SelfAudit: Heterogeneous Multi-Dimensional Reflection for Complex Document Auditing

Complex document auditing requires the integration of information across chapters, verification of multiple rules, and fine-grained reasoning over long contexts. Existing LLM-based methods often struggle with task decomposition, distinguishing different types of errors, and maintaining the quality of intermediate results. We propose SelfAudit, an end-to-end multi-agent framework built around Heterogeneous Multi-Dimensional Confidence Decomposition with Dependency-Aware Error Attribution (HCD-EA). HCD-EA estimates a three-dimensional confidence vector consisting of Retrieval Fidelity, Evidence Support, and Rule Compliance, and uses the minimum of these dimensions to identify the weakest part of the auditing process. If the confidence falls below a predefined threshold, HCD-EA uses the predefined dependency graph, locates the earliest failing node in topological order, and re-executes only the affected sub-pipeline. A heterogeneous verifier provides an independent confidence estimate, reducing the risk of circular self-evaluation. Experiments on ContractNLI, CUAD, and LegalBench-RAG show that SelfAudit outperforms 13 baselines across five categories, achieving a weighted F1 of 0.844 on ContractNLI. Ablation studies demonstrate the contribution of each component, while selective re-execution reduces LLM calls by 17% compared with blind retry.