Question about ground truth label

[Trae1ounG]

Thank you for your excellent work!

May I ask if you have ever directly used the model-generated y instead of replacing it with ground truth? Would this approach enable effective training?

[jibo27]

Thank you for your interest!
We always replace y with the ground truth label, so we don't use the model-generated y.
This is an interesting idea, but I don't think this would be very effective. Since y is generated by the model itself, its probability—whether the result is correct or incorrect—would be high. This would lead to a high reward based on our PR definition. In this case, there would be no effective reward signal for training the model.
Please correct me if I misunderstood.