Ongoing attempt at the ARC-AGI benchmark — the François Chollet-authored test suite that requires solving novel visual reasoning puzzles from a handful of training examples each, designed to be genuinely resistant to the standard deep-learning playbook.
Iterating on a staged pipeline approach: symmetry and transformation inference across training pairs, multi-phase candidate generation (symbolic + program-space + learned priors), and held-out scoring via IoU and exact-match. Currently at pipeline iteration 11. Writeup: ARC Prize 2025 project page.