Starting from a policy with no intra-class style preference, RFT using exact policy gradients can maintain style symmetry without introducing style bias.