ALARM & CrowdLLM: Streamline LLMs for Quality Decision-Making

Large language models (LLMs) have demonstrated unprecedented capabilities in understanding and reasoning across complex tasks. They offer the potential to build cost-effective, accurate, and reliable AI systems that assist humans in processing massive amounts of information and making timely, high-quality decisions. For example, LLM-based systems can enable smart-home monitoring by detecting anomalies from real-time surveillance video and recommending preventive actions. However, significant challenges remain in streamlining LLMs for robust and trustworthy decision-making, as they are often prone to uncertainty, bias, and limited diversity in their outcomes. In this talk, we introduce two frameworks designed to address these challenges: ALARM, an LLM-based anomaly detection system with integrated uncertainty quantification, and CrowdLLM, a framework that combines LLMs with generative models to create a “digital population” serving as a cost-effective digital twin for applications such as crowdsourcing, social simulation, and recommendation systems. We conclude with a discussion of future directions toward dependable and interpretable LLM-based decision systems.