OpenAI ↗ · 2025-06-18 18:00 · UTC+8
Toward understanding and preventing misalignment generalization
Media reports OpenAI Source language: English
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.