Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs
https://arxiv.org/pdf/2512.09742