Jerry Yao-Chieh Hu

We present a unified theoretical framework for standard and high-order flow matching and prove their minimax optimality.

We establish the universal approximation capability of single-layer, single-head self- and cross-attention mechanisms