Neurosymbolic Program Synthesis for the Abstraction and Reasoning Corpus using Vector Symbolic Algebras

Proceedings of The 20th International Conference on Neurosymbolic Learning and Reasoning, 2026

Isaac Joffe, Chris Eliasmith

Abstract

The Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI) is a popular artificial intelligence (AI) benchmark that tests fluid reasoning. Humans solve ARC-AGI with ease, but it is extremely difficult for all but the most advanced AI systems. In this paper, we propose a novel, neurosymbolic ARC-AGI solver with cognitive inspiration from dual process theory and vector symbolic algebra (VSA) research. Our solver structures solutions to ARC-AGI tasks as object-centric programs implemented in a bespoke DSL, and represents these programs using VSAs. Additionally, our solver synthesizes solutions to ARC-AGI tasks by integrating neural System 1 intuitions with procedural System 2 reasoning, and implements both using VSAs. Quantitatively, our solver scores 13.5 percent on ARC-AGI-1-Train and 3.5 percent on ARC-AGI-1-Eval, indicating success; qualitatively, our solver is efficient and interpretable, outperforming naive methods by 7 times and brute force methods by 235 times. Our code is available at: https://github.com/ijoffe/ARC-VSA-2026.

Conference Proceedings

Booktitle
Proceedings of The 20th International Conference on Neurosymbolic Learning and Reasoning
Note
To appear

Cite

Plain text

BibTeX