“Approaches based on discrete program search and program synthesis are performing better on the ARC benchmark than large language models.”