Researchers have made a significant discovery about how large language models (LLMs) internally represent rhetorical questions, which are asked to persuade or signal stance rather than seek information. By conducting a linear probing study on two social-media datasets with distinct discourse contexts, they found that rhetorical signals emerge early in the representation process. Notably, these signals are most stably captured by the last-token representation, providing insight into the inner workings of LLMs. This finding has implications for understanding how LLMs process and generate text, particularly in contexts where rhetorical questions are commonly used, such as social media and political discourse1. The ability to effectively represent rhetorical questions can impact the development of more sophisticated language models, and this study contributes to the ongoing effort to improve LLMs. This matters to practitioners because it can inform the development of more effective natural language processing systems, which can have significant geopolitical implications.