Group relative policy optimization (GRPO)

A reinforcement learning method for language models. For each question it samples a group of answers and scores each one against the group's average, so no separate model is needed to estimate the expected score.