arXiv subjects

Gianmario Voria

Publications and source records attributed to Gianmario Voria.

11 recordsLinked to original sources

arXiv2026

How Do Developers Maintain and Evolve Their Agents' Instructions? An Empirical Study

Context. Autonomous coding agents are increasingly used in software development, shifting parts of the engineering process to AI assistance. While this automation brings clear benefits, it introduces challenges in governance, traceability, and control over agent behavior. Agent Context Files (ACFs) have emerged as a practical mechanism to guide agents through structured instructions, yet little is known about how these artifacts are maintained and how their evolution relates to code development. Objective. This paper plans to investigate the evolution of ACFs and their role in agent-driven development. Specifically, we (1) classify ACF changes through a taxonomy grounded in software maintenance theory, (2) analyze how different types of changes are associated with code quality outcomes, and (3) examine their temporal patterns across the development lifecycle. Method. We conduct a large-scale mining study combining repositories with ACFs and agent-generated commits. We reconstruct ACF evolution at the commit level, classify changes using a qualitative approach, and analyze their association with code quality metrics. Statistical analyses and hypotheses are used to evaluate differences across maintenance categories, to inform future design of ACFs for governing autonomous coding agents.

cs.SE↗

arXiv2026

SCOPE: A Dataset of Stereotyped Prompts for Counterfactual Fairness Assessment of LLMs

Large Language Models (LLMs) now serve as the foundation for a wide range of applications, from conversational assistants to decision support tools, making the issue of fairness in their results increasingly important. Previous studies have shown that LLM outputs can shift when prompts reference different demographic groups, even when intent and semantic content remain constant. However, existing resources for probing such disparities rely primarily on small, template-based counterfactual examples or fixed sentence pairs. These benchmarks offer limited linguistic diversity, narrow topical coverage, and little support for analyzing how communicative intent affects model behavior. To address these limitations, we introduce SCOPE (Stereotype-COnditioned Prompts for Evaluation), a large-scale dataset of counterfactual prompt pairs designed to enable systematic investigation of group-sensitive behavior in LLMs. SCOPE contains 241,280 prompts organized into 120,640 counterfactual pairs, each grounded in one of 1,438 topics and spanning nine bias dimensions and 1,536 demographic groups. All prompts are generated under four distinct communicative intents: Question, Recommendation, Direction, and Clarification, ensuring broad coverage of common interaction styles. This resource provides a controlled, semantically aligned, and intent-aware basis for evaluating fairness, robustness, and counterfactual consistency.

cs.SE↗

arXiv2026

Bias Ahead: Sensitive Prompts as Early Warnings for Fairness in Large Language Models

Large Language Models (LLMs) are being increasingly integrated into software systems, offering powerful capabilities but also raising concerns about fairness. Existing fairness benchmarks, however, focus on stereotype-specific associations, which limit their ability to anticipate risks in diverse, real-world contexts. In this paper, we propose sensitive prompts as a new abstraction for fairness evaluation: inputs that are not inherently biased but are more likely to elicit biased or inadequate responses due to the sensitivity of their content. We construct and release SensY, a dataset of 12,801 prompts, categorized as sensitive and non-sensitive, spanning seven thematic domains, combining synthetic generation and real user inputs. Using this dataset, we query three open-source LLMs and manually analyze 4,500 responses to evaluate their adequacy in answering sensitive prompts. Results show that while models often provide factually correct answers, they frequently fail to acknowledge the ethical, relational, or contextual implications of sensitive inputs. In addition, we develop an automated classifier for predicting prompt sensitivity, achieving robust performance on sensitive prompts. Our findings demonstrate that prompt sensitivity can serve as an effective early-warning mechanism for fairness risks in LLMs. This perspective shifts fairness assessment from reactive mitigation toward preventive design, enabling developers to anticipate and manage bias before deployment.

cs.SE↗

arXiv2026

Tracing Stereotypes in Pre-trained Transformers: From Biased Neurons to Fairer Models

The advent of transformer-based language models has reshaped how AI systems process and generate text. In software engineering (SE), these models now support diverse activities, accelerating automation and decision-making. Yet, evidence shows that these models can reproduce or amplify social biases, raising fairness concerns. Recent work on neuron editing has shown that internal activations in pre-trained transformers can be traced and modified to alter model behavior. Building on the concept of knowledge neurons, neurons that encode factual information, we hypothesize the existence of biased neurons that capture stereotypical associations within pre-trained transformers. To test this hypothesis, we build a dataset of biased relations, i.e., triplets encoding stereotypes across nine bias types, and adapt neuron attribution strategies to trace and suppress biased neurons in BERT models. We then assess the impact of suppression on SE tasks. Our findings show that biased knowledge is localized within small neuron subsets, and suppressing them substantially reduces bias with minimal performance loss. This demonstrates that bias in transformers can be traced and mitigated at the neuron level, offering an interpretable approach to fairness in SE.

cs.SE↗

arXiv2026

Once Upon a Team: Investigating Bias in LLM-Driven Software Team Composition and Task Allocation

LLMs are increasingly used to boost productivity and support software engineering tasks. However, when applied to socially sensitive decisions such as team composition and task allocation, they raise concerns of fairness. Prior studies have revealed that LLMs may reproduce stereotypes; however, these analyses remain exploratory and examine sensitive attributes in isolation. This study investigates whether LLMs exhibit bias in team composition and task assignment by analyzing the combined effects of candidates' country and pronouns. Using three LLMs and 3,000 simulated decisions, we find systematic disparities: demographic attributes significantly shaped both selection likelihood and task allocation, even when accounting for expertise-related factors. Task distributions further reflected stereotypes, with technical and leadership roles unevenly assigned across groups. Our findings indicate that LLMs exacerbate demographic inequities in software engineering contexts, underscoring the need for fairness-aware assessment.

cs.SE↗

arXiv2025

Toward Systematic Counterfactual Fairness Evaluation of Large Language Models: The CAFFE Framework

Nowadays, Large Language Models (LLMs) are foundational components of modern software systems. As their influence grows, concerns about fairness have become increasingly pressing. Prior work has proposed metamorphic testing to detect fairness issues, applying input transformations to uncover inconsistencies in model behavior. This paper introduces an alternative perspective for testing counterfactual fairness in LLMs, proposing a structured and intent-aware framework coined CAFFE (Counterfactual Assessment Framework for Fairness Evaluation). Inspired by traditional non-functional testing, CAFFE (1) formalizes LLM-Fairness test cases through explicitly defined components, including prompt intent, conversational context, input variants, expected fairness thresholds, and test environment configuration, (2) assists testers by automatically generating targeted test data, and (3) evaluates model responses using semantic similarity metrics. Our experiments, conducted on three different architectural families of LLM, demonstrate that CAFFE achieves broader bias coverage and more reliable detection of unfair behavior than existing metamorphic approaches.

cs.SE↗

arXiv2025

RECOVER: Toward Requirements Generation from Stakeholders' Conversations

Stakeholders' conversations in requirements elicitation meetings hold valuable insights into system and client needs. However, manually extracting requirements is time-consuming, labor-intensive, and prone to errors and biases. While current state-of-the-art methods assist in summarizing stakeholder conversations and classifying requirements based on their nature, there is a noticeable lack of approaches capable of both identifying requirements within these conversations and generating corresponding system requirements. These approaches would assist requirement identification, reducing engineers' workload, time, and effort. To address this gap, this paper introduces RECOVER (Requirements EliCitation frOm conVERsations), a novel conversational requirements engineering approach that leverages natural language processing and large language models (LLMs) to support practitioners in automatically extracting system requirements from stakeholder interactions. The approach is evaluated using a mixed-method study that combines performance analysis with a user study involving requirements engineers, targeting two levels of granularity. First, at the conversation turn level, the evaluation measures