Skip to main content
Artificial Intelligence

Toward understanding and preventing misalignment generalization

OpenAI Blog · · 4 views

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.

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.