Large Language Models Struggle to Describe the Haystack without Human Help: Human-in-the-loop Evaluation of LLMs