A community-driven showcase of ARC-AGI methods, harnesses, and approaches for ARC-AGI. Open to anyone - submit via Pull Request.
This is a curated showcase of methods, not a ranking. Although self-reported scores are collected with each submission, there's no reliable way to validate them - so we don't display them on the website. Only ARC Prize Verified scores are shown, with a number and badge. We want to highlight the system and the idea, not the score.
This is a showcase of novel, general-purpose, open methods for ARC-AGI (v1, v2, or v3) - not a competition and not a ranking. We highlight the system and the idea, not the score. The goal is to promote interesting, net-new work.
A submission belongs here if it meets all three bars:
- General-purpose. The method should plausibly work on tasks or games it has never seen. Per-task hardcoding, human-derived answer keys, or lookup tables of solutions don't belong here - even with a high score.
- The system is open - not just the output. The code that produces your results must be public and reproducible. Publishing only generated solutions (e.g. a file of per-task Python) without the system that created it is not enough.
- Novel contribution. A new harness, search strategy, training method, or approach - one new idea. Novel wrappers around frontier models are welcome.
Although we collect scores, self-reported scores are untrustworthy by design. We don't run Community Leaderboard code, nor verify scores against the ARC-AGI Semi-Private sets. So:
- We collect your score and cost in
submission.yaml. For ARC-AGI-3 we collect a requiredscorecard_urland derive the score from it instead of taking a self-reported number. - We do not display self-reported or derived scores, or costs, on the website.
- The only numbers shown are ARC Prize Verified scores (with the verified badge). For verified scores, see the official leaderboard.
But an untrustworthy score doesn't make a submission any less interesting. We rely on community verification to surface what's genuinely cool about these methods - we judge the idea and the method, not the number.
Fork the repo, copy submissions/.example/ to submissions/<your-id>/, fill in submission.yaml, and open a Pull Request. A maintainer reviews and merges submissions on a weekly basis.
See CONTRIBUTING.md for the full how-to - required fields, the submission schema, and the review process.