Ambiguity Detection and Elimination in Automated Executable Process Modeling

Researchers have developed a framework to detect and eliminate ambiguities in natural-language specifications converted to executable BPMN process models by large language models. The method identifies behavioral inconsistencies through KPI analysis, diagnoses gateway logic problems, and repairs source text through evidence-based refinement, reducing variability in regenerated model behavior.

This research addresses a critical vulnerability in AI-driven automation: the gap between human intent and machine execution. When large language models generate executable business process models from natural-language text, subtle ambiguities in source specifications can produce structurally valid BPMN models with fundamentally different runtime behaviors. This divergence creates operational risk in mission-critical domains like healthcare, where process variability directly impacts outcomes.

The framework tackles this problem through an iterative diagnosis-and-repair loop. Rather than attempting impossible semantic validation against non-existent ground truth, the method empirically detects behavioral instability by regenerating and simulating models multiple times, then analyzing key performance indicator distributions. When inconsistencies emerge, model-based diagnosis techniques pinpoint the exact gateway logic causing divergence and trace it back to specific narrative segments in the original specification.

This approach has immediate implications for enterprise automation initiatives relying on LLM-generated workflows. Organizations deploying AI to convert process documentation into executable models face hidden risks if the source material remains ambiguous. Healthcare applications, financial compliance workflows, and supply chain processes all depend on consistent execution. The research demonstrates this framework reduces behavioral variability in diabetic nephropathy guidance policies, suggesting broad applicability across regulated industries.

Looking forward, this work establishes a pattern for human-in-the-loop AI validation in mission-critical automation. Rather than trusting LLM outputs blindly, organizations can implement feedback loops that detect instability and systematically improve source specifications. The closed-loop validation approach could become standard practice as enterprises integrate generative AI deeper into process automation infrastructure.

• →LLM-generated BPMN models from ambiguous specifications produce structurally valid but behaviorally inconsistent results across repeated executions.
• →The framework detects behavioral inconsistency through empirical KPI distribution analysis rather than attempting semantic correctness proofs.
• →Model-based diagnosis maps divergent gateway logic back to specific narrative segments for targeted text repair.
• →Testing on healthcare process policies demonstrates the method reduces behavioral variability in regenerated models.
• →This establishes a practical validation pattern for enterprise automation systems relying on AI-generated executable specifications.