TailProp: content-adaptive light- and heavy-tailed propagation for vision

Jiahao Kong, Zihan Li

Science-inspired vision models show that explicit propagation dynamics can provide structured and interpretable alternatives to conventional token mixing. Existing formulations, ho…