Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BPMN Assistant: An LLM-Based Approach to Business Process Modeling

Josip Tomo Licardo, Nikola Tankovic, Darko Etinger

TL;DR

BPMN Assistant tackles the difficulty of editing BPMN diagrams with large language models by decoupling process logic from BPMN XML syntax through a structured JSON intermediate representation. This enables atomic editing operations via function calls and a validation loop to ensure soundness before XML generation. Across multiple models, the JSON-based approach outperforms direct XML regeneration in editing tasks, with substantial latency reductions of approximately $43\%$ and output-token reductions over $75\%$, while requiring more input context. Notably, open-weight models like DeepSeek V3 show meaningful gains (up to about $50\%$ editing success) when using the JSON workflow, highlighting implications for on-prem and privacy-preserving deployments. The work demonstrates a practical path toward reliable, interactive BPMN modeling with constrained LLM interactions and structured representations that generalize beyond proprietary systems.

Abstract

This paper presents BPMN Assistant, a tool that leverages Large Language Models for natural language-based creation and editing of BPMN diagrams. While direct XML generation is common, it is verbose, slow, and prone to syntax errors during complex modifications. We introduce a specialized JSON-based intermediate representation designed to facilitate atomic editing operations through function calling. We evaluate our approach against direct XML manipulation using a suite of state-of-the-art models, including GPT-5.1, Claude 4.5 Sonnet, and DeepSeek V3. Results demonstrate that the JSON-based approach significantly outperforms direct XML in editing tasks, achieving higher or equivalent success rates across all evaluated models. Furthermore, despite requiring more input context, our approach reduces generation latency by approximately 43% and output token count by over 75%, offering a more reliable and responsive solution for interactive process modeling.

BPMN Assistant: An LLM-Based Approach to Business Process Modeling

TL;DR

BPMN Assistant tackles the difficulty of editing BPMN diagrams with large language models by decoupling process logic from BPMN XML syntax through a structured JSON intermediate representation. This enables atomic editing operations via function calls and a validation loop to ensure soundness before XML generation. Across multiple models, the JSON-based approach outperforms direct XML regeneration in editing tasks, with substantial latency reductions of approximately and output-token reductions over , while requiring more input context. Notably, open-weight models like DeepSeek V3 show meaningful gains (up to about editing success) when using the JSON workflow, highlighting implications for on-prem and privacy-preserving deployments. The work demonstrates a practical path toward reliable, interactive BPMN modeling with constrained LLM interactions and structured representations that generalize beyond proprietary systems.

Abstract

This paper presents BPMN Assistant, a tool that leverages Large Language Models for natural language-based creation and editing of BPMN diagrams. While direct XML generation is common, it is verbose, slow, and prone to syntax errors during complex modifications. We introduce a specialized JSON-based intermediate representation designed to facilitate atomic editing operations through function calling. We evaluate our approach against direct XML manipulation using a suite of state-of-the-art models, including GPT-5.1, Claude 4.5 Sonnet, and DeepSeek V3. Results demonstrate that the JSON-based approach significantly outperforms direct XML in editing tasks, achieving higher or equivalent success rates across all evaluated models. Furthermore, despite requiring more input context, our approach reduces generation latency by approximately 43% and output token count by over 75%, offering a more reliable and responsive solution for interactive process modeling.

Paper Structure

This paper contains 34 sections, 2 equations, 5 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Research Objectives
  3. Contribution
  4. Related Work
  5. Process Extraction and Prompting Strategies
  6. Interactive Modeling Tools
  7. Automated Generation and Evaluation Frameworks
  8. System Architecture
  9. Backend
  10. Model Integration and Selection
  11. Supported BPMN Elements
  12. BPMN Layout Server
  13. Frontend
  14. Data Flow
  15. JSON Representation of BPMN Diagrams
  16. ...and 19 more sections

Figures (5)

  • Figure 1: BPMN Assistant system: (a) Component architecture showing key system elements and (b) request sequence flow showing the temporal order of interactions.
  • Figure 2: The web application interface featuring a dual-panel design: a chat interface on the left and a BPMN canvas on the right.
  • Figure 3: Latency comparison between JSON and XML-based BPMN generation.
  • Figure 4: Token usage comparison for input and output tokens in JSON and XML-based BPMN generation. (a) shows the higher input token requirement for JSON prompts. (b) shows the significantly more concise output generated by the JSON approach.
  • Figure 5: Success rate comparison for diagram editing tasks between JSON and XML-based BPMN representations.