Training Agents 3: Reinforcement Learning
GRPO trains agents by sampling multiple completions per prompt and using the group's relative scores as the training signal. This method moves beyond imitation learning, allowing models to learn from their own varied trajectories.
Read article