1 paper
Eight Suzuki, Hideitsu Hino, Noboru Murata
Transformers achieve strong performance, but their internal computations remain opaque. We view each Transformer layer as a dynamic graph whose nodes are token representations and…