Toward understanding and preventing misalignment generalization
来自 OpenAI News
· 2025-06-18
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.