Training efficiency. Because the boundary conditions match, the loss remains the same. This allows the MLFM to be initialized directly from a pretrained MDM, avoiding the compute cost of pretraining…