Neural Cellular Automata for ARC-AGI

TL;DR

This paper investigates Neural Cellular Automata's ability to perform complex, grid-based tasks requiring precise transformations and generalization, using gradient training on the ARC-AGI benchmark to evaluate their potential for abstract reasoning.

Contribution

It introduces a gradient-based training method for NCAs applied to ARC-AGI tasks, providing insights into their capabilities and limitations for abstract reasoning and self-organizing systems.

Findings

Gradient-trained NCAs show promise for ARC tasks

Design modifications impact NCA performance

Insights into NCA behavior for generalization

Abstract

Cellular automata and their differentiable counterparts, Neural Cellular Automata (NCA), are highly expressive and capable of surprisingly complex behaviors. This paper explores how NCAs perform when applied to tasks requiring precise transformations and few-shot generalization, using the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI) as a domain that challenges their capabilities in ways not previously explored. Specifically, this paper uses gradient-based training to learn iterative update rules that transform input grids into their outputs from the training examples and apply them to the test inputs. Results suggest that gradient-trained NCA models are a promising and efficient approach to a range of abstract grid-based tasks from ARC. Along with discussing the impacts of various design modifications and training constraints, this work examines the behavior and properties of NCAs applied to ARC to give insights for broader applications of self-organizing systems.