Feedback during programming is desirable, but its usefulness depends on immediacy and relevance to the task. Unit and regression testing are practices to ensure programmers can obtain feedback on their changes; however, running a large test suite is rarely fast, and only a few results are relevant.

Identifying which tests in a test suite are most relevant to a change helps detect defects earlier during programming and selecting tests that serve as examples to help programmers understand particular code.

In this work, we describe an approach to evaluate how well large language models (LLMs) and embedding models can judge the relevance of a test to a change. We construct a dataset by applying faulty variations of real-world code changes and measuring whether the model could nominate the failing tests beforehand.

We found that, while embedding models perform best on such a task, even simple information retrieval models are surprisingly competitive. In contrast, pre-trained LLMs are of limited use as they focus on confounding aspects like coding styles.

We argue that the high computational cost of AI models is not always justified, and tool developers should also consider non-AI models for code-related retrieval and recommendation tasks. Lastly, we generalize from unit tests to live examples and outline how our approach can benefit live programming environments.