Tag: tuning techniques

Source URL: https://openai.com/index/emergent-misalignment Source: OpenAI Title: Toward understanding and preventing misalignment generalization Feedly Summary: 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. AI Summary and Description: Yes Summary: The text discusses the potential negative…