This repository provides the implementation for Poetiq's state-of-the-art (SOTA) reasoning system, enabling reproduction of its record-breaking submissions to the ARC-AGI-1 and ARC-AGI-2 benchmarks. It targets AI researchers and engineers focused on advanced reasoning capabilities and benchmark performance, offering a direct path to understanding and leveraging top-tier solutions for complex abstract reasoning tasks.

How It Works

The project facilitates replication of Poetiq's approach to achieving SOTA performance on abstract reasoning challenges. While specific algorithmic details are elaborated in linked blog posts, the core functionality revolves around a sophisticated reasoning engine designed to tackle the ARC-AGI benchmark's demanding problem sets. This approach has proven effective in surpassing existing benchmarks and establishing new performance standards.

Quick Start & Requirements

• Primary install: Set up a Python 3.11+ virtual environment, activate it, and run pip install -r requirements.txt.
• Prerequisites: API keys for models such as Gemini and OpenAI are mandatory. A .env file in the root directory must contain these keys.
• Execution: Modify constants in main.py to configure the problem set and run python main.py.
• Links:
  • Launch Post: Traversing the Frontier of Superintelligence
  • Follow-up Post: Poetiq Shatters ARC-AGI-2 State of the Art at Half the Cost

Highlighted Details

• Enables reproduction of record-breaking submissions to the ARC-AGI-1 and ARC-AGI-2 benchmarks.
• Achieves state-of-the-art (SOTA) reasoning performance on official leaderboards.
• Provides verifiable results for public and private evaluation datasets.

Maintenance & Community

For inquiries or discussions regarding AI reasoning, contact poetiq@poetiq.ai. Users are encouraged to cite the project's launch post for academic or research use.

Licensing & Compatibility

The provided README does not specify a software license. This absence necessitates caution regarding usage, modification, and distribution, particularly for commercial applications or integration into closed-source projects.

Limitations & Caveats

The solver's functionality is dependent on external proprietary AI models (e.g., Gemini, OpenAI), requiring valid API keys and incurring associated costs. The primary focus is on reproducing benchmark results, and general-purpose reasoning capabilities beyond the ARC-AGI scope are not detailed.