Diffusion Models for ARC-AGI: A Retrospective (christopherhwood.com)

In a recent retrospective, a machine learning researcher shared their exploration of applying diffusion models to the ARC-AGI abstract reasoning benchmark, marking their first project after learning about diffusion techniques. The researcher initially attempted to leverage signal processing methods but found that unknown transformations between inputs and outputs required a learning-based approach. By conceptualizing ARC grids as “clean” and “corrupted” signals, they formulated a model to reverse the corruption through a deterministic forward process, transitioning away from standard stochastic diffusion. This work is significant as it highlights the potential of diffusion models within the AI/ML community, especially in structured transformation tasks, where conventional methods had dominated. Through a series of iterations, the researcher developed a latent diffusion framework, employing autoencoders and a UNet denoiser, leading to major performance improvements. Their innovative architecture used a “clue giver” to guide solvers through noise, enabling the model to explore diverse denoising strategies. Though the project solved one task and nearly cracked another, key engineering challenges remain, underscoring the transformative capabilities of tailored diffusion models in understanding complex data relationships and paving the way for future research in this area.