Large language model applications in disaster management: An interdisciplinary review

Disasters increasingly challenge urban resilience, demanding advanced computational approaches for effective information management and response coordination. This interdisciplinary review systematically assesses Large Language Model (LLM) applications in disaster management, analyzing 70 LLM-focused studies within the broader landscape of AI-driven disaster management. Our analysis establishes a phase-based framework spanning detection, tracking, analysis, and action, and reveals three critical gaps in current disaster management solutions: limited advancement beyond disaster response to include preparedness, recovery, and mitigation phases; insufficient integration across diverse stakeholder groups and available resources; and inadequate transformation of situation awareness data into actionable insights. Leveraging cross-modal semantic reasoning, knowledge graph-constrained entity extraction, and advanced code generation, LLMs are well positioned to overcome information ambiguity and verification challenges often encountered in rapidly evolving disaster contexts. These capabilities also enable automation in disaster investigation and communication, effectively orchestrating diverse analytical tools and resources. To harness these advantages and promote further progress, we introduce the "3M" framework for intelligent disaster information management: multi-modal data fusion for integrated assessment, multi-source information validation for robust truth-finding, and multi-agent collaboration in physical-virtual disaster systems. This framework provides a systematic foundation for advancing next-generation LLM-driven disaster management research and practice in increasingly complex contexts