Several FONDA subprojects perform research close to the interface between scientists and workflow systems. Recently, Large Language Models (LLMs) have proven helpful in supporting such programming interfaces by taking over parts of the necessary coding after adequate prompting by the developer. In an explorative project during phase I, we already showed that general-purpose LLMs, such as ChatGPT, empowers users to explain the intention of DAWs using workflow definition languages.
In this team, we will further refine and evaluate strategies for using LLMs for different tasks at the human-computer interface to workflow systems. The concrete study cases emerge from the subprojects that participate in the team and will include tasks such as DAW design, adaptation, debugging, porting, or understanding. By gathering the competencies from the subprojects, we will be able to jointly research new strategies for using LLMs in such settings, for instance, by fine-tuning models with task-specific examples, by specialized prompting strategies, or by re-using models generated for other programming purposes (e.g., Github Copilot).