Structural and Algorithmic Limitations In Llm-Driven Bpmn Generation: A Rapid Review of Empirical Evidence

Business Process Management (BPM) demands formalized models, yet manual BPMN (Business Process Model and Notation) construction remains labor-intensive and error-prone. Large Language Models (LLMs) present opportunities for automated text-to-model translation, though their effectiveness remains underexplored. This rapid review synthesizes 11 empirical investigations published from 2023 to 2025, revealing a nascent field dominated by GPT-4 (55% adoption) and prompt-based methodologies (64% prevalence). Our analysis uncovers critical structural deficiencies: LLMs demonstrate fundamental incapacity for graph-theoretic reasoning and set-theoretic spatial optimization required for valid BPMN layouts. Specifically, 55% of surveyed approaches fail with non-trivial process complexity, while 64% lack reproducible artifacts. All testable implementations exhibit severe degradation when confronted with hierarchical structures, parallel gateways, or spatial coordinate systems inherent to BPMN Diagram Interchange specifications. We conclude that current approaches conflate linguistic pattern matching with algorithmic reasoning—a categorical error that mandates hybrid architectures combining LLM semantic extraction with specialized graph layout algorithms. The field urgently requires standardized benchmarks, rigorous graph-theoretic validation metrics, and transparent reproducibility to transcend proof-of-concept limitations.