1 paper · 1 filter
Daichi Hayakawa, Issei Sato
In this study, we provide constructive proof that Transformers can recognize and generate hierarchical language efficiently with respect to model size, even without the need for a…