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

A team of researchers has proposed a method to make large language models simulate human decision-making in persuasion games more accurately by anchoring them to mathematical models from cognitive science and economics, according to a paper submitted in 2026 [1]. The approach, called Equation-to-Behavior Prompting, guides large language models to replicate specific cognitive models of belief updating, such as Bayesian inference and Grether's α-β model, in strategic interactions based on legal decision-making [1]. The work addresses a known limitation: while artificial intelligence systems are increasingly used for safety evaluations and training, they often fail to capture the full range of human behavior in strategic settings [1]. The study found that large models could approximate equation-based specifications through prompting alone, but smaller models could not [1]. To close this gap, the researchers applied reinforcement learning in a technique they term Equation-to-Behavior RL. This training reduced belief error by 26.5% in out-of-distribution parameterizations for small models [1]. The research draws on a long tradition of using formal rules to understand persuasion and learning. Argumentation theory, for instance, studies how conclusions are supported or undermined through logical reasoning and encompasses forms of dialogue such as deliberation and negotiation [5]. Similarly, procedural rhetoric examines how rule-based systems and processes can make claims about how the world works, a concept formalized by Ian Bogost in his 2007 book "Persuasive Games" [4]. The cognitive models used in the paper also intersect with broader psychological theories. Social learning theory, developed by Albert Bandura, posits that people acquire behaviors and attitudes through observation and imitation within a social context, a process that includes vicarious reinforcement through observed rewards and punishments [2]. The new prompting method attempts to encode such structured decision-making patterns directly into language model behavior. The researchers demonstrated a practical benefit of this diversity. Training small models to consider different kinds of decision-makers improved average belief change by 2.5% to 12% over training that assumed only Bayesian reasoning, even when the models were later tasked with persuading GPT-5-mini [1]. The paper suggests the technique could improve human simulations for training and evaluation in increasingly realistic settings and enable novel research into more complicated mathematical models of human decision-making [1]. The preprint was posted on arXiv, an open-access repository for electronic preprints that has hosted over two million articles as of late 2021 and receives about 24,000 submissions per month [10].