MULTIVERSE: Exposing Large Language Model Alignment Problems in Diverse Worlds

Understand

Large Language Model (LLM) alignment aims to ensure that LLM outputs match with human values.

• Researchers have demonstrated the severity of alignment problems with a large spectrum of jailbreak techniques that can induce LLMs to produce malicious content during conversations.
• Finding the corresponding jailbreaking prompts usually requires substantial human intelligence or computation resources.
• In this paper, we report that LLMs have different levels of alignment in various contexts.