Program Synthesis Using Inductive Logic Programming for the Abstraction and Reasoning Corpus.

The article focuses on ILPAR, a program synthesis system based on Inductive Logic Programming (ILP), designed to address the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI), an unsolved benchmark requiring strong generalization and reasoning. ILPAR uses a small Domain Specific Language (DSL) of object-centric abstractions to represent input-output grid transformations as sequences of logic programs in First-Order Logic, enabling it to learn from few examples and generalize to unseen tasks. Evaluated on a sample of 20 unseen ARC-AGI tasks, ILPAR achieved 30% accuracy, demonstrating competitive performance compared to other symbolic AI approaches, though with higher computational cost. The system’s limitations include a restricted DSL coverage of human core knowledge priors and a fixed maximum program length, with plans to extend the DSL, improve efficiency, and integrate deep learning to guide program search.