A new paradigm requires new tools and methods.
There is a lack of appropriate development methods and tools. Even though there are many libraries available for training models, e.g. TensorFlow, their correct use poses great challenges for developers without deep knowledge of the methods encoded in them, and it requires automation. At the same time, tools that are usually used in software development for the analysis of software correctness are hardly available. Therefore, the group is researching new methods and tools that support developers in meeting these challenges and help to ensure that AI can be used by a larger circle of developers and transferred to a broader set of applications.
The second object of our research is the development of new software engineering methods to enhance the AI itself. For example, we are exploring ways to introduce ideas of reactive programming and modular programming into the data processing pipeline powering the learning of AI models. Respectively, we want to explore new programming languages to support the third wave of AI that integrates learned, modeled and built in knowledge and cognitive models to achieve human-like properties. For example, by continuously extending existing knowledge and “thinking” logically. The von Neumann machine model on which today’s programming is based describes calculations on a level too low to be suitable for the complexity of third-wave AI. However, a new suitable model has not yet been systematically researched. Possibly, the same architectures that are used for the pipeline design can be applied to the composition of different components of third wave AI.
Courses:
Recent Papers:
- Where Do LLMs Still Struggle? An In-Depth Analysis of Code Generation Benchmarks. AM Sharifloo, M Heydari, P Kazerooni, D Maninger, M Mezini.
- A critical study of what code-llms (do not) learn. A Anand, S Verma, K Narasimhan, M Mezini.
- Integrating Symbolic Execution into the Fine-Tuning of Code-Generating LLMs. M Sakharova, A Anand, M Mezini.