Benchmarking Causal Study to Interpret Large Language Models for Source Code
This work addresses the problem of interpretability and confounding bias in LLM evaluations for software engineering, offering a more rigorous benchmarking approach, though it is incremental in applying causal methods to an existing domain.
The paper tackles the lack of causal inference in evaluating Large Language Models for source code tasks by introducing a benchmarking strategy called Galeras, which includes curated testbeds for three software engineering tasks, and demonstrates through a case study on ChatGPT that prompt semantics has a positive causal effect on performance with an average treatment effect of ≈3% and identifies confounders like prompt size with correlations of ≈0.412%.
One of the most common solutions adopted by software researchers to address code generation is by training Large Language Models (LLMs) on massive amounts of source code. Although a number of studies have shown that LLMs have been effectively evaluated on popular accuracy metrics (e.g., BLEU, CodeBleu), previous research has largely overlooked the role of Causal Inference as a fundamental component of the interpretability of LLMs' performance. Existing benchmarks and datasets are meant to highlight the difference between the expected and the generated outcome, but do not take into account confounding variables (e.g., lines of code, prompt size) that equally influence the accuracy metrics. The fact remains that, when dealing with generative software tasks by LLMs, no benchmark is available to tell researchers how to quantify neither the causal effect of SE-based treatments nor the correlation of confounders to the model's performance. In an effort to bring statistical rigor to the evaluation of LLMs, this paper introduces a benchmarking strategy named Galeras comprised of curated testbeds for three SE tasks (i.e., code completion, code summarization, and commit generation) to help aid the interpretation of LLMs' performance. We illustrate the insights of our benchmarking strategy by conducting a case study on the performance of ChatGPT under distinct prompt engineering methods. The results of the case study demonstrate the positive causal influence of prompt semantics on ChatGPT's generative performance by an average treatment effect of $\approx 3\%$. Moreover, it was found that confounders such as prompt size are highly correlated with accuracy metrics ($\approx 0.412\%$). The end result of our case study is to showcase causal inference evaluations, in practice, to reduce confounding bias. By reducing the bias, we offer an interpretable solution for the accuracy metric under analysis.