Evaluating Multimodal LLMs for Accessible Disaster Assistance
A recent preprint on arXiv (2608.14651) explores the effectiveness of open-weight Multi-Modal Large Language Models (MM-LLMs) in delivering reliable and actionable disaster support through both text and audio formats. This research highlights a significant shortcoming in emergency services, which often overlook individuals with specific access and functional needs, such as those who are hard of hearing, pregnant women, mothers with young children, and elderly individuals with dementia. By utilizing the latest advancements in AI, particularly MM-LLMs capable of integrating text, audio, images, and video in a chatbot framework, the study assesses the consistency of these systems across various communication methods. The analysis employs real emergency alert scenarios across four modalities, aiming to enhance the development of more inclusive AI-driven disaster communication solutions.
Key facts
- The paper is a preprint on arXiv with ID 2608.14651.
- It evaluates open-weight MM-LLMs for disaster assistance.
- The study focuses on text and audio modalities.
- Target populations include hard-of-hearing individuals, pregnant women, mothers with toddlers, and elderly with dementia.
- The research uses real emergency alert scenarios.
- The analysis covers four different modalities.
- The goal is to assess consistency and actionability of outputs.
- The paper is categorized as 'new' in the announcement type.
Entities
Institutions
- arXiv