Comment by throwaway4aday

That's essentially what R1 Zero is showing:

> Notably, it is the first open research to validate that reasoning capabilities of LLMs can be incentivized purely through RL, without the need for SFT.