1
The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment
提出“背锅假说”揭示大模型微调中涌现错位的泛化机制，并探讨缓解方法，挑战传统认知。
arXiv:2606.06667v1 Announce Type: new Abstract: The mechanisms behind LLMs' broad over-generalization beyond training examples remain unclear. Emergen…