Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AuditAgent: Expert-Guided Multi-Agent Reasoning for Cross-Document Fraudulent Evidence Discovery

Songran Bai, Bingzhe Wu, Yiwei Zhang, Chengke Wu, Xiaolong Zheng, Yaze Yuan, Ke Wu, Jianqiang Li

TL;DR

AuditAgent tackles cross-document financial fraud evidence localization by embedding auditing-domain priors into a multi-agent reasoning framework. It combines a variational Bayesian prior model, a prior-guided multi-route retrieval, and a two-phase multi-expert reasoning pipeline to locate and assemble fraud evidence across lengthy financial disclosures. Experiments on FinFraud-Real, a CSRC-based dataset, show substantial improvements in recall and interpretability over general-purpose agents and single LLMs, with ablations highlighting the value of subject priors and domain-specific models. The work demonstrates the practical impact of integrating domain knowledge and structured priors for robust, transparent automated financial forensics in real regulatory contexts.

Abstract

Financial fraud detection in real-world scenarios presents significant challenges due to the subtlety and dispersion of evidence across complex, multi-year financial disclosures. In this work, we introduce a novel multi-agent reasoning framework AuditAgent, enhanced with auditing domain expertise, for fine-grained evidence chain localization in financial fraud cases. Leveraging an expert-annotated dataset constructed from enforcement documents and financial reports released by the China Securities Regulatory Commission, our approach integrates subject-level risk priors, a hybrid retrieval strategy, and specialized agent modules to efficiently identify and aggregate cross-report evidence. Extensive experiments demonstrate that our method substantially outperforms General-Purpose Agent paradigm in both recall and interpretability, establishing a new benchmark for automated, transparent financial forensics. Our results highlight the value of domain-specific reasoning and dataset construction for advancing robust financial fraud detection in practical, real-world regulatory applications.

AuditAgent: Expert-Guided Multi-Agent Reasoning for Cross-Document Fraudulent Evidence Discovery

TL;DR

AuditAgent tackles cross-document financial fraud evidence localization by embedding auditing-domain priors into a multi-agent reasoning framework. It combines a variational Bayesian prior model, a prior-guided multi-route retrieval, and a two-phase multi-expert reasoning pipeline to locate and assemble fraud evidence across lengthy financial disclosures. Experiments on FinFraud-Real, a CSRC-based dataset, show substantial improvements in recall and interpretability over general-purpose agents and single LLMs, with ablations highlighting the value of subject priors and domain-specific models. The work demonstrates the practical impact of integrating domain knowledge and structured priors for robust, transparent automated financial forensics in real regulatory contexts.

Abstract

Financial fraud detection in real-world scenarios presents significant challenges due to the subtlety and dispersion of evidence across complex, multi-year financial disclosures. In this work, we introduce a novel multi-agent reasoning framework AuditAgent, enhanced with auditing domain expertise, for fine-grained evidence chain localization in financial fraud cases. Leveraging an expert-annotated dataset constructed from enforcement documents and financial reports released by the China Securities Regulatory Commission, our approach integrates subject-level risk priors, a hybrid retrieval strategy, and specialized agent modules to efficiently identify and aggregate cross-report evidence. Extensive experiments demonstrate that our method substantially outperforms General-Purpose Agent paradigm in both recall and interpretability, establishing a new benchmark for automated, transparent financial forensics. Our results highlight the value of domain-specific reasoning and dataset construction for advancing robust financial fraud detection in practical, real-world regulatory applications.

Paper Structure

This paper contains 17 sections, 2 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Machine Learning for Financial Statement Fraud Detection
  4. Financial Domain Benchmarks
  5. LLMs-based Long Document Understanding
  6. Data and Task: FinFraud-Real
  7. Methodology: AuditAgent
  8. Variational Bayesian-Based Subject Prior Modeling
  9. Prior-Guided Multi-Route Retrieval
  10. Multi-Expert Reasoning for Evidence Generation
  11. Experiments
  12. Experimental Setup
  13. Results
  14. Overall Performance
  15. Effectiveness of Subject Prior Modeling
  16. ...and 2 more sections

Figures (4)

  • Figure 1: AuditAgent consists of three key stages --- (1) Variational Bayesian-Based Subject Prior Modeling; (2) Prior-Guided Multi-Route Retrieval; (3) Multi-Expert Reasoning for Evidence Generation.
  • Figure 2: Performance breakdown over cases with different number of input tokens.
  • Figure 3: Fraudulent prior distribution over historical A-share data obtained by Subject Prior Modeling.
  • Figure 4: Case Study to show (1) The superiority of AuditAgent comparing with General-Purpose Agent; (2) The superiority of r1-like reasoning models for analyzing evidences.