Title: AuditFlow: Executable Symbolic Environments for Structured Financial Reporting Verification

Abstract:

Verifying structured financial audits presents a significant challenge for language-model agents, as accuracy relies not merely on textual analysis but on the integration of structured evidence. To succeed, a model must map reported facts to specific taxonomy concepts, navigate calculation or dimensional relationships, and recalculate expected values prior to applying audit rules. In response, we introduce AuditFlow, a multi-agent framework grounded in graph theory that distinguishes between adaptive search processes and deterministic verification.

AuditFlow constructs a symbolic environment by combining a static US-GAAP taxonomy graph with a dynamic XBRL filing graph. This environment is accessible via typed tools designed for fact retrieval, taxonomy traversal, numerical validation, and rule evaluation. The system employs a hierarchical review process: two junior auditors examine each case independently from regulatory and evidentiary perspectives, while a senior auditor intervenes to resolve conflicts and may initiate further investigations. Final reports are synthesized through evidential aggregation to generate an audit verdict, expected values, an evidence trail, and a trustworthiness score.

In evaluations using a FinMR sample derived from FinAuditing under GPT-5.5, AuditFlow achieved a joint audit accuracy of 82.09%, surpassing the strongest baseline by 14.93 points. The critical role of the symbolic environment was highlighted when removing deterministic checks caused accuracy to plummet to 17.91%, demonstrating that this component performs verification tasks that the model cannot reliably execute on its own.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC