Large language models in route choice: effects of feedback information on decision behavior and system outcomes

Abstract

This paper develops a multi-agent simulation framework based on large language models (LLMs) to study repeated routechoice behavior. In this framework, multiple LLM agents repeatedly make route-choice decisions under two feedback conditions: full-information feedback (FI) and partial-information feedback (PI). The experiment employs seven LLMs and compares differences in group behavior in a setting with 18 agents and 100 rounds of repeated decision-making. We analyze the stability and dynamic of route-choice behavior using RMSSD, deviation degree, and an efficiency index. The results show that LLMs exhibit significant behavioral heterogeneity in repeated route-choice tasks. Compared with FI, PI effectively reduces system fluctuations in most models and, in some cases, brings group behavior closer to the theoretical equilibrium state. Further analysis shows that the information structure changes group dynamics mainly by affecting the frequency and direction of individual strategy switching. Excessive information may induce synchronous switching behavior in some models, thereby amplifying system fluctuations. The LLM-based multi-agent simulation results highlight that more information does not necessarily lead to better collective outcomes. In contrast, excessive information may induce synchronized strategy switching among LLM agents, thereby amplifying system fluctuations and reducing behavioral stability.