That is sort of like saying why apply reinforcement learning (most of the modern improvements in LLMs have come from RL) to any raw LLM. They are refining it to perform better at certain tasks.