ARC Prize 2026 – ARC-AGI-2 Submission

This repository contains my solutions and experiments for the ARC Prize 2026 competition on Kaggle.

About ARC

The Abstraction and Reasoning Corpus (ARC) is a benchmark to evaluate AI’s ability to reason and generalize to novel, human-solvable tasks. (Epoch AI)

Project Goals

• Learn and implement reasoning-based AI
• Build rule-based and programmatic solvers
• Achieve generalization to unseen ARC tasks

Progress

• ✅ Joined ARC Prize 2026
• ✅ Built and submitted baseline model (zero output)
• ✅ Upgraded to a simple rule-based model (color mapping)
• ✅ Submission pipeline matches required JSON format
• ✅ Automated format check and output verification

Tech Stack

• Python 3
• NumPy

How to Use

1. Run the notebook arc_baseline_zero_model.ipynb.
2. The notebook will generate submission.json in the required format for Kaggle.
3. (Optional) Use the provided evaluation utilities to test accuracy on the validation set.

Next Steps

• Add more advanced rule-based and pattern-based reasoning
• Explore program synthesis and learning-based approaches
• Visualize and analyze model predictions

Insights

ARC is about discovering and applying rules from small examples, not just scaling deep learning. Generalization and reasoning are key. (GitHub)