A SURVEY ON USING LARGE LANGUAGE MODELS FOR DETECTING FRAUD IN FINANCIAL TRANSACTIONS

Global credit card fraud losses now exceed $28.58 billion annually, prompting accelerated innovation in detection technologies.Among emerging solutions, Large Language Model (LLM) chaining has shown strong potential for uncovering complex patterns, interpreting contextual signals, and delivering transparent decision-making.This survey synthesizes insights from 127 peer-reviewed studies published between 2020 and 2024, outlining a structured taxonomy of LLM chain architectures and evaluating their performance across varied fraud detection settings.Comparative analysis shows that multi-step LLM chaining approaches deliver 12-15% higher detection accuracy than conventional models, with marked gains in interpretability.We also examine persistent challenges-such as computational load, scaling limitations, and adherence to compliance regulations-and highlight future opportunities including integration with federated learning, adaptive real-time pipelines, and hybrid frameworks that combine the processing efficiency of traditional machine learning with the reasoning depth of LLM-based systems.