SpecSlot Decouples Reading, Allocation and Routing in Pixel-Space Generative Transformers

The Read, Allocate, Route paper argues that one-token-per-patch pixel-space transformers conflate what the model reads, what each token predicts and how many tokens join global attention — making an 8x8 patch grid 16x more expensive than 16x16 attention on 256x256 images. SpecSlot splits those roles across separate token sets, letting detail-heavy regions get more compute without blowing up the quadratic attention cost.