Skip to content
lkaesbergPublic

About

SPaRC, EMNLP 2025: 1,000 grid pathfinding puzzles that test spatial reasoning in LLMs. pip install sparc-puzzle to evaluate any model with the CLI or Inspect AI. Humans solve 98.0% of the test puzzles, the best model 83.6%.

Topics

Resources

Stars

6 stars

Watchers

1 watching

Forks

Latest commit

Β 

History

101 Commits

Folders and files

Repository files navigation

SPaRC Logo

SPaRC: A Spatial Pathfinding Reasoning Challenge

A comprehensive toolkit for spatial reasoning puzzle solving and model evaluation

Overview

SPaRC provides a comprehensive framework for evaluating language models on spatial reasoning tasks inspired by "The Witness" puzzle game. This package includes tools for dataset processing, solution validation, and model evaluation with beautiful terminal output.

Installation

Install the package from PyPI:

pip install sparc-puzzle

Or install from source:

git clone https://github.com/lkaesberg/SPaRC.git
cd SPaRC
pip install -e .

Quick Start

1. Testing a Model on the Dataset

Run the complete benchmark on your model:

sparc --api-key "your-openai-api-key" --model "gpt-4" --batch-size 5

Key Features:

  • πŸ”„ Resume Support: Automatically saves progress and resumes from where you left off
  • ⚑ Batching: Process multiple puzzles concurrently for faster evaluation
  • 🎨 Rich Output: Beautiful terminal interface with progress tracking
  • πŸ›‘ Graceful Shutdown: Press Ctrl+C to stop after current batch

Example with different endpoints:

# OpenAI API
sparc --api-key "sk-..." --model "gpt-4"

# Custom endpoint (e.g., local model)
sparc --api-key "your-key" --base-url "http://localhost:8080/v1" --model "llama-3.3-70b"

# Resume interrupted session
sparc --api-key "your-key" --model "gpt-4"  # Automatically resumes

# Fresh start (ignore previous results)
sparc --api-key "your-key" --model "gpt-4" --overwrite

2. Using the Validation API

Use SPaRC's validation functions in your own code:

from sparc.validation import extract_solution_path, validate_solution, analyze_path
from sparc.prompt import generate_prompt
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("lkaesberg/SPaRC", "all", split="test")
puzzle = dataset[0]

# Generate prompt for your model
puzzle_prompt = [
                  {
                    "role": "system",
                    "content": "You are an expert at solving puzzles games.",
                  },
                  {
                    "role": "user", 
                    "content": generate_prompt(puzzle)
                  }
                ]


# Your model generates a response
model_response = "... model response with path coordinates ..."

# Extract the path from model response
extracted_path = extract_solution_path(model_response, puzzle)
# Returns: [{"x": 0, "y": 2}, {"x": 0, "y": 1}, ...]

# Validate against ground truth
is_correct = validate_solution(extracted_path, puzzle)
# Returns: True/False

# Get detailed analysis
analysis = analyze_path(extracted_path, puzzle)
# Returns: {
#   "starts_at_start_ends_at_exit": True,
#   "connected_line": True,
#   "non_intersecting_line": True,
#   "no_rule_crossing": True,
#   "fully_valid_path": True
# }

3. Running SPaRC with Inspect AI

SPaRC is also an Inspect AI task. Inspect is an optional extra, so install it with the package. From a checkout, run pip install -e ".[inspect]", then:

inspect eval sparc --model openai/gpt-4.1-mini            # all 500 test puzzles
inspect eval sparc_hard --model openai/gpt-4.1-mini       # the 89 hard (level 5) puzzles

With a normal (non-editable) install, for example pip install "sparc-puzzle[inspect]", the task names carry the package prefix: inspect eval sparc/sparc and inspect eval sparc/sparc_hard.

Since task version 2-A, the prompt states the star, polyshape and negative polyshape rules the way the puzzles were built. Add -T prompt=paper for the paper's original wording. See sparc/README.md for options, the changelog, scoring and reproduced numbers.

Since version 0.7.0, the CLI's step-by-step and visual prompts use the same corrected rule wording. Each keeps its original wording under a _paper name, for example --prompt gym_step_paper or --prompt single_shot_visual_paper.

