1. take a model and prompt e.g. "name this color: #D02027" 2. generate multiple outputs 3. score outputs 4. calculate group average 5. encourage outputs that beat the average full version with interactive visualizations: adaptive-ml.com/post/grpo-simp…