REGREACT: Self-Correcting Multi-Agent Pipelines for Structured Regulatory Information Extraction

A team of researchers has introduced RegReAct, a novel multi-agent AI framework designed to solve the persistent challenge of extracting structured, machine-readable compliance criteria from complex regulatory documents. Unlike single-pass language models like GPT-4o, which often hallucinate structural elements and fail to capture hierarchical relationships, RegReAct decomposes the task into seven specialized stages. Each stage employs an Observe-Diagnose-Repair (ODR) loop that continuously validates the AI's outputs against the original source text. This self-correcting mechanism is powerful enough to identify and fix not only the model's own hallucinations but also cross-reference errors present within the regulations themselves.

To ensure structural accuracy, RegReAct constructs a typed criterion graph that preserves the hierarchical nature of legal requirements. For completeness, it actively resolves external dependencies by retrieving, summarizing, and embedding referenced legal content directly into its outputs, creating self-contained results. In a practical evaluation, the team applied RegReAct to three EU Taxonomy Delegated Acts, successfully constructing a comprehensive dataset comprising 242 regulatory activities with over 4,800 hierarchical criteria, thresholds, and enriched source summaries. The system demonstrably outperformed a GPT-4o single-pass baseline across all structural and semantic evaluation metrics, proving the efficacy of its multi-agent, self-correcting approach for high-stakes document analysis.

The researchers have committed to making the code and the resulting dataset publicly available, providing a valuable resource for both the AI research community and professionals in legal tech, compliance, and regulatory analysis. This work represents a significant step toward reliable, automated processing of dense legal and regulatory texts, where accuracy and traceability are paramount.