Toward understanding and preventing misalignment generalization

2025-06-18来源 OpenAI BlogEnglish
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.

原文：https://www.a8q9.cn/en/article/toward-understanding-and-preventing-misalignment-generalization-68498f.html

相关文章