Researcher, University of California, Berkeley
5 papers at NeurIPS 2025
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
We develop private mechanisms for Euclidean Jordan algebras and private algorithms for symmetric cone programming, this leads to private algorithms for semidefinite programming.
