Enabling LLMs and MLLMs to carry out interpretable evaluation with reinforcement learning.