arXiv subjects

Antonio Mastropaolo

Publications and source records attributed to Antonio Mastropaolo.

At least 19 recordsLinked to original sources

arXiv2026

Pattern over Pixels: Measuring Pattern Completion Bias in Multimodal Code Generation

Multimodal large language models (MLLMs) are increasingly used to translate webpage screenshots into front-end code, but repeated UI patterns may sway them toward visually incorrect yet pattern-consistent outputs. In this work, we test how repeated webpage patterns hurt MLLM accuracy on an objective screenshot-to-code fill-in-the-blank task. We introduce the first benchmark for visual pattern-completion bias, where one localized element in a repeated UI pattern is perturbed and the model must recover the masked width or font-size value from the screenshot and HTML context. Starting from 30 webpages curated from the Design2Code dataset, we build 1,440 evaluated screenshots spanning structural card and text-style patterns under standard and noise-overlaid conditions. We evaluate five frontier MLLMs and find that all are strongly biased toward the repeated baseline. Mean bias rate reaches 69.78% on card-width perturbations and 80.22% on text font-size perturbations, while mean accuracy is only 21.17% and 7.89%, respectively. Codex-5.3 performs best but still drops from 68.61% accuracy on cards to 13.89% on text, while Flash-3.0 reaches 96.11% bias on text. Noise, subtler perturbations, and boundary positions further increase bias rate. Reasoning analysis further shows that greater reasoning effort correlates with lower bias, yet qualitative evidence reveals that models can identify the anomalous element and still override it with the pattern-consistent answer. Our results identify a concrete failure mode in multimodal code generation and show that its severity is strongly associated with visual saliency

cs.SE↗

arXiv2026

Recovering Fine-Grained Code Change Rationale from Multiple Software Artifacts

Understanding the reasons behind past code changes is critical for refactoring, code review, and debugging. However, code change rationale is often fragmented, inconsistently documented, and scattered across heterogeneous artifacts. We address this challenge with two contributions. First, we conduct an empirical study of nine rationale components from an established taxonomy and trace where they are documented across artifacts associated with 63 commits from five widely used open-source Java projects. Seven components appear in practice, and rationale is highly fragmented: commit messages and pull requests primarily capture GOAL, while NEED and ALTERNATIVE are more often found in issues and pull requests. No single artifact type consistently captures all components, which presents the need for cross-document reasoning. Second, we introduce ARGUS, an LLM-based approach that identifies sentences expressing GOAL, NEED, and ALTERNATIVE across a commit's artifacts and synthesizes them into concise rationale summaries. ARGUS achieved 51.4% overall precision and 93.2% recall for rationale identification and generated summaries rated as accurate relative to reference summaries. Experiments across different LLMs showed varying identification performance but consistently accurate summaries. A user study with 12 Java programmers found these summaries useful for understanding unfamiliar code changes and supporting code review, documentation, debugging, and maintenance.

cs.SE↗

arXiv2026

Not All Tokens Matter: Data-Centric Optimization for Efficient Code Summarization

The rapid advancement of Large Language Models (LLMs) has revolutionized software engineering automation, particularly in automated code summarization, which enhances program comprehension and supports development activities. However, training LLMs for code summarization remains computationally expensive, with performance deteriorating on longer inputs-challenges that intensify when handling millions of code-comment pairs. We investigate strategic data optimization through targeted token reduction to minimize computational overhead while maintaining summary quality. We compare three token-level reduction techniques-(i) Abstract Syntax Tree (AST) representations, (ii) Function Signatures, and (iii) CrystalBLEU-guided pruning-combined with semantic filtering, evaluating them on Java and Python in standalone and cascaded reduction settings. Our findings reveal highly language-dependent optimal strategies: AST-based optimization achieves 37% performance improvements in Java with 56-73% token reduction but shows up to 49% degradation in Python. Conversely, Function Signatures perform poorly in Java but optimally in Python, achieving 83% token reduction while maintaining quality. CrystalBLEU demonstrates cross-language robustness with 60-72% reduction. These results challenge assumptions about cross-language transferability, demonstrating that which tokens are kept matters more than how many are removed, making language-aware token curation essential for efficient code summarization.

cs.SE↗

arXiv2026

Large Language Models for Code Generation from Multilingual Prompts: A Curated Benchmark and a Study on Code Quality

