Pre-trained language models (PLMs) have demonstrated remarkable abilities in coding tasks, establishing themselves as a state-of-the-art technique in machine learning for code. However, due to their deep neural network-based structure, PLMs function as black-box systems, making it crucial to understand the types of information they actually learn. Recent studies indicate that PLMs possess cross-lingual capabilities, allowing them to generalize to unseen programming languages and outperform monolingual models when trained in a multilingual setting. Nonetheless, the reasons behind these cross-lingual abilities remain largely uncharted and remain open questions. In this paper, we explore this phenomenon through a syntactic perspective. Specifically, we build on our prior work, the AST-Probe, a probing methodology that evaluates whether a PLM encodes the complete grammatical structure of a programming language. This probe identifies a syntactic subspace within the PLM’s vector representations, which is then used to reconstruct ASTs. We extend this approach in two ways. First, we conducted experiments on eight programming languages and eight PLMs and found that: (1) this syntactic structure can be extracted in all cases, (2) CodeBERT and GraphCodeBERT excel at encoding ASTs, and (3) syntactic knowledge resides in the middle layers of all PLMs, with a distribution that is independent of the programming language. Secondly, we mathematically adapt the AST-Probe to a multilingual setting and apply it to CodeBERT. Our findings provide evidence that CodeBERT learns cross-lingual representations of programming languages syntax.

Evaluating the alignment of large language models (LLMs) with user-defined coding preferences is a challenging endeavor that requires a deep assessment of LLMs’ outputs. Existing methods and benchmarks rely primarily on automated metrics and static analysis tools, which often fail to capture the nuances of user instructions and LLM outputs. To address this gap, we introduce the LLM-as-a-Judge evaluation framework and present CodeUltraFeedback, a comprehensive dataset for assessing and improving LLM alignment with coding preferences. CodeUltraFeedback consists of 10,000 coding instructions, each annotated with four responses generated from a diverse pool of 14 LLMs. These responses are annotated using GPT-3.5 as a judge, with both ranking-based scores and detailed textual feedback across five distinct coding preferences. Our analysis reveals that responses from GPT-3.5 and GPT-4 are consistently rated higher than those from open-weight models, underscoring substantial alignment gaps between closed- and open-weight LLMs. In turn, we explore the usage of CodeUltraFeedback as feedback data to fine-tune and align CodeLlama-7B-Instruct using supervised fine-tuning (SFT) and reinforcement learning from AI feedback (RLAIF) with direct preference optimization (DPO). The resulting aligned model achieves an average alignment improvement of 22.7% and 29.7% when evaluated with GPT-3.5 and GPT-4 judges, respectively. Notably, our aligned CodeLlama-7B-Instruct surpasses much larger models, such as CodeLlama-13B and 34B, in alignment with coding preferences. Despite not being explicitly trained for functional correctness, it also achieves a 10.5% and 26.6% relative improvement in Pass@1 and Pass@10 on the HumanEval+ benchmark. Our contributions demonstrate the practical value of preference tuning in code generation and set the stage for further progress in model alignment and RLAIF for automated software engineering.

Large Language Models (LLMs) and other automated techniques have been increasingly used to support software developers by generating software artifacts such as code snippets, patches, and comments. However, accurately assessing the correctness of these generated artifacts remains a significant challenge. On one hand, human evaluation provides high accuracy but is labor-intensive and lacks scalability. On the other hand, many automatic evaluation metrics are scalable and require minimal human effort, but they often fail to accurately reflect the actual correctness of generated software artifacts.In this paper, we present SE-Jury, the first evaluation metric for LLM-as-Ensemble-Judge specifically designed to accurately assess the correctness of generated software artifacts. SE-Jury first defines five distinct evaluation strategies, each implemented as an independent judge. A dynamic team selection mechanism then identifies the most appropriate subset of judges as a team to produce a final correctness score through ensembling. We evaluate SE-Jury across a diverse set of software engineering (SE) benchmarks that span three popular SE tasks: code generation, automated program repair, and code summarization. Results demonstrate that SE-Jury consistently achieves a higher correlation with human judgments, with improvements ranging from 29.6%–140.8% over existing automatic metrics. SE-Jury reaches agreement levels with human annotators that are close to inter-annotator agreement in code generation and program repair. These findings underscore SE-Jury’s potential as a scalable and reliable alternative to human evaluation in these SE tasks.

In code review, generating structured and relevant comments is crucial for identifying code issues and facilitating accurate code changes that ensure an efficient code review process. Well-crafted comments not only streamline the code review itself but are also essential for subsequent tasks like code refinement, where the code is modified to satisfy the input review comment. Although various AI-based approaches aimed to automate comment generation, their effectiveness remains limited by the quality of the training data. Existing code review datasets are often noisy and unrefined, posing limitations to the learning potential of AI models and hindering the automation process. To address these challenges, we propose a curation pipeline designed to enhance the quality of the largest publicly available code review dataset. We begin by establishing an evaluation framework, incorporating specific criteria and categories to empirically study the initial quality of the dataset. Using a large language model (LLM)-driven approach, we then apply our curation pipeline to refine the dataset. A comparative analysis of the newly curated dataset, based on the same evaluation framework, demonstrates substantial improvements in the clarity and conciseness of the comments. Additionally, we assess the impact of the curated dataset on automating downstream tasks, specifically comment generation and code refinement. Our findings show that the curated dataset leads to enhanced model performance in generating more accurate comments. Curated comments are also more useful as they lead to more accurate code refinement.

