“Current LLMs perform poorly in multi-agent simulations because they lack sufficient training data on when to participate in group conversations, resulting in behavior that is either too quiet or too talkative.”