This research paper explores efficient neurally-guided program induction for solving tasks within the ARC-AGI open-world problem domain. Three paradigms are examined: learning the grid space, learning the program space, and learning the transformation space. The authors thoroughly investigate the first two, finding the program space approach (GridCoder) most effective, though limited by structural generalization issues. A novel probabilistic program enumeration search algorithm is presented, utilizing transformer-based token sequences. Finally, the paper proposes learning the transformation space as a potential solution to overcome GridCoder's limitations, providing preliminary experimental support.

https://arxiv.org/pdf/2411.17708