Large language models (LLMs) demonstrate impressive capabilities to generate accurate code snippets given natural language intents in a zero-shot manner, i.e., without the need for specific fine-tuning. While prior studies have highlighted the advantages of fine-tuning LLMs, this process incurs high computational costs, making it impractical in resource-scarce environments, particularly for models with billions of parameters. To address these challenges, previous research explored in-context learning (ICL) and retrieval-augmented generation (RAG) as strategies to guide the LLM generative process with task-specific prompt examples. However, ICL and RAG introduce inconveniences, such as the need for designing contextually relevant prompts and the absence of learning task-specific parameters, thereby limiting downstream task performance. In this context, we foresee parameter-efficient fine-tuning (PEFT) as a promising approach to efficiently specialize LLMs to task-specific data while maintaining reasonable resource consumption. In this article, we deliver a comprehensive study of PEFT techniques for LLMs in the context of automated code generation. Our comprehensive investigation of PEFT techniques for LLMs reveals their superiority and potential over ICL and RAG across a diverse set of LLMs and three representative Python code generation datasets: Conala, CodeAlpacaPy, and APPS. Furthermore, our study highlights the potential for tuning larger LLMs and significant reductions in memory usage by combining PEFT with quantization. Therefore, this study opens opportunities for broader applications of PEFT in software engineering scenarios.

Motivated by recent work on lifelong learning applications for language models (LMs) of code, we introduce CodeLL, a lifelong learning dataset focused on code changes. Our contribution addresses a notable research gap marked by the absence of a long-term temporal dimension in existing code change datasets, limiting their suitability in lifelong learning scenarios. In contrast, our dataset aims to comprehensively capture code changes across the entire release history of open-source software repositories. In this work, we introduce an initial version of CodeLL, comprising 71 machine-learning-based projects mined from Software Heritage. This dataset enables the extraction and in-depth analysis of code changes spanning 2,483 releases at both the method and API levels. CodeLL enables researchers studying the behaviour of LMs in lifelong fine-tuning settings for learning code changes. Additionally, the dataset can help studying data distribution shifts within software repositories and the evolution of API usages over time.

Pre-trained language models (PLMs) have become a prevalent technique in deep learning for code, utilizing a two-stage pre-training and fine-tuning procedure to acquire general knowledge about code and specialize in a variety of downstream tasks. However, the dynamic nature of software codebases poses a challenge to the effectiveness and robustness of PLMs. In particular, world-realistic scenarios potentially lead to significant differences between the distribution of the pre-training and test data, i.e., distribution shift, resulting in a degradation of the PLM's performance on downstream tasks. In this paper, we stress the need for adapting PLMs of code to software data whose distribution changes over time, a crucial problem that has been overlooked in previous works. The motivation of this work is to consider the PLM in a non-stationary environment, where fine-tuning data evolves over time according to a software evolution scenario. Specifically, we design a scenario where the model needs to learn from a stream of programs containing new, unseen APIs over time. We study two widely used PLM architectures, i.e., a GPT2 decoder and a RoBERTa encoder, on two downstream tasks, API call and API usage prediction. We demonstrate that the most commonly used fine-tuning technique from prior work is not robust enough to handle the dynamic nature of APIs, leading to the loss of previously acquired knowledge i.e., catastrophic forgetting. To address these issues, we implement five continual learning approaches, including replay-based and regularization-based methods. Our findings demonstrate that utilizing these straightforward methods effectively mitigates catastrophic forgetting in PLMs across both downstream tasks while achieving comparable or superior performance.

The objective of pre-trained language models is to learn contextual representations of textual data. Pre-trained language models have become mainstream in natural language processing and code modeling. Using probes, a technique to study the linguistic properties of hidden vector spaces, previous works have shown that these pre-trained language models encode simple linguistic properties in their hidden representations. However, none of the previous work assessed whether these models encode the whole grammatical structure of a programming language. In this paper, we prove the existence of a syntactic subspace, lying in the hidden representations of pre-trained language models, which contain the syntactic information of the programming language. We show that this subspace can be extracted from the models’ representations and define a novel probing method, the AST-Probe, that enables recovering the whole abstract syntax tree (AST) of an input code snippet. In our experimentations, we show that this syntactic subspace exists in five state-of-the-art pre-trained language models. In addition, we highlight that the middle layers of the models are the ones that encode most of the AST information. Finally, we estimate the optimal size of this syntactic subspace and show that its dimension is substantially lower than those of the models’ representation spaces. This suggests that pre-trained language models use a small part of their representation spaces to encode syntactic information of the programming languages.