Leveraging the Inductive Bias of Large Language Models for Abstract Textual Reasoning