Rybinski, Fabian

The quality of Large Language Model (LLM) outputs depends largely on the structure and design of the input prompts. While Chain-of-Thought (CoT) prompting already shows established improvements, the potential of structured approaches based on formal models remains largely unexplored.

Process models from business process modeling offer established structuring paradigms with clearly defined elements such as activities, gateways, events and data flows. These formal structures could serve as templates for the design of LLM prompts in order to systematically think through and process complex tasks. Different process modeling languages such as BPMN, DMN or Petri nets offer different structuring approaches that could be suitable for different types of tasks.

This thesis will therefore investigate how process models can be used as structuring elements for LLM prompting. The aim is to improve the output quality of LLMs by developing process model-based prompting strategies.

The work will cover the following steps:

Research: First, an initial research is conducted to identify existing approaches in the literature in the field of prompt engineering and structured LLM input.

Analysis: The relevant process model elements and prompting techniques will be analyzed and evaluated for their combinability and suitability for different task types based on selected criteria.

Conception: Based on the analysis, a concept will be developed that enables the systematic transfer of process model structures into prompting templates. Combinations with existing techniques such as CoT will also be developed.

Prototype & evaluation: Finally, a prototype implementation will be developed and various prompting strategies will be evaluated.