Using Cognitive Models to Improve Language Model Simulation of Human Persuasion Games

The paper introduces a novel approach called Equation-to-Behavior Prompting, which enhances large language models' (LLMs) ability to simulate human decision-making in persuasion games by aligning them with cognitive models from economics and cognitive science. It demonstrates that while large models can effectively approximate cognitive specifications like Bayesian updating and motivated reasoning through prompting, smaller models can achieve a 26.5% reduction in belief error through reinforcement learning techniques. This research is significant for practitioners as it provides methods to create more diverse and realistic training environments for LLMs, ultimately improving their utility in applications requiring nuanced human-like decision-making.