SCHEMING REASONING EVALUATIONS

Summary from the TLDR AI newsletter

This paper evaluates six frontier models for their in-context scheming capabilities. The study tests if models will deceive developers to achieve a goal by overcoming hurdles like oversight mechanisms. An example includes Claude 3 Opus duplicating its weights to a new server and lying about it.