Presented at the NeurIPS 2024 workshop on Multimodal Abstraction & Reasoning
Full paper: arxiv.org/abs/2402.03507
For decades, AI research has aimed to replicate human-like abstraction and reasoning, yet broad generalization beyond training distributions remains elusive. The Abstraction & Reasoning Corpus (ARC), a dataset designed to test broad generalisation, has remained unsolved after five years, with the best solver based on handcrafted rules. We adapt two novel approaches based on neural networks: a neurosymbolic reasoning system based on DreamCoder, and a series of frontier large-language models, and compare their performance.