Rigidity in LLM Bandits with Implications for Human-AI Dyads

This repository accompanies the paper “Rigidity in LLM Bandits with Implications for Human-AI Dyads” .It provides complete experimental data, model outputs, and analysis scripts used to evaluate how large language models (LLMs) exhibit decision rigidity and bias amplification under controlled two-armed bandit paradigms. The dataset quantify how LLMs (DeepSeek, GPT-4.1, Gemini-2.5) adapt, or fail to adapt, across different decoding regimes defined by temperature and top-p sampling.By treating LLMs as participants in reinforcement-learning tasks, the study identifies systematic tendencies toward stubborn exploitation and reduced exploration, modeled via hierarchical Rescorla–Wagner fits in Stan.本开源仓库配套于论文《大语言模型老虎机任务中的决策刚性及其对人机二元交互的启示》。本仓库提供完整的实验数据、模型输出结果与分析脚本，用于评估大语言模型（Large Language Model，LLM）在受控双臂老虎机范式下所表现出的决策刚性与偏差放大效应。 本数据集量化了DeepSeek、GPT-4.1、Gemini-2.5等多款大语言模型在由温度（temperature）与top-p采样（top-p sampling）定义的不同解码策略下的适应与适应失效情况。本研究将大语言模型视作强化学习任务的参与主体，识别出其系统性的顽固利用倾向与探索行为减少特征，并通过Stan统计建模框架中的分层Rescorla-Wagner模型完成拟合建模。