UC San Diego

There have been several proposed methods for artificial intelligence systems. This thesis focuses on two approaches: library learning (neurosymbolic) and reasoning-based language models. We begin by proposing a dual-system for artificial intelligence -- where the system requires both knowledge and composition -- before introducing our two key settings that we frame through this dual-system. First, the symbolic domain is formally defined with program synthesis before highlighting an ambitious strategy in library learning. This is followed by original work studying a novel neurosymbolic approach to the ARC-AGI benchmark. Then, we introduce the neural domain -- where we frame reasoning as search and follow with work that studies how post-training data influences downstream reasoning behavior in chess. This thesis concludes by solidifying a unified framework for artificial intelligence that is based on identified similarities between the principled, theoretical system proposed in library learning and its practical counterpart in reasoning models. From this unified view, we see that language model breakthroughs have slowly converged to the library learning meta-architecture -- and we propose future directions of study based on the remaining discrepancies.