4 July 2025 in IEEE Transactions on Software Engineering
Large Language Models (LLMs) have been recently exploited as judges for complex natural language processing tasks, such as Q&A (Question & Answer). The basic idea is to delegate to an LLM the assessment of the “quality” of the output provided by an automated technique (often another LLM) for tasks for which: (i) quantitative metrics would only tell part of the story, and; (ii) a large-scale human-based evaluation would be too expensive. LLMs-as-a-judge, if proven effective for a specific task, can also unlock new possibilities for automation, with several LLMs proposing a solution for a given instance of the task (e.g., an answer to a question) and others judging and deciding what is the best output to show the user. We study the effectiveness of LLMs-as-a-judge for two code-related tasks, namely code generation and code summarization. The rationale for choosing these tasks is two-fold. First, quantitative metrics are usually not enough for the assessment of code summarizers/generators. For example, it is well documented that metrics such as BLEU are quite weak proxies for the quality of the generated summaries. Second, even state-of-the-art techniques still struggle with handling complex instances of these tasks (e.g., summarizing a quite long / complex function), making them good candidates for benefiting from more advanced solutions envisioning collaboration among LLMs. For code generation, we check whether eight LLMs are able to judge the correctness of 1,405 Java methods and 1,281 Python functions generated by the same LLMs or implemented by humans. For code summarization, we compare the judgment of five LLMs to those provided by nine humans for ç1.2k summaries, related to both Java and Python functions. Our findings show that GPT-4-turbo is the best LLM in terms of judging capabilities for both tasks, with “smaller” LLMs featuring tens of billions parameters not being able to cope with judging tasks. However, even the best-performing LLM frequently misjudges the correctness of the code and summary quality.
21 March 2024 in IEEE Transactions on Software Engineering
Neural Language Models of Code, or Neural Code Models (NCMs), are rapidly progressing from research prototypes to commercial developer tools. As such, understanding the capabilities and limitations of such models is becoming critical. However, the abilities of these models are typically measured using automated metrics that often only reveal a portion of their real-world performance. While, in general, the performance of NCMs appears promising, currently much is unknown about how such models arrive at decisions. To this end, this paper introduces docode , a post hoc interpretability method specific to NCMs that is capable of explaining model predictions. docode is based upon causal inference to enable programming language-oriented explanations. While the theoretical underpinnings of docode are extensible to exploring different model properties, we provide a concrete instantiation that aims to mitigate the impact of spurious correlations by grounding explanations of model behavior in properties of programming languages. To demonstrate the practical benefit of docode , we illustrate the insights that our framework can provide by performing a case study on two popular deep learning architectures and ten NCMs. The results of this case study illustrate that our studied NCMs are sensitive to changes in code syntax. All our NCMs, except for the BERT-like model, statistically learn to predict tokens related to blocks of code (e.g., brackets, parenthesis, semicolon) with less confounding bias as compared to other programming language constructs. These insights demonstrate the potential of docode as a useful method to detect and facilitate the elimination of confounding bias in NCMs.