Submitted by X.Y. Han 1 A Theoretical Framework for Auxiliary-Loss-Free Load Balancing of Sparse Mixture-of-Experts in Large-Scale AI Models University of Chicago 2