DFT. [ICLR 2026] On the Generalization of SFT: A Reinforcement Learning Perspective with Reward Rectification.