LLMs Encode Harmfulness and Refusal Separately | Read Paper on Bytez