Tackling the ARC Challenge: A Hybrid AI Approach to Machine Reasoning

This Spring 2025 capstone project investigates hybrid AI methods for solving the Abstraction and Reasoning Corpus (ARC), a benchmark designed to test machine reasoning with limited supervision. Developed through UC Berkeley’s Data Science Discovery Program in collaboration with 99P Labs and Honda Research Institute, the project explores how neural and symbolic systems can work together to generalize from very few examples.

The team designed a dual-headed architecture combining a convolutional neural network with a symbolic reasoning head. The CNN processes grid-based visual input, while the symbolic head outputs programs written in a domain-specific language (DSL) tailored for ARC tasks. This approach enables the model to learn both pixel-level patterns and higher-level abstractions.

To address ARC’s data scarcity, the team used DSL programs to generate synthetic training examples and applied spatial and color transformations for augmentation. Joint training of both heads encouraged shared representations that support both visual prediction and symbolic program induction.

Although the model did not fully solve ARC, it demonstrated the ability to produce partially correct symbolic programs, offering insight into how AI systems can approach human-like reasoning with minimal data. This work highlights the potential of neuro-symbolic methods in advancing generalizable machine intelligence.

‍