This research proposes a hybrid framework that integrates Large Language Models (LLMs) with SMT solvers to enable automated legal analysis for financial regulatory compliance. Our approach encompasses formal representation of financial regulatory frameworks, leveraging LLM's natural language understanding capabilities and SMT solver's logical reasoning power for more accurate and efficient compliance verification.

Through application to actual financial cases, we demonstrate the framework's effectiveness in identifying regulatory violations and proposing minimally-changing compliant solutions. This work represents a successful integration of symbolic AI and neural AI, opening new possibilities for fintech and legal tech domains.

This paper introduces a neuro-symbolic approach for automated financial legal analysis and compliance verification. We combine the powerful language understanding capabilities of LLMs with the logical reasoning power of formal verification tools (SMT solvers) to create a system capable of understanding complex legal texts and identifying compliance issues.

The innovation of this research lies in leveraging LLM's self-correction mechanisms, enabling the system to automatically verify and refine its generated logical representations. Through evaluation on financial regulatory cases, we demonstrate the superiority of our method compared to traditional approaches, earning the ACM SIGSOFT Distinguished Paper Award.

This paper presents an innovative architecture based on agentic AI and symbolic reasoning for automated compliance correction and optimization. We construct a multi-agent system where each agent specializes in specific compliance aspects, collaborating to identify and correct legal violations.

The core innovation of our system is implementing an adaptive correction pipeline capable of generating minimally-changing compliant solutions under multiple constraints. We employ SMT solvers to ensure all generated solutions comply with legal requirements while maintaining practical feasibility. This research is currently under peer review in the TOSEM journal's "Agentic AI in Software" special issue.