KL Divergence in Post-Training — Reverse versus Forward