ARC-NCA: Towards Developmental Solutions to the Abstraction and Reasoning Corpus (2505.08778v1)

Abstract: The Abstraction and Reasoning Corpus (ARC), later renamed ARC-AGI, poses a fundamental challenge in artificial general intelligence (AGI), requiring solutions that exhibit robust abstraction and reasoning capabilities across diverse tasks, while only few (with median count of three) correct examples are presented. While ARC-AGI remains very challenging for artificial intelligence systems, it is rather easy for humans. This paper introduces ARC-NCA, a developmental approach leveraging standard Neural Cellular Automata (NCA) and NCA enhanced with hidden memories (EngramNCA) to tackle the ARC-AGI benchmark. NCAs are employed for their inherent ability to simulate complex dynamics and emergent patterns, mimicking developmental processes observed in biological systems. Developmental solutions may offer a promising avenue for enhancing AI's problem-solving capabilities beyond mere training data extrapolation. ARC-NCA demonstrates how integrating developmental principles into computational models can foster adaptive reasoning and abstraction. We show that our ARC-NCA proof-of-concept results may be comparable to, and sometimes surpass, that of ChatGPT 4.5, at a fraction of the cost.

Overview of ARC-NCA: Developmental Solutions for the Abstraction and Reasoning Corpus

The paper "ARC-NCA: Towards Developmental Solutions to the Abstraction and Reasoning Corpus" explores the efficacy of Neural Cellular Automata (NCA) in tackling the challenges posed by the ARC benchmark, which tests artificial general intelligence (AGI) systems' ability to perform abstraction and reasoning with minimal examples. The research works under the hypothesis that developmental computation—exemplified by NCAs—can offer a pathway to improved problem-solving capabilities not merely based on data extrapolation but through emergent, self-organizing processes.

Key Contributions and Results

The primary contribution of this paper lies in the introduction of ARC-NCA, a framework leveraging NCAs and an enhanced variation referred to as EngramNCA. ARC-NCA aims to explore whether the developmental nature of NCAs can accommodate the demands of abstraction and reasoning benchmarks like ARC-ag. The NCA are configured to mirror biological systems' dynamic and adaptive characteristics; EngramNCA introduces hidden memory states for more refined abstraction capabilities.

The paper reports that ARC-NCA achieves solve rates competitive with, and at times surpassing, the ChatGPT 4.5—a popular LLM—while requiring significantly lower computational resources. Specifically, EngramNCA v3 demonstrated the highest solve rate at approximately 12.9% using individual model outputs, and further improved with union models like NCA $\cup$ EngramNCA v3 reaching up to 15.3%. The paper emphasizes the cost-effectiveness of ARC-NCA models, achieving similar performance to ChatGPT 4.5 at a fraction of the operational costs—about 1000 times less expensive per task.

Technical Insights and Model Architecture

The paper goes into substantial detail explaining the technicalities of ARC-NCA and the proposed models. Standard NCAs operate on lattices where cells react to local interactions steered by neural networks. In contrast, EngramNCA extends this by incorporating dual-state cells, simulating memory transfer processes akin to those in biological systems.

Several architectural augmentations are introduced to better suit ARC-specific challenges such as position variance and the requirement for precise interaction with grid edges. EngramNCA variants experimented with different sensing mechanisms, and handling both toroidal and non-toroidal information processing to adapt cell behavior suitably.

Interpretation and Implications

The findings support the notion that developmental approaches can bridge the gap between human-like reasoning and current AI systems, particularly for few-shot learning benchmarks like ARC-AGI. Such approaches embody a potential paradigm shift from brute-force or purely data-driven methodologies to more adaptive, human-mimicking strategies within AGI research.

The successful application of NCAs in this context highlights the promise of cellular automata as model architectures that can handle complex visual reasoning tasks through their inherent emergent properties. By drawing inspiration from developmental biology, the ARC-NCA framework aligns well with the direction AI needs to pursue for achieving more generalized intelligence.

Future Directions

Although ARC-NCA shows promising results, there is room for further research, especially regarding the ARC-AGI-2 benchmark, which introduces even more intricate reasoning challenges. Extensions of this work may involve pre-training strategies leveraging criticality or the incorporation of such developmental models within hybrid systems alongside LLMs.

Exploring NCA’s role in latent representation processing offers additional avenues for enhancing human-like abstraction and reasoning capabilities. Through these future developments, ARC-NCA holds potential to significantly advance AI’s adaptability in performing complex cognitive tasks.

This paper indicates a renewed interest in developmental computation as a viable approach to AI's open problems, remaining central to AGI advancements and artificial life research.