Tune a model with rewards: let it answer, score the answers, and make the good ones more likely. REINFORCE, PPO’s clipped ratio, GRPO’s group scores and DPO’s preference pairs, one prompt and four bars.
This interactive visualization needs JavaScript.