1 paper
Ziwu Sun, Zhen Gao, Jiyong Zhang +1
We propose a new perspective on policy optimization: rather than reweighting all samples by their importance ratios, an optimizer should select which samples are trustworthy enough…