OpenAI ↗ · 2025-06-18 18:00 · UTC+8
Toward understanding and preventing misalignment generalization
Company news OpenAI Source language: English
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.
Excerpt
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.
Short excerpt from the source — open the original for the full text.