I'm pretty sure these models also went through RL post-training, the question you ask still needs to be interpreted as an instruction.