Toward understanding and preventing misalignment generalization

核心要点速览

We study how training on incorrect responses can cause broader misalignment in language models and identify an internal feature driving this behavior—one that can be reversed with minimal fine-tuning.