Multi-Agent Framework for Audit Risk Assessment with Explicit Uncertainty and Evidence Conflict Modeling
Researchers have developed UMAR, a novel multi-agent framework designed to improve audit risk assessment by explicitly modeling uncertainty and evidence conflict. UMAR utilizes three specialized agents—MD&A Text Agent, Financial Ratio Agent, and CAM Agent—to generate independent risk scores with calibrated uncertainty. These scores are then aggregated using Dempster-Shafer theory, which quantifies inter-agent disagreement. Evaluations on 3,200 firm-year observations from SEC 10-K filings demonstrated UMAR's superior performance over baseline models, achieving an AUROC of 0.782 and a PR-AUC of 0.341, while also providing actionable insights through conflict patterns. AI
IMPACT This framework could improve the accuracy and interpretability of financial risk assessments by leveraging multi-agent systems and uncertainty quantification.