SAGE: An LLM-driven Self Reflective Agentic Framework for Fraud Detection

In three linesSAGE is an LLM-driven multi-agent framework for fraud detection. It coordinates three specialized agents using a six-layer Data Diagnostic Tree and Markov decision process guided by natural-language gradients. On five fraud datasets and five LLM backbones, SAGE improves F1 by 40.86% on average and wins 96% of method-dataset comparisons.

Summary generated by Claude — human-verified