Explain GRPO and why dropping the value network matters at scale.

GRPO compares several sampled responses within a group and normalizes their rewards, avoiding a separate value network. This saves memory at scale but still depends on reward quality and sufficiently diverse samples.

Use equations or tensor shapes where they clarify the claim, then name an experiment or ablation that would distinguish competing explanations.