A new benchmark study evaluates the effectiveness of gradient-boosted decision trees (GBDTs) versus large language model (LLM) agents in payment fraud detection. Conducted on a local CPU without the need for GPUs or cloud accounts, the benchmark assesses latency, cost, and reproducibility. The results indicate that GBDTs remain superior for synchronous payment authorizations, while agents are better suited for asynchronous tasks. The study provides detailed metrics including a PR-AUC of 0.847 and ROC-AUC of 0.931 for the GBDT model, based on synthetic data that simulates real-world transaction features. The source code for the benchmark is publicly available on GitHub, allowing for independent verification of results.
Fraud DetectionBenchmarkLLM+1
Fraud DetectionBenchmarkLLMGradient Boosted Decision Trees
Benchmark Study Reveals GBDTs Outperform LLM Agents in Payment Fraud Detection
Towards Data Science· Sandeep Bharadwaj Mannapur· Friday, June 26, 2026
More Articles From This Day
Generative AIAI Agents+2
Generative AIAI AgentsEnterprise AdoptionMicrosoft Azure
Microsoft Launches Discovery to Transform Scientific Research with Agentic AI
Microsoft has introduced Discovery, a cloud-based enterprise agentic AI platform aimed at enhancing scientific research across data-heavy fields such as chemistry, materials science, and life sciences. Launched in private preview at Microsoft Build 2025 and made generally available in June 2026, Discovery addresses the challenges of synthesizing vast amounts of research data by providing AI-driven digital lab assistants. These agents actively participate in the scientific process, from hypothesis formation to data analysis and experimental design, while ensuring enterprise-grade governance and security. The platform's central Discovery Engine facilitates knowledge mapping and multi-agent orchestration, enabling rapid movement from hypothesis to discovery by consolidating fragmented data.