Evaluating needle in a haystack the responses are abnormal

Hello, Author. When I changed the question in the "searching for a needle in the haystack" evaluation from one about the needle to a different question (for example, "please describe this video"), the LLM (Large Language Model) outputs were all garbled characters. Only when the query pertains to the needle does the output come out as expected. What could be causing this issue?