Abstract
This paper addresses the challenge of enhancing artificial intelligencereasoning capabilities, focusing on logicality within the Abstraction andReasoning Corpus (ARC). Humans solve such visual reasoning tasks based on theirobservations and hypotheses, and they can explain their solutions with a properreason. However, many previous approaches focused only on the grid transitionand it is not enough for AI to provide reasonable and human-like solutions. Byconsidering the human process of solving visual reasoning tasks, we haveconcluded that the thinking process is likely the abductive reasoning process.Thus, we propose a novel framework that symbolically represents the observeddata into a knowledge graph and extracts core knowledge that can be used forsolution generation. This information limits the solution search space andhelps provide a reasonable mid-process. Our approach holds promise forimproving AI performance on ARC tasks by effectively narrowing the solutionspace and providing logical solutions grounded in core knowledge extraction.