Research Seminar by LIN Yun | SE for (AI for SE)

SE for (AI for SE)
 

Language models are transforming all aspects of the software development process, including code generation, editing, testing, and debugging. When developing and training code language models, there is a need for a unified process involving data collection and cleaning, model training, updating, and adjustment. Therefore, we anticipate that analysis techniques for model training will emerge as a new layer of infrastructure to answer questions such as: "How did the model make a specific prediction?", "How was the model's prediction gradually learned during training?", and "What actions can we take to modify and improve a particular prediction?". This talk presents our work on code editing—a project-wide, interactive editing approach targeted at generation—as a case study to demonstrate how we analyze models, trace training samples, and construct a digital twin environment to test code editing models. The talk concludes with a discussion on how AI is reshaping the paradigms of software development.