Calibrating Behavioral Parameters with Large Language Models

Can Large Language Models serve as behavioral functions for multi-agent economic simulation? We address this through eight experiments testing whether LLMs generate canonical behavioral biases when reasoning about synthetic financial scenarios. Using four models (GPT-4o, GPT-4o-mini, Claude-3.5-Haiku, Gemini-2.5-Pro) across 24,000 agent-scenario pairs, we find a striking pattern: LLMs exhibit \textit{less} behavioral bias than humans on most measures. Disposition ratios range from 0.01--0.21 versus the human benchmark of 1.6; loss aversion $\lambda$ ranges from 1.12--1.90 versus 2.25. However, behavioral profiles successfully shift LLM behavior toward human-like patterns, with loss-averse profiles achieving $\lambda = 3.0$ and herding-prone profiles reaching 90\% herding rates. These findings suggest LLMs may serve as \textit{calibratable} behavioral agents for mechanism design research, enabling scalable simulation of strategic interactions with heterogeneous, psychologically-grounded decision-makers.