This paper delves into the intricacies of code summarization using advanced transformer-based language models. Through empirical studies, we evaluate the efficacy of code summarization by altering function and variable names to explore whether models truly understand code semantics or merely rely on textual cues. We have also introduced adversaries like dead code and commented code across three programming languages (Python, Javascript, and Java) to further scrutinize the model’s understanding. Ultimately, our research aims to offer valuable insights into the inner workings of transformer-based LMs, enhancing their ability to understand code and contributing to more efficient software development practices and maintenance workflows.

Fine-tuning all the layers of a pre-trained neural language encoder (either using all the parameters or using parameter-efficient methods) is often the de-facto way of adapting it to a new task. We show evidence that for different downstream language tasks, fine-tuning only a subset of layers is sufficient to obtain performance that is close to and often better than fine-tuning all the layers in the language encoder. We propose an efficient metric based on the diagonal of the Fisher information matrix (FIM score), to select the candidate layers for selective fine-tuning. We show, empirically on GLUE and SuperGLUE tasks and across distinct language encoders, that this metric can effectively select layers leading to a strong downstream performance. Our work highlights that task-specific information corresponding to a given downstream task is often localized within a few layers, and tuning only those is sufficient for strong performance. Additionally, we demonstrate the robustness of the FIM score to rank layers in a manner that remains constant during the optimization process.