1 paper
Junkai Zhang, Zihao Wang, Lin Gui +7
Reinforcement fine-tuning (RFT) often suffers from reward over-optimization, where a policy model hacks the reward signals to achieve high scores while producing low-quality output…