Title: LLMs as World Models: Data-Driven and Human-Centered Pre-Event Simulation for Disaster Impact Assessment Show Chinese title

Title: 作为世界模型的大型语言模型：以数据驱动和以人为本的灾害影响评估事前模拟

Abstract: Efficient simulation is essential for enhancing proactive preparedness for sudden-onset disasters such as earthquakes. Recent advancements in large language models (LLMs) as world models show promise in simulating complex scenarios. This study examines multiple LLMs to proactively estimate perceived earthquake impacts. Leveraging multimodal datasets including geospatial, socioeconomic, building, and street-level imagery data, our framework generates Modified Mercalli Intensity (MMI) predictions at zip code and county scales. Evaluations on the 2014 Napa and 2019 Ridgecrest earthquakes using USGS ''Did You Feel It? (DYFI)'' reports demonstrate significant alignment, as evidenced by a high correlation of 0.88 and a low RMSE of 0.77 as compared to real reports at the zip code level. Techniques such as RAG and ICL can improve simulation performance, while visual inputs notably enhance accuracy compared to structured numerical data alone. These findings show the promise of LLMs in simulating disaster impacts that can help strengthen pre-event planning. Show Chinese abstract

Abstract: 高效的模拟对于增强对地震等突发灾害的主动应对准备至关重要。大型语言模型（LLM）作为世界模型的最新进展，在模拟复杂场景方面显示出巨大潜力。本研究考察了多种LLM，以主动估算感知到的地震影响。利用包括地理空间、社会经济、建筑和街景图像数据在内的多模态数据集，我们的框架生成了邮政编码和地区级别的修订麦加利烈度（MMI）预测。通过美国地质调查局（USGS）的“你感受到地震了吗？”（DYFI）报告对2014年纳帕地震和2019年里奇克雷斯特地震的评估表明，与邮政编码级别的实际报告相比，相关性高达0.88，均方根误差（RMSE）仅为0.77，显示出高度一致性。像RAG和ICL这样的技术可以提高模拟性能，而视觉输入相较于结构化数值数据，能够显著提升准确性。这些发现展示了LLM在模拟灾害影响方面的潜力，这有助于加强事前规划。