About Entropy Trend

[TU2021]

Thanks for your great work!

In DeepscaleR reproduce, someone and me, confused on the training trend about the training entropy trend, just like this link: https://github.com/agentica-project/deepscaler/issues/15

Can you share your training entropy log ? It seems that it is the core difference to reproduce deepscaleR and your work (maybe, because you directly use their code).

[nick7nlp]

When reproducing DeepScaleR, we observed that the entropy loss showed the same trend as in DeepScaleR but differed from the trend exhibited by our training strategy (FastCuRL).

[TU2021]

Interesting observation: in the DeepscaleR official repo, entropy increases （stage1 2） and then decreases （stage3）, while in FastCuRL it is the opposite （stage1 decrease，stage234 increase）.
Do you have any novel understanding of this difference?