Imagine you are tasked with improving the performance of an underperforming LLM. What approach would you take?