbolton/sketchvlm-physics-ball-drop

SketchVLM: Physics Ball Drop Dataset

This dataset is part of the SketchVLM project, introduced in the paper: SketchVLM: Vision language models can annotate images to explain thoughts and guide users.

Project Page | GitHub | Interactive Demo

Description

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 reasoning.

The Physics Ball Drop dataset (based on PHYRE) is one of the benchmarks developed to evaluate visual reasoning. In this task, models must predict the trajectory of a ball through various obstacles and provide a visual explanation of the path.

Usage

As documented in the official GitHub repository, you can download the images and metadata locally using the Hugging Face CLI:

huggingface-cli download loganbolton/sketchvlm-physics-ball-drop --repo-type dataset --local-dir datasets/ball_drop

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},
}