1
LLMs Encode Harmfulness and Refusal Separately
最新研究发现：LLM能分离编码“有害性”与“拒绝行为”，揭示安全机制新维度。
arXiv:2507.11878v5 Announce Type: replace Abstract: LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond j…
2
Harnessing Textual Refusal Directions for Multimodal Safety
利用文本拒绝方向增强多模态AI安全性的前沿研究，为模型防御提供新思路。
arXiv:2606.31876v1 Announce Type: new Abstract: To improve safety in Large Language Models (LLMs) we can either perform post-training alignment or exp…