Controlled Chain of Thought: Eliciting Role-Play Understanding in LLM Through Prompts

Tabletop Role Playing Games (TRPG) are games that require players to become the characters they play through engaging in role-play. The challenge of training AI lies in the requirement of it not only understanding the explicitly stated game rules, but also the implicit ones that come with role-playing. Previous studies endeavouring said challenge allude to aspects of role-play, but do not emphasise its role in their methods, indicating that the definition of role-play and how it is to be employed remains unclear. This short paper aims to investigate a proposed definition of role-play based on previous research and employ its use on Large Language Model LLM, eliciting an understanding thereof through a novel prompting method dubbed Controlled Chain of Thought (CCoT). CCoT allows for the LLM to highlight absence of information in inputs given to it by generating questions, which become the template for its chain of thoughts when answered. This paper presents an initial pilot testing of CCoT as well as opens up the discussion of how a definition of role-play can be beneficial for future studies.