Diffusion Models for ARC AGI: A Surprisingly Difficult Task

This post details an attempt to solve the ARC AGI challenge using a diffusion model. The author adapted a fine-tuned autoregressive language model into a diffusion model, enabling non-sequential generation. While the diffusion approach achieved modestly better pixel accuracy, it didn't translate to improved task success rates. The key bottleneck was identified as the lack of efficient caching in the diffusion model's architecture, making it slower than the autoregressive baseline. Future work will focus on improving caching and developing more efficient candidate generation strategies.