Is cold start SFT always necessary before GRPO

hi,
i have gone through all the GRPO papers(except deepseek R1, where initially they trained R0 purely via GRPO) and in all of those they use cold start SFT before GRPO, i wanted to ask that if a model was good at following the output formatting and had no language issues, wouldn’t continued pre training be an alternative (if you wanted to add some extra domain knowledge), why didn’t anyone use it?