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

In three linesResearchers propose Equation-to-Behavior Prompting to guide LLMs to simulate diverse cognitive models (Bayesian, motivated reasoning, Grether's α-β model). Large models approximate these specifications via prompting, but small models fail. RL training reduces belief error by 26.5% and improves performance by 2.5–12% on legal persuasion games.

Summary generated by Claude — human-verified