SPaRC Dataset

Paper • Website • Code • Spatial-Gym • Solver • Generator

A grid-based puzzle dataset for benchmarking LLMs spatial reasoning capabilities.

Configs

all is the benchmark: 500 train and 500 test puzzles that mix all rule types. It is not the union of the other configs. The per-rule configs (dots, gaps, stars, stones, triangles, polys, polys_ylops, stones_stars, dots_stars_polys, gaps_dots_triangles) hold separate puzzles with only some rule types. The paper uses them for its per-rule results (Table 8).

The images config holds pictures of all 2,000 puzzles: the 1,000 from all and the 1,000 from the rule configs, with the same train/test split. It has only an id column and one image column per drawing type, so join it with any other config on id. The dataset viewer opens on this config. The text configs have no images, so loading them stays as small as before.

Column	What it shows	plot_type in sparc-visualization
original	The puzzle as the visual SPaRC tasks show it to models	original
start_end_markers	Start and exit marked with S and E	start_end_marked
axis_labels	Row and column numbers along the edges	coordinate_grid
axis_labels_and_start_end	Axis labels plus S and E	coordinate_grid_and_start_end_marked
cell_coordinates	The coordinates of every cell	path_cell_annotated
text_symbols	Cell coordinates, with every symbol written as a word	text
low_contrast	Faded colors	low_contrast
low_contrast_with_cell_coordinates	Faded colors with cell coordinates	low_contrast_and_path_cell_annotated
low_resolution	Scaled down and back up	low_resolution
low_resolution_with_cell_coordinates	Lower resolution with cell coordinates	low_resolution_and_path_cell_annotated
rotated	Board rotated by 15 degrees	rotated
rotated_with_cell_coordinates	Rotated board with cell coordinates	rotated_and_path_cell_annotated
black_frame	A black frame around the board	black_frame
black_frame_with_cell_coordinates	Black frame with cell coordinates	black_frame_and_path_cell_annotated

The first six columns are the drawings from Is Visual Prompting All You Need? (Findings of EMNLP 2026). All images were drawn with sparc-visualization 0.2.2. The JSONL files in the repository root hold the same puzzles as plain text.

Data Schema

Each record (JSON) includes:

• id (string): unique puzzle identifier
• difficulty_level (int) & difficulty_score (float)
• grid_size: { "height": H, "width": W }
• polyshapes: JSON string mapping shape IDs to binary grids
• puzzle_array: 2D array with cell codes (e.g., S, E, +, P-O-112)
• solution_count (int) and solutions list (with index, path, pathLength)
• text_visualization: YAML-style summary

Sample Entry

{
  "id": "10e68dc3a6fbfcdf",
  "difficulty_level": 3,
  "difficulty_score": 2.3261,
  "grid_size": {"height":3,"width":4},
  "polyshapes":"{\"112\":[[0,1,...]]}",
  "puzzle_array":[
    ["+","+","+","E"],
    ["+","P-O-112","+","o-O"],
    ["S","N","+","G"]
  ],
  "solution_count": 10,
  "solutions": [ { "index":0, "pathLength":34, ... } ],
  "text_visualization": "..."
}

Citation Information

If you use the dataset in any way, please cite the following paper. Preprint: https://arxiv.org/abs/2505.16686

@inproceedings{kaesberg-etal-2025-sparc,
  title     = "{SP}a{RC}: A Spatial Pathfinding Reasoning Challenge",
  author    = "Kaesberg, Lars Benedikt and Wahle, Jan Philip and Ruas, Terry and Gipp, Bela",
  booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
  month     = nov,
  year      = "2025",
  address   = "Suzhou, China",
  publisher = "Association for Computational Linguistics",
  url       = "https://aclanthology.org/2025.emnlp-main.526/",
  doi       = "10.18653/v1/2025.emnlp-main.526",
  pages     = "10359--10390"
}

Related Papers

Two follow-up papers use the SPaRC puzzles. BibTeX for both is on the website.

• Mind the Gap Between Spatial Reasoning and Acting! Step-by-Step Evaluation of Agents With Spatial-Gym turns SPaRC into a step-by-step environment for agents.
• Is Visual Prompting All You Need? Studying VLM Spatial Reasoning under Progressive Visual Scaffolds studies vision-language models on SPaRC with visual scaffolds. Findings of EMNLP 2026.

For a plain-language walk-through of all three papers, see Spatial Reasoning in AI Models, Taken Apart on Towards Data Science.