Sample a group of responses, score each, and turn rewards into advantages using the group's own mean as the baseline — no critic network.
Each bar is one sampled response's reward; the dashed line is the group mean. GRPO turns rewards into advantages A_i = (r_i − mean)/std: above-average responses are pushed up, below-average pushed down, with no separate value network. Drag the rewards — make them all equal and watch the signal collapse to zero.
Illustrative model for building intuition, not a live system.