Explainable AML Triage with LLMs: Evidence Retrieval and Counterfactual Checks
Researchers propose an explainable AML triage framework using LLMs to handle transaction alerts while mitigating hallucinations and ensuring compliance. The approach emphasizes evidence-constrained decision-making to improve auditability and governance.
Researchers have introduced a new framework for anti-money laundering (AML) triage that leverages large language models (LLMs) to handle the vast volumes of transaction alerts generated by monitoring systems. The framework aims to address the challenges of hallucinations, weak provenance, and unfaithful explanations that arise from unconstrained LLM outputs in regulated workflows.
The proposed system treats AML triage as an evidence-constrained decision process, ensuring that the models' outputs are grounded in verifiable data. This approach not only enhances the accuracy and reliability of the triage process but also aligns with strict audit and governance requirements. By focusing on evidence retrieval and counterfactual checks, the framework provides a more transparent and accountable method for investigators to review and act on alerts.
The research highlights the potential for LLMs to revolutionize AML triage by making the process more efficient and compliant. However, the success of this framework will depend on its ability to integrate seamlessly with existing regulatory systems and gain the trust of financial institutions. Future developments may include real-world testing and refinement based on feedback from AML professionals.