GRPO loss通过配置系数置为1这块，有个小细节想确认下

[CccccchenJD]

如果你们实现过程中，采用的①，此时系数趋于∞
如果实现中，采用的②，此时系数趋近于-1，虽然不等于-3/2，但应该影响不大。

[SonicZun]

Thank you for your interest in our work, and apologies for the delayed reply.
In the paper, we did not separately address the "truncated noise schedule" for simplicity.
Our actual implementation is consistent with Equation ② in your derivation.