Why not SFT-cold-start first?

[dszpr]

Hi! Many appreciate for the great work!
In the training of Deepseek-R1, they first used high-quality COT data for SFT-cold-start in Step-I, and then applied GRPO for training in Step-II.
However, in R1-V project, you directly used GRPO without SFT-cold-start. Why is that?