Using large language models to generate executable BPMN models based on text descriptions: an overview of approaches, limitations, and validation methods
TL;DR
It is argued that edge (sequence flow) generation is the weakest link once nodes are fixed, and typical structural failure modes (dangling nodes, disconnects, gateway violations, etc.) and causes tied to autoregressive generation are summarized.
Abstract
The paper addresses the use of large language models (LLMs) to automatically generate executable business processes in BPMN 2.0 from unstructured natural-language descriptions, with deployment to process engines such as Camunda Platform in mind. The text-to-BPMN XML mapping task is stated and decomposed into subproblems: extracting activities, events, and gateways; recovering control flow and branching; ensuring valid sequence flows and conformance to the BPMN specification. We survey process representations (BPMN XML, JSON as an intermediate format, graph-based models) and LLM adaptation methods: prompt engineering, instruction tuning, and fine-tuning. We argue that edge (sequence flow) generation is the weakest link once nodes are fixed, and summarize typical structural failure modes (dangling nodes, disconnects, gateway violations, etc.) and causes tied to autoregressive generation. A staged pipeline is proposed—separate generation of node set V and edge set E followed by post-validation — together with a three-level validation scheme: syntactic (BPMN XSD), structural (graph invariants), and executable (Camunda deploy and run). The article outlines a feedback-enabled pipeline architecture and discusses applicability and limitations.