It's a clever idea: The authors instruct multiple LLM-simulated agents to assume different roles in a Socratic dialogue to figure out how to solve problems proposed by the user. Think of it as a <i>team</i> of LLM-simulated agents, each with a different set of goals and responsibilities (e.g., one may be creative, another may be analytical, another may be a proofreader, etc.).<p>Demo (video): <a href="https://vimeo.com/821089266" rel="nofollow">https://vimeo.com/821089266</a><p>Code: <a href="https://github.com/RunzheYang/SocraticAI">https://github.com/RunzheYang/SocraticAI</a>