Large Language Models (LLMs) perform differently on identical programming tasks when prompted in different natural languages, a phenomenon known as language bias. While this behavior has been widely studied for general text generation, its impact on code generation quality and programming conventions remains largely unexplored. We investigate how the language used to describe programming tasks affects the source code generated by GPT-4o mini, DeepSeek, and Claude. Our study comprises 460 coding tasks spanning Python (230) and Java (230). We translate and manually curate the original English prompts into Chinese, Hindi, Spanish, and Italian while preserving their technical meaning. We evaluate the generated code using multiple dimensions, including functional correctness through test pass rates, structural quality using established code metrics, issues detected by static analysis tools, and lexical characteristics such as the language used in identifiers and comments. Our results show that (i) English prompts do not consistently produce the best functional correctness or code quality, (ii) the impact of prompt language depends on both the programming language and the LLM, and (iii) generated code frequently mixes English with the prompt language in comments and string literals. These findings provide the first curated multilingual benchmark for studying language bias in code generation and offer insights for developing more robust multilingual code generation systems.

cs.SE↗

arXiv2026

Quantize with Confidence? An Empirical Study of Quantization for Code Generation

The growing adoption of local inference frameworks such as Ollama has made it increasingly common for developers to run large code models on laptops and other resource-constrained hardware. In these settings, post-training quantization is essential for reducing memory footprint and enabling practical deployment, yet its impact on generated code remains insufficiently understood. We empirically evaluate six state-of-the-art quantization methods (GPTQ, AWQ, QuIP#, AQLM, BitsAndBytes, and GGUF) on two representative large code model families, Qwen2.5-Coder and CodeLlama, using the multilingual McEval and CoderEval benchmarks for Python and Java. We assess functional correctness (pass@1) together with maintainability, reliability, security, and structural complexity. We also introduce a novel analysis of robustness under varying prompt complexity, characterized by Shannon entropy and token length. Our results show that quantization techniques differ meaningfully in their impact on correctness and code quality. AQLM consistently matches or exceeds the full-precision baseline, whereas QuIP# exhibits the largest correctness degradation, particularly on complex prompts. Security attributes remain stable across models, benchmarks, and programming languages, while robustness to prompt complexity varies across techniques. These findings provide practical guidance for selecting quantization strategies for deploying large code models on resource-constrained hardware and highlight the importance of evaluating quantized models beyond functional correctness.

cs.SE↗

arXiv2026

Prior Bias in Vision Language Models on UML Diagram Interpretation

Vision Language Models (VLMs) are increasingly applied to software engineering artifacts, especially UML class diagrams whose meaning depends on visual notation. Yet, it is unclear whether VLMs actually read such diagrams or instead answer from pretrained priors about how classes typically relate. We introduce a controlled UML benchmark in which each prior-conforming diagram is paired with its prior-conflicting counterpart that (1) preserves the same class names and layout while (2) reverses only the relation arrow. We evaluate eight open-source VLMs from two model families, InternVL3.5 and Qwen3, alongside two closed-source frontier models GPT-5.4 and GPT-5.4 Mini. Across the eight open-source models, reversing the arrow reduces relation-direction accuracy by 33.48% on average, while GPT-5.4 Mini retains a 10% gap. In the harder three-class condition, accuracy drops sharply by 45.28% for open-source models, and even 18.62% for the GPT-5.4 family on average. Scaling provides only limited improvements and is family-dependent. Our benchmark presents a diagnostic prior-driven failure in diagram-grounded software understanding. Our artifact is available at https://anonymous.4open.science/r/UMLKnowledgeConflict-8461.

cs.CV↗

arXiv2026

Parameter-Efficient Multi-Task Fine-Tuning in Code-Related Tasks

Large Language Models (LLMs) have proven highly effective in automating software engineering tasks, bridging natural language and code semantics to achieve notable results in code generation and summarization. However, their scale incurs substantial computational costs, making full fine-tuning impractical. Parameter-Efficient Fine-Tuning (PEFT) methods like QLoRA enable efficient specialization with lower resource demands. Recent studies show QLoRA-optimized Large Code Models (LCMs) perform strongly across diverse tasks, yet it remains unclear whether this effectiveness persists when a single model is QLoRA fine-tuned for multiple code-related tasks. The interaction between Multi-task fine-tuning and QLoRA optimization, and how transfer learning affects correctness and quality of generated artifacts, remains largely unexplored. We investigate Multi-task QLoRA fine-tuning across three representative tasks: code generation, translation, and summarization. We evaluate functional correctness through execution-based and similarity-based metrics, complemented by comprehensive code quality analysis--an aspect largely overlooked in prior work. Our findings show that Multi-task QLoRA effectively leverages transfer learning, achieving competitive or superior performance at the 1.5B, 3B, and 7B configurations relative to both Single-task QLoRA and Multi-task full fine-tuning. Larger models demonstrate more consistent balance between correctness and quality, whereas smaller models preserve functionality but exhibit a higher incidence of quality-related issues.

