Sentimental Agents: Combining Sentiment Analysis and Non-Bayesian Updating for Cooperative Decision-Making

With ongoing exploration of Large Language Model(LLM)-based multi-agent systems, it is becoming increasingly important to understand and interpret the dynamics of agent interactions and their beliefs, particularly when designed to emulate diverse roles and perspectives or to engage in debates. At present, there are no unified solutions that can systematically interpret and analyze the beliefs and interactions of these agents. This study introduces Sentimental Agents, a framework designed to support decision-making by providing multiple perspectives on a topic. Agents within this framework are equipped with a mental model of self, articulated in natural language. We have integrated sentiment analysis with a non-Bayesian updating mechanism to interpret and analyze the agents' beliefs and interactions systematically. A collective viewpoint is achieved when the update is marginal. We have adapted this framework for a simulated scenario in the Human Resource domain, implementing a conceptual tool known as the Artificial Board of Advisors (ABA). A key focus of this simulation is the application of ABA in the assessment of candidates for roles, showcasing its potential application in a theoretical HR recruiting environment.