CLI Reference

Basic Usage

sparc --api-key "your-key" [OPTIONS]

Options

Option Default Description
--api-key Required OpenAI API key or your model's API key
--base-url https://api.openai.com/v1 API endpoint URL
--model gpt-4 Model name to evaluate
--temperature 1.0 Generation temperature
--batch-size 5 Number of concurrent requests
--results-file <model>.jsonl File to save results
--overwrite False Ignore existing results and start over
--verbose False Show detailed output for each puzzle
--max-new None Process at most this many new puzzles
--gym False Use step-by-step gym mode instead of single-shot
--gym-traceback False Enable traceback visualization in gym mode
--run-name None Suffix for output filename (e.g., experiment1)

Examples

# Basic evaluation
sparc --api-key "sk-..." --model "gpt-4"

# High throughput with larger batches
sparc --api-key "sk-..." --model "gpt-3.5-turbo" --batch-size 20

# Conservative approach with lower temperature
sparc --api-key "sk-..." --model "gpt-4" --temperature 0.1

# Verbose output to see each puzzle result
sparc --api-key "sk-..." --model "gpt-4" --verbose

# Custom results file
sparc --api-key "sk-..." --model "claude-3" --results-file "claude_results.json"

# Process only 10 new puzzles
sparc --api-key "sk-..." --model "gpt-4" --max-new 10

# Step-by-step gym mode (agent receives feedback after each move)
sparc --api-key "sk-..." --model "gpt-4" --gym

# Gym mode with traceback (shows path history in observations)
sparc --api-key "sk-..." --model "gpt-4" --gym --gym-traceback

# Named experiment run
sparc --api-key "sk-..." --model "gpt-4" --run-name "experiment1"

Core Functions

extract_solution_path(solution_text: str, puzzle_data: Dict) -> List[Dict[str, int]]

Extracts coordinate path from model response text.

validate_solution(extracted_path: List[Dict[str, int]], puzzle_data: Dict) -> bool

Validates if the extracted path matches any ground truth solution.

analyze_path(solution_path: List[Dict[str, int]], puzzle: Dict) -> Dict

Provides detailed analysis of path validity and rule compliance.

generate_prompt(puzzle_data: Dict) -> str

Generates the formatted prompt for a puzzle.

Contributing

Contributions are welcome! Please feel free to submit pull requests or open issues.

Citation

If you use SPaRC in your research, please cite:

@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"
}

If you use the step-by-step environment, please also cite Spatial-Gym:

@misc{kaesberg2026spatialgym,
  title  = {Mind the Gap Between Spatial Reasoning and Acting! Step-by-Step Evaluation of Agents With Spatial-Gym},
  author = {Kaesberg, Lars Benedikt and Yang, Tianyu and Bauer, Niklas and Ruas, Terry and Wahle, Jan Philip and Gipp, Bela},
  year   = {2026},
  eprint = {2604.09338},
  archivePrefix = {arXiv}
}

If you use the visual scaffolds, please also cite:

@misc{kaesberg2026visualprompting,
  title  = {Is Visual Prompting All You Need? Studying VLM Spatial Reasoning under Progressive Visual Scaffolds},
  author = {Kaesberg, Lars Benedikt and Yang, Tianyu and Wunderlich, Florian Valentin and Ruas, Terry and Kurzawe, Daniel and Wahle, Jan Philip and Gipp, Bela},
  year   = {2026},
  eprint = {2608.21170},
  archivePrefix = {arXiv},
  note   = {Accepted at Findings of EMNLP 2026}
}

Related work from our group

Paper Venue Links
SPaRC: A Spatial Pathfinding Reasoning Challenge EMNLP 2025 (Main) ACL Anthology Β· arXiv:2505.16686 Β· Code Β· Dataset
Mind the Gap Between Spatial Reasoning and Acting! Step-by-Step Evaluation of Agents With Spatial-Gym arXiv preprint, 2026 arXiv:2604.09338 Β· Code
Is Visual Prompting All You Need? Studying VLM Spatial Reasoning under Progressive Visual Scaffolds Findings of EMNLP 2026 arXiv:2608.21170
Voting or Consensus? Decision-Making in Multi-Agent Debate Findings of ACL 2025 ACL Anthology Β· arXiv:2502.19130

We also wrote two plain-language articles on Towards Data Science. Spatial Reasoning in AI Models, Taken Apart walks through all three papers, and Fix the Image, Not the Prompt explains why changing the image helps vision models more than rewording the prompt.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Links

About

SPaRC, EMNLP 2025: 1,000 grid pathfinding puzzles that test spatial reasoning in LLMs. pip install sparc-puzzle to evaluate any model with the CLI or Inspect AI. Humans solve 98.0% of the test puzzles, the best model 83.6%.

Topics

Resources

Stars

6 stars

Watchers

1 watching

Forks

Releases

Packages

Used by

Contributors

Languages