Reproduce “Vector Symbolic Algebras for the Abstraction and Reasoning Corpus” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Isaac Joffe, Chris Eliasmith
Researchers from the University of Waterloo developed a neurosymbolic solver for the Abstraction and Reasoning Corpus (ARC-AGI) using Vector Symbolic Algebras (VSAs) to model human dual-process cognition, achieving 10.8% task accuracy on ARC-AGI-1-Train and outperforming large language models on several related, simpler benchmarks.
Key insights
- VSAs provide a unified neurosymbolic framework that bridges symbolic reasoning and connectionist learning, enabling robust distributed representations for abstract concepts and facilitating both structured manipulation and neural-like operations.
- Explicitly modeling human dual-process cognition (System 1 intuition guiding System 2 reasoning) through VSA-powered heuristics leads to a computationally efficient and interpretable AI solver for complex abstraction tasks.
- Neurosymbolic approaches, by design, can address fundamental limitations of large language models (LLMs) in tasks requiring precise numerical reasoning, explicit rule-following, and high sample efficiency, offering a distinct advantage for fluid intelligence problems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.