Symbolic picture: Colourbox
Developing software is one thing. But an increasingly important part of the work of programmers is testing the software for errors. Just how important this is was demonstrated recently when a faulty update led to IT system failures worldwide.
Large language models are changing the process of software development. At the University of Passau, researchers are working with practitioners to explore their potential and develop new tools.
AI makes it possible to analyse and improve source code in a way that was previously impossible."
Professor Gordon Fraser, University of Passau
In the DeepCode project, for example, Gordon Fraser, Professor of Software Engineering II, is working with msg systems ag to investigate AI models that are capable of efficiently analysing source code. The project, which is funded by the Bavarian Research Foundation, utilises various deep learning techniques, a branch of artificial intelligence that works with artificial neural networks. The first results are now available.
"AI makes it possible to analyse and improve source code in a way that was previously impossible," explains Professor Fraser. The potential for improving code quality is enormous, but in order to realise this potential, data, models and tools based on them are needed. The project team has made significant progress in this area over the past year - both in the collection of data, the creation of tools to process this data efficiently and the development of various AI models.
The team from the practice tested the developed models in pilot projects - with success: msg systems ag reports that the feedback from the practice was consistently positive, particularly with regard to the quality of the proposed code improvements. The team was able to demonstrate an increase in quality of over 20 per cent by fine-tuning with data from msg. "The results so far have exceeded our expectations. We see great potential for the application of our methods in software development, especially in the quality assurance and maintenance of software," says Florian Kandlinger, who is co-responsible for the project on the msg side.
DeepCode is not the only project in which Professor Fraser is looking for new ways to improve software quality. The professor focuses his research on errors in software development and is pursuing three main approaches: how AI can provide support when writing the source code, how the technology can help to improve the code and how young programmers can be sensitised to the topic of testing. He recently launched a project with which he successfully took part in a competition organised by the Bavarian-Czech Academic Agency. This is also about using fully automated test methods for quality assurance in software development.
Research on large language models at the University of Passau
The projects are part of an interdisciplinary research focus at the University of Passau on the topic of large language models. Researchers from several faculties are investigating the future collaboration between humans and machines as well as the associated ethical and legal issues.
Professor Steffen Herbold, who holds the Chair of AI Engineering at the University of Passau, explains that language models have a disruptive effect, particularly in the field of software development. He predicts that language models will become a normal component of new software systems. The technology will also be able to support software development and reliably solve simple and routine tasks. Florian Lemmerich, Professor of Applied Machine Learning at the University of Passau, is researching how the quality of large language models can be evaluated.
This text was machine-translated from German.
Professor Gordon Fraser
How can we find and prevent software errors?
How can we find and prevent software errors?
Professor Gordon Fraser has held the Chair of Software Engineering II at the University of Passau since 2017. After completing his doctorate at Graz University of Technology, he conducted research at Saarland University and the University of Sheffield. His research and teaching focusses on issues relating to software analysis, software development and the didactics of programming.
