Why is hyperparameter optimization so hard for reinforcement learning, and how is AgileRL changing this?

Combining GRPO with evolutionary hyperparameter optimization to squeeze the most out of small LLMs