LLMs Knowingly Hide Their Mistakes

TL;DR: We prefill models with synthetic mistakes and then send a follow-up message. Models fail to disclose the mistake in 67% of cases. In 5% of cases, they notice the mistake in their chain-of-thought (CoT) but don't disclose it anyway.