So is GRPO that much better because it ascribes feedback to a whole tight band of ‘quality’ ranges of on-policy answers while the band tends towards improvement

So is GRPO that much better because it ascribes feedback to a whole tight band of ‘quality’ ranges of on-policy answers while the band tends towards improvement in the aggregate, or is it just faster algorithm = more updates for a given training duration?