未提及具体数据集名称
收藏资源简介:
本文提出了一种名为CrimeMind的新型LLM驱动ABM框架,用于模拟多模态城市环境中的城市犯罪。该框架将日常活动理论(RAT)融入代理工作流程,使其能够处理丰富的多模态城市特征并对犯罪行为进行推理。为解决LLM代理在评估环境安全性方面的挑战,论文收集了一个小规模的人工标注数据集,并通过一种无需训练的文本梯度方法将CrimeMind的感知与人类判断进行对齐。实验结果表明,CrimeMind在犯罪热点预测和空间分布准确性方面优于传统的ABM和深度学习基线。
This paper proposes a novel LLM-driven ABM framework named CrimeMind for simulating urban crime in multimodal urban environments. The framework integrates Routine Activity Theory (RAT) into the agent workflow, enabling it to handle rich multimodal urban features and reason about criminal behaviors. To address the challenges faced by LLM agents in evaluating environmental safety, this paper collects a small-scale manually annotated dataset, and aligns CrimeMind's perceptions with human judgments through a training-free textual gradient method. Experimental results demonstrate that CrimeMind outperforms traditional ABM and deep learning baselines in terms of crime hotspot prediction and spatial distribution accuracy.

- 1CrimeMind: Simulating Urban Crime with Multi-Modal LLM Agents清华大学电子工程系,香港浸会大学社会学系 · 2025年