cs.SE↗

arXiv2026

Rethinking Software Empirical Studies with Structural Causal Models

Causal Inference offers a fundamental approach for advancing empirical software engineering (ESE) beyond traditional statistical association, enabling researchers to rigorously identify and quantify causal relationships in software experiments. This paper introduces CausalSE, a framework that operationalizes Judea Pearl's causal inference paradigm in ESE context. The paper focuses on Structural Causal Models (SCMs) to address the limitations of classical statistical methods in mitigating confounding bias. Through a case study using the Galeras dataset and propensity score matching, we demonstrate how CausalSE disentangles the effect of prompt engineering strategies on code generation outcomes in a popular LLM (i.e., GPT-3). The results reveal that while associational analyses can suggest improvements in certain interventions (e.g., more complex prompts), causal analysis often does not find a significant treatment effect, highlighting the risk of false positives when confounding is not addressed. By providing a tutorial-based methodology and a real-world case study, this work equips software researchers with practical tools to design, analyze, and interpret software experiments with methodological rigor, ultimately enabling more informed and actionable conclusions in both research and practice.

cs.SE↗

arXiv2026

Prompt-Driven Code Summarization: A Systematic Literature Review

Software documentation is essential for program comprehension, developer onboarding, code review, and long-term maintenance. Yet producing quality documentation manually is time-consuming and frequently yields incomplete or inconsistent results. Large language models (LLMs) offer a promising solution by automatically generating natural language descriptions from source code, helping developers understand code more efficiently, facilitating maintenance, and supporting downstream activities such as defect localization and commit message generation. However, the effectiveness of LLMs in documentation tasks critically depends on how they are prompted. Properly structured instructions can substantially improve model performance, making prompt engineering-the design of input prompts to guide model behavior-a foundational technique in LLM-based software engineering. Approaches such as few-shot prompting, chain-of-thought reasoning, retrieval-augmented generation, and zero-shot learning show promise for code summarization, yet current research remains fragmented. There is limited understanding of which prompting strategies work best, for which models, and under what conditions. Moreover, evaluation practices vary widely, with most studies relying on overlap-based metrics that may not capture semantic quality. This systematic literature review consolidates existing evidence, categorizes prompting paradigms, examines their effectiveness, and identifies gaps to guide future research and practical adoption.

cs.SE↗

arXiv2026

Developers and Generative AI: A Study of Self-Admitted Usage in Open Source Projects

The availability of generative Artificial Intelligence (AI) tools such as ChatGPT or GitHub Copilot is reshaping the way in which software is developed, evolved, and maintained. Oftentimes, developers leave traces of such an usage in software artifacts. This allows not only to understand how AI is used in software development, but also to let others be aware how such software artifacts were created, e.g., for licensing or trustworthiness purposes. This paper-building upon our preliminary work presented at MSR 2024-aims at qualitatively investigating on the self-admitted use of two very popular generative AI tools - ChatGPT and GitHub Copilot - in software development. To this aim, we mined GitHub for such traces, by looking at commits, issues and pull requests (PRs). Then, through a manual coding, we create a taxonomy of 64 different ChatGPT and GitHub Copilot usage tasks, grouped into 7 categories. By repeating our previous analysis two years after and by extending it to GitHub Copilot, we show how the usage avenues have been expanded, the extent to which developers perceived such a generative AI usage useful, and whether some concerns occurring more than one year ago are no longer present. The taxonomy of tasks we derived from such a qualitative study provided (i) developers with valuable insights into how generative AI can be integrated into their workflows, and (ii) researchers with a clear overview of tasks that developers perceive as well-suited for automation.

cs.SE↗

arXiv2026

An Empirical Study on the Effects of System Prompts in Instruction-Tuned Models for Code Generation

