Toward understanding and preventing misalignment generalization

【RSS 原文 · 未经 AI 筛选或翻译】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.