Hi, do you have any idea to avoid reward hacking?

[syguan96]

Thanks for sharing this project!
I found that training is not stable enough, and the model tends to hack the reward model. Could you share any ideas about these?

[fikry102]

Thanks for sharing this project! I found that training is not stable enough, and the model tends to hack the reward model. Could you share any ideas about these?

你好，我感觉是不是因为用的单个reward model，所以还是比较容易出现hacking?
我之前做的一个文生图工作也发现了类似的问题：https://arxiv.org/abs/2512.00743

方便加个好友交流下吗？