@misc{zhao2025llms, title={LLMs Encode Harmfulness and Refusal Separately}, author={Jiachen Zhao and Jing Huang and Zhengxuan Wu and David Bau and Weiyan Shi}, year={2025}, eprint={2507.11878}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2507.11878} }