Cognitive Models Enhance LLM Simulation of Human Persuasion

Large language models (LLMs) are increasingly used to simulate human behavior for safety evaluations and training purposes. However, these simulations often fail to capture the full spectrum of human decision-making, which can range from rational Bayesian updating to biased motivated reasoning. To address this limitation, researchers suggest leveraging mathematical models from cognitive science and economics to guide LLMs in mimicking a broader range of human behaviors. The proposed approach, termed "Equation-to-Behavior Prompting," involves instructing LLMs to adhere to specific mathematical models of human decision-making. This method was evaluated on persuasion games based on legal scenarios, demonstrating that large models can approximate various equation-based specifications, including Bayesian updating and models of motivated reasoning. Smaller models, however, struggled with this prompting technique. To improve smaller models, "Equation-to-Behavior RL" was introduced, using reinforcement learning to train them to follow mathematical rules. This significantly reduced belief error in out-of-distribution parameterizations. The study highlights that these enhanced simulations can create more diverse training environments, leading to improved performance in persuading other LLMs. This work paves the way for more realistic human simulations, benefiting AI training and evaluation, and enabling deeper research into complex human decision-making models.