Researchers have developed a non-transformer model that achieves superior performance on the ARC-AGI-1 benchmark using significantly less compute than traditional systems. This approach utilizes recurrent memory and latent reasoning to perform cost-efficient rule inference at scale.