Using Cognitive Models to Improve Language Model Simulation of Human Persuasion Games
Researchers have developed a new method called Equation-to-Behavior Prompting to guide large language models (LLMs) in simulating diverse human decision-making behaviors, moving beyond simple Bayesian updating. This approach was tested on persuasion games, showing that larger models can approximate specified cognitive models through prompting. For smaller models, a reinforcement learning technique, Equation-to-Behavior RL, significantly reduced belief errors, particularly in out-of-distribution scenarios. Training smaller models with these diverse decision-maker simulations improved their average belief change compared to training solely on Bayesian models, even when interacting with models like GPT-5-mini. AI
IMPACT Enhances LLM training and evaluation by enabling more realistic simulations of diverse human decision-making.