Increasing the number of LLM calls in compound inference systems can initially improve performance but may later lead to a decline due to varying query difficulties.
