SocialMindChange

SocialMindChange is a pioneering benchmark introduced to assess the advanced social intelligence of large language models (LLMs). Unlike traditional Theory of Mind (ToM) evaluations that merely require models to report on evolving mental states, SocialMindChange tasks LLMs with an active role: to generate dialogue that strategically influences and shifts another character's mental-state trajectory towards a predefined goal. This benchmark addresses a critical gap in current LLM capabilities, pushing them beyond passive observation to active social action. It operates by placing an LLM as one character within a multi-character social context across five connected scenes, requiring it to produce consistent dialogue while maintaining awareness of all participants' evolving beliefs, feelings, and intentions. This capability is crucial for developing more sophisticated, interactive, and socially adept AI systems in areas like empathetic chatbots, persuasive agents, and complex narrative generation.