## Unexpected Side Effects in Conversational Agents A recent study has highlighted how the use of certain stylistic features in prompts for conversational agents based on large language models (LLMs) can generate unexpected side effects. The research, presented on arXiv, focuses on analyzing how requesting a specific stylistic feature, such as conciseness, can influence other features, such as the perceived expertise of the agent. The researchers examined 12 frequently used stylistic features, finding that prompting for conciseness leads to a significant reduction in the perception of expertise. This result suggests that the different stylistic features are not independent, but deeply interconnected. ## CASSE: A Dataset for Studying Stylistic Effects To support future research in this area, a dataset called CASSE (Conversational Agent Stylistic Side Effects) has been created, which collects data related to these complex interactions. The study also evaluated mitigation strategies based on prompts and activation steering, finding that, although they can partially restore suppressed traits, they often worsen the primary intended style. These results challenge the assumption of faithful stylistic control in large language models and highlight the need for multi-objective and more principled approaches to safe and targeted stylistic steering in conversational agents. The research opens new perspectives for the development of more effective and reliable conversational agents.