Instruction-tuned Language Models (ILMs) have become essential components of modern AI systems, demonstrating exceptional versatility across natural language and reasoning tasks. Among their most impactful applications is code generation, where ILMs -- commonly referred to as Code Language Models (CLMs) -- translate human intent into executable programs. While progress has been driven by advances in scaling and training methodologies, one critical aspect remains underexplored: the impact of system prompts on both general-purpose ILMs and specialized CLMs for code generation. We systematically evaluate how system prompts of varying instructional detail, along with model scale, prompting strategy, and programming language, affect code assistant. Our experimental setting spans 360 configurations across four models, five system prompts, three prompting strategies, two languages, and two temperature settings. We find that (1) increasing system-prompt constraint specificity does not monotonically improve correctness -- prompt effectiveness is configuration-dependent and can help or hinder based on alignment with task requirements and decoding context; (2) for larger code-specialized models, few-shot examples can degrade performance relative to zero-shot generation, contrary to conventional wisdom; and (3) programming language matters, with Java exhibiting significantly greater sensitivity to system prompt variations than Python, suggesting language-specific prompt engineering strategies may be necessary.

cs.SE↗

arXiv2026

Towards Comprehensive Benchmarking Infrastructure for LLMs In Software Engineering

Large language models for code are advancing fast, yet our ability to evaluate them lags behind. Current benchmarks focus on narrow tasks and single metrics, which hide critical gaps in robustness, interpretability, fairness, efficiency, and real-world usability. They also suffer from inconsistent data engineering practices, limited software engineering context, and widespread contamination issues. To understand these problems and chart a path forward, we combined an in-depth survey of existing benchmarks with insights gathered from a dedicated community workshop. We identified three core barriers to reliable evaluation: the absence of software-engineering-rich datasets, overreliance on ML-centric metrics, and the lack of standardized, reproducible data pipelines. Building on these findings, we introduce BEHELM, a holistic benchmarking infrastructure that unifies software-scenario specification with multi-metric evaluation. BEHELM provides a structured way to assess models across tasks, languages, input and output granularities, and key quality dimensions. Our goal is to reduce the overhead currently required to construct benchmarks while enabling a fair, realistic, and future-proof assessment of LLMs in software engineering.

cs.SE↗

arXiv2026

Evaluating the Impact of Post-Training Quantization on Large Language Models for Code Generation

Large Language Models (LLMs) have shown an impressive capability in code generation. The LLM effectiveness generally increases with its size: The higher the number of LLM's trainable parameters the better its ability to implement code. However, when it comes to deploying LLM-based code generators, larger LLMs pose significant challenges related to their memory (and, consequently, carbon) footprint. A previous work by Wei et al. proposed to leverage quantization techniques to reduce the memory footprint of LLM-based code generators without substantially degrading their effectiveness. In short, they studied LLMs featuring up to 16B parameters, quantizing their precision from floating point 32 bits down to int 8 bits and showing their limited impact on code generation performance. Given the fast pace at which LLM capabilities and quantization techniques are evolving, in this work we present a differentiated replication of the work by Wei et al. in which we consider (i) on the one side, more recent and larger code-related LLMs, of up to 34B parameters; (ii) the latest advancements in model quantization techniques, which allow pushing the compression to the extreme quantization level of 2 bits per model parameter and; (iii) different types of calibration datasets to guide the quantization process, including code-specific ones. Our empirical evaluation reveals that the new frontier for LLM quantization is 4-bit precision, resulting in an average memory footprint reduction of 70% compared to the original model without observing any significant decrease in performance. Additionally, when the quantization becomes even more extreme (3 and 2 bits), a code-specific calibration dataset helps to limit the loss of performance.

cs.SE↗

arXiv2025

Toward Explaining Large Language Models in Software Engineering Tasks

Recent progress in Large Language Models (LLMs) has substantially advanced the automation of software engineering (SE) tasks, enabling complex activities such as code generation and code summarization. However, the black-box nature of LLMs remains a major barrier to their adoption in high-stakes and safety-critical domains, where explainability and transparency are vital for trust, accountability, and effective human supervision. Despite increasing interest in explainable AI for software engineering, existing methods lack domain-specific explanations aligned with how practitioners reason about SE artifacts. To address this gap, we introduce FeatureSHAP, the first fully automated, model-agnostic explainability framework tailored to software engineering tasks. Based on Shapley values, FeatureSHAP attributes model outputs to high-level input features through systematic input perturbation and task-specific similarity comparisons, while remaining compatible with both open-source and proprietary LLMs. We evaluate FeatureSHAP on two bi-modal SE tasks: code generation and code summarization. The results show that FeatureSHAP assigns less importance to irrelevant input features and produces explanations with higher fidelity than baseline methods. A practitioner survey involving 37 participants shows that FeatureSHAP helps practitioners better interpret model outputs and make more informed decisions. Collectively, FeatureSHAP represents a meaningful step toward practical explainable AI in software engineering. FeatureSHAP is available at https://github.com/deviserlab/FeatureSHAP.

cs.SE↗

arXiv2025

Smart but Costly? Benchmarking LLMs on Functional Accuracy and Energy Efficiency

The rapid advancement of AI technologies and their accelerated adoption in software development necessitates a systematic evaluation of their environmental impact alongside functional correctness. While prior studies have examined sustainability in large language models, existing approaches lack systematic frameworks for evaluating accuracy-energy trade-offs in Code Language Models (CLMs). In this paper, we present a framework, BRACE, to benchmark CLMs on a unified scale of energy efficiency and functional correctness (referred to as accuracy). We benchmark 22 state-of-the-art models on code generation and summarization tasks, proposing two rating methods: Concentric Incremental Rating Circles (CIRC) and Observation to Expectation Rating (OTER). CIRC provides deterministic Euclidean-based rankings with static trade-offs that are robust to outliers, and OTER offers trend-aware evaluation with dynamic trade-offs that capture the complex correlation between energy and accuracy, each offering a distinct perspective and addressing the problem in a unique way. These rating methods enable us to rate LLMs on a 1-5 scale reflecting their combined capabilities in terms of energy efficiency and functional correctness. Our analysis reveals models generally perform better in the code summarization tasks as they are not enforced to generate a grammar-based and syntactically correct output. Also, we find that models' size does not have a significant impact on their ratings, indicating that if models utilize their parameters efficiently, they can be ranked higher on these scales. The proposed BRACE framework empowers practitioners to make evidence-based model selections that balance sustainability with task requirements, guiding rating choice -- CIRC for deterministic comparisons or OTER for trend-aware evaluation -- based on deployment priorities.

cs.SE↗

arXiv2025

A Systematic Literature Review of Parameter-Efficient Fine-Tuning for Large Code Models

The rise of Artificial Intelligence (AI)-and particularly Large Language Models (LLMs) for code-has reshaped Software Engineering (SE) by enabling the automation of tasks such as code generation, bug detection, and repair. However, these models require significant computational resources for training and fine-tuning, posing challenges for real-world adoption in resource-constrained environments. To address this, the research community has increasingly turned to Parameter-Efficient Fine-Tuning (PEFT)-a class of techniques that enables the adaptation of large models by updating only a small subset of parameters, rather than the entire model. In this Systematic Literature Review (SLR), we examine the growing application of PEFT techniques-across a wide range of software engineering tasks. We analyze how these methods are used to optimize various deep learning (DL) architectures, focusing on their impact on both performance and efficiency. Our study synthesizes findings from 28 peer-reviewed papers, identifying patterns in configuration strategies and adaptation trade-offs. The outcome of this review is a comprehensive taxonomy that categorizes PEFT usage by task type, distinguishing between generative (e.g., Code Summarization) and non-generative (e.g., Code Clone Detection) scenarios. Our findings aim to inform future research and guide the practical deployment of PEFT in sustainable, AI-powered software development. Our artifacts are publicly available at https://github.com/alvi75/SLR-PEFT

cs.SE↗

arXiv2025

On the Effectiveness of LLM-as-a-judge for Code Generation and Summarization

Large Language Models have been recently exploited as judges for complex natural language processing tasks, such as Q&A. The basic idea is to delegate to an LLM the assessment of the "quality" of the output provided by an automated technique 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 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, 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 ~1.2k 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.

cs.SE↗

arXiv2025

Is Quantization a Deal-breaker? Empirical Insights from Large Code Models

The growing scale of large language models (LLMs) not only demands extensive computational resources but also raises environmental concerns due to their increasing carbon footprint. Model quantization emerges as an effective approach that can reduce the resource demands of LLMs by decreasing parameter precision without substantially affecting performance (e.g., 16 bit to 4 bit). While recent studies have established quantization as a promising approach for optimizing large code models (LCMs), a specialized subset of LLMs tailored for automated software engineering, their findings offer only limited insights into its practical implications. Specifically, current investigations focus only on the functional correctness of the code generated by quantized models, neglecting how quantization impacts critical aspects of code quality such as reliability, maintainability, and security. To bridge this gap, our study investigates the effects of quantization on the qualitative aspects of automatically generated code. We apply Activation-aware Weight Quantization (AWQ) to two widely used code models, CodeLlama and DeepSeekCoder, to generate Java and Python code. Using state-of-the-art static analysis tools, we evaluate software quality metrics and static features including cyclomatic complexity, cognitive complexity, and lines of code. Our findings reveal that quantization is a robust technique that not only preserves functional correctness, but also retains key qualitative code attributes sought after by developers, such as maintainability and structural simplicity.

cs.SE↗