Training affine maps to transfer linear features from large to small models often destroys the very information intended for transfer. This early-stage research suggests that standard bidirectional loss leads to feature collapse.