IME LLMs hallucinate much less when given a text that it can summarize or answer questions about vs. asking it to generate something