OpenAI 研究：如何理解和预防模型泛化偏差

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.