How to Beat ARC-AGI by Combining Deep Learning and Program Synthesis

Program synthesis means searching for a program that explains the examples of a task and can handle a new input. In these 2024 speaker notes, François Chollet proposes using deep learning to guide that search. A learned model could suggest promising program structures or interpret an input into pieces that a search system can manipulate.

The motivation is the number of possible programs: blindly testing combinations becomes impractical as they grow. Chollet's proposed combination uses learned pattern recognition to narrow the choices, then explicit search to construct a solution. This explains a research direction behind ARC Prize, while keeping Chollet's prediction that such ideas would crack the benchmark within a few years distinct from an already demonstrated result.