Self-attention in transformers does two jobs at once, according to a new arXiv preprint: it decides where to read and how to mix. The authors argue that bundling those functions creates three costs: O(L^2) training, an O(T) decode state, and position signals bound to the training length. The proposed model, FON, is described as a draft-revise dual-organ language model that removes sequence-level quadratic attention.

FON's design separates the two roles rather than assigning them to one mechanism. The title indicates a draft-revise structure, but the abstract gives no further architectural detail. It also notes that existing attention-free alternatives have their own limitations; the provided text stops before specifying them.

Because the source is a single announcement abstract, this digest can only report the motivation and the proposed direction. No experimental comparisons or results are included in the available text.