Ciro Santilli @cirosantilli 40

This section is about unofficial ARC-AGI-like problem sets.

These are interesting from both a:

github.com/neoneye/arc-dataset-collection contains a fantastic collection of such datasets, with visualization at: neoneye.github.io/arc/

A subsets of problems that they curate from competitions.

Creator of FrontierMath.

Socials:

The extreme overfitting case of training is to have a map where each input leads to one output.

However it is cool that this overfit does not allow you to compute the final input for which there is no known output.

This therefore forces the creation of more general solution rules.

While in some cases solutions can work for any input, in many others they require specific assumptions about input, but the model could simply check that the assumptions apply to all inputs and use them for the final algorithm.

Bibliography:

Some mentions at: arcprize.org/blog/arc-prize-2025-results-analysis section "Zero-Pretraining Deep Learning Methods".

www.kaggle.com/code/allegich/eda-statistical-analysis-and-feature-extraction has a very basic feature extraction.

AI code generation benchmark in which part of the benchmark includes producing a formal Lean proof of the implementation. Sweet.