ARC-AGI-2 Abstraction Dataset

A dataset containing 120 ARC-AGI-2 evaluation tasks with solvers written in CompDSL—a purely functional, simply-typed DSL for grid transformation programs. Each solver is guaranteed to terminate with polynomial complexity bounds and maintains equational reasoning properties.

What is CompDSL?

CompDSL is a purely functional, simply-typed DSL for grid programs embedded in disciplined Python. Every solver is guaranteed to terminate in O(n^d) time with no mutation, recursion, or unbounded loops.

Core idea: Write programs using only pure functions, comprehensions, and the single combinator fold_repaint—enforced via AST validators and type checking.

Dataset Overview

This dataset contains solvers for all 120 ARC-AGI-2 evaluation tasks, generated using an automated approach combining coding agents with abstraction-refinement techniques.

Status: 116 of 120 solvers pass interpolation (correctly solve all training examples). Identity baselines are provided for the remaining 4 tasks (142ca369, 21897d95, 271d71e2, da515329) as placeholders for future improvements.

Each task bundle includes:

• Task Solution (solution.py): Final working Python implementation in CompDSL
• Abstraction Code (abstractions.py): Reusable abstraction functions and pipelines
• Analysis Report (abstractions.md): Detailed abstraction process and DSL specification

For details on the refinement pipeline, see the companion compositional-program-synthesis repository.

Directory Structure

arc-agi-2-abstraction-dataset/
├── tasks/               # Self-contained task bundles
│   ├── 195c6913/
│   │   ├── solution.py
│   │   ├── abstractions.py
│   │   └── abstractions.md
│   ├── 1ae2feb7/
│   │   ├── solution.py
│   │   ├── abstractions.py
│   │   └── abstractions.md
│   └── ...
├── dsl/                 # Typed DSL: docs, registry, validators
├── check_consistency.py # Dataset integrity checker
├── CHANGELOG.md
└── README.md            # This file

CompDSL Specification

The dsl/ directory contains the full specification, validators, and operation registry.

Documentation:

• dsl/README.md — Quick reference with syntax cheat sheet
• dsl/DSL.md — Authoritative specification for contributors
• dsl/DSL_Research_Note.md — Design rationale and theoretical foundations

Validation: Run check_lambda_types.py and validate_dsl.py in the dsl/ directory. See dsl/DSL.md for details.

Task Bundles

Each tasks/<task-id>/ directory contains:

• solution.py (required): Solver function. 116 of 120 pass all training examples; 4 are identity baselines.
• abstractions.py (optional): Reusable helper functions (component analysis, symmetry detection, morphological ops).
• abstractions.md (optional): Report with DSL specification validated by check_lambda_types.py.

Identity baselines omit the optional files.

Common Patterns

Per-item local reasoning: fold_repaint over objects, applying local masks
Global infer → local render: First pass builds structure, second pass paints
Search/evaluation (rare): Generate candidates, score, select best

Domain abstractions: Component analysis, symmetry detection, morphological ops, geometric/color reasoning, neighborhood rules.

Usage

# Run a solver
from tasks.1ae2feb7.solution import solve_1ae2feb7
result = solve_1ae2feb7(input_grid)

# Use abstraction helpers
from tasks.1ae2feb7.abstractions import repeat_last_nonzero_block
result = repeat_last_nonzero_block(grid)

Check repository consistency: python check_consistency.py

Citation

If you use this dataset in your research, please cite:

@dataset{arc_agi_2_abstraction_2025,
  title={ARC-AGI-2 Abstraction Dataset},
  description={Generated solutions and abstractions for ARC-AGI-2 tasks},
  year={2025},
  url={https://github.com/cristianoc/arc-agi-2-abstraction-dataset}
}

License

This dataset is released under the MIT License. See LICENSE for details.

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for detailed guidelines on:

• Adding new tasks and abstractions
• Repository structure and naming conventions
• Running consistency checks
• Code quality guidelines
• Pull request process

Please feel free to submit issues, feature requests, or pull requests following the guidelines in CONTRIBUTING.md.

Acknowledgments

We are grateful to:

• The ARC-AGI-2 challenge organizers for creating and stewarding the benchmark
• François Chollet for encouraging deeper exploration of abstraction-refinement approaches
• Michael Hodel for introducing the author to ARC, sparking formative discussions, and publishing the original ARC solver corpus (arc-dsl)
• Eric Pang for feedback on the abstraction refinement work and sharing DreamCoder-inspired ARC-AGI program synthesis (arc_agi, blog post)
• Peter O'Hearn for ongoing conversations about ARC and reasoning in AI
• The broader ARC research community for continually sharing ideas, analyses, and solver techniques