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

2026-06-16 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Social Sciences & Behavioral Studies · Depth: Expert, quick

Summary

A new approach, "Equation-to-Behavior Prompting," guides large language models to simulate diverse human decision-making in strategic interactions, addressing the common failure of current LLM simulations to cover the full spectrum of human behavior, from Bayesian updating to motivated reasoning. This method leverages mathematical models from cognitive science and economics. Researchers found that large models can approximate complex equation-based specifications, including Bayesian updating, affine distortion, motivated updating, and Grether's α-β model, through prompting. Small models, however, struggled with this approach. To overcome this, the study introduces "Equation-to-Behavior RL," a reinforcement learning method that reduces belief error by 26.5% in out-of-distribution parameterizations for small models. These enhanced simulations create diverse training environments, improving average belief change by 2.5%-12% over Bayesian-only training when persuading models like GPT-5-mini.

Key takeaway

For AI Scientists developing human-like simulations, you should integrate cognitive models to capture the full spectrum of human decision-making beyond simple Bayesian updating. This approach, using "Equation-to-Behavior Prompting" or "Equation-to-Behavior RL" for smaller models, significantly enhances the diversity and realism of simulated agents. Your training environments will become more robust, leading to improved model performance in complex strategic interactions and better safety evaluations.

Key insights

Cognitive models and RL can guide LLMs to simulate diverse human decision-making, improving training environments.

Principles

• Human decision-making is diverse, not just Bayesian.
• Mathematical models can specify complex behaviors.
• Small models need RL for complex behavior simulation.

Method

"Equation-to-Behavior Prompting" uses mathematical models to guide LLMs. For small models, "Equation-to-Behavior RL" applies reinforcement learning to adhere to these rules.

In practice

• Simulate diverse human biases for safety evaluations.
• Create varied training environments for LLMs.
• Improve persuasion capabilities of small models.

Topics

• Cognitive Models
• Language Model Simulation
• Persuasion Games
• Reinforcement Learning
• Human Decision-Making
• Safety Evaluation

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Mental-R1: Aligning LLM Reasoning for Mental Health Assessment — Aligning LLM reasoning with human cognitive processes via specialized reinforcement learning improves mental health assessment accuracy.
• Sell better with AI — Human decisions are primarily narrative-driven, requiring compelling stories to influence outcomes.
• Reinforcement Learning Course — This course provides a structured deep dive into Reinforcement Learning, from fundamentals to advanced applications.
• Reinforcement Learning with Human Feedback (RLHF) in 4 minutes — RLHF uses human feedback and reinforcement learning to align LLM outputs with human preferences.
• From Next-Word Prediction to World Models — AI is evolving from single-model next-word prediction to complex, efficient, and agentic "world models."
• Data News — Week 25.43 — Modern AI systems combine pre-training and reinforcement learning (RL) to achieve advanced reasoning and behavioral capabilities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.