1
Dialectics of Alignment: Harnessing Unsafe Knowledge for Dynamic Safety Routing
批判性审视LLM对齐中“擦除”范式的局限，主张利用不安全知识实现动态安全路由，拓展模型认知边界。
arXiv:2606.00686v1 Announce Type: new Abstract: The prevailing paradigm in large language model (LLM) alignment operates via erasure, filtering unsafe…