AI-BASED AUTOMATION OF DECISION LOGIC REPRESENTATION: BRIDGING GAP IN AUTOMATED DECISION MODELING VALIDATION

Abstract

The study aims to resolve the "modeling bottleneck" in Business Process Management by developing an automated method for ensuring the correctness of Decision Model and Notation (DMN) tables generated by Large Language Models (LLMs). The primary goal is to determine whether shifting the AI's role from a pure generator to a validator within a closed-loop system can overcome the structural limitations inherent in stochastic models.

The research employs a comparative experimental design using a "Mutation Testing" approach. We analyze two distinct workflows: (1) Static Generation, where test cases are fixed, and (2) Dynamic Paired Generation, where the LLM regenerates both the decision logic (DMN XML) and the validation criteria (Test Cases JSON) simultaneously upon failure. The methodology integrates concepts from "Struc-Bench" for structural analysis and utilizes a deterministic DMN engine (Camunda) for execution-based verification.

The experiments demonstrate that "out-of-the-box" LLM generation fails in approximately 4.5% of cases due to semantic drift and structural hallucinations when validated against static benchmarks. However, the proposed "Dynamic Paired Generation" workflow achieved a 100% convergence rate across 200 cycles. The system successfully identified and corrected both syntactic errors (XML schema violations) and logical errors (Hit Policy violations) without human intervention.

The study introduces the concept of "Dynamic Paired Generation" for neuro-symbolic systems. Unlike traditional "Chain-of-Thought" prompting, this approach leverages the mutual consistency between two independent structural representations (Logic and Examples) to filter out hallucinations, proving that dynamic validation is superior to static prompting for structured data tasks.

The proposed framework provides a blueprint for "Self-Healing" decision management systems. It allows non-technical business analysts to convert natural language policies into executable, error-free DMN models, significantly reducing the time and cost of regulatory compliance and process automation.