Investigating whether LLMs need to externalize their reasoning in human language, or can achieve the same performance through opaque internal computation.