In the past years, we have seen an explosion of Machine Learning techniques applied to software engineering tasks. The vast majority of these approaches is applied to program code, using Large Language Models (LLMs) to predict token sequences in specific contexts. However, the dynamic nature of programs is hardly exploited; on the contrary, interpreting and predicting the semantics of code remains a big challenge for LLMs.

In this paper, I follow a very different path. I train machine learning models from the dynamic behavior of programs, thereby obtaining program-specific models that can predict how the specific software behaves. Specifically, we learn how synthesized changes to code (mutations) lead to changes in behavior (notably, the program output). The resulting models can then predict how a given change affects the program behavior, but also where which change may be needed for a desired change in behavior. When a programmer says: “I want this button to be green”, the model will suggest where and how to make a code change.

In the long run, having models automatically learn the relationship between code and behavior will open new paths in automating programming tasks such as debugging, software maintenance, or program understanding.