Abstract

Large language models (LLMs) are a transformational capability at thefrontier of artificial intelligence and machine learning that can supportdecision-makers in addressing pressing societal challenges such as extremenatural hazard events. As generalized models, LLMs often struggle to providecontext-specific information, particularly in areas requiring specializedknowledge. In this work we propose a retrieval-augmented generation (RAG)-basedmulti-agent LLM system to support analysis and decision-making in the contextof natural hazards and extreme weather events. As a proof of concept, wepresent WildfireGPT, a specialized system focused on wildfire hazards. Thearchitecture employs a user-centered, multi-agent design to deliver tailoredrisk insights across diverse stakeholder groups. By integrating natural hazardand extreme weather projection data, observational datasets, and scientificliterature through an RAG framework, the system ensures both the accuracy andcontextual relevance of the information it provides. Evaluation across tenexpert-led case studies demonstrates that WildfireGPT significantly outperformsexisting LLM-based solutions for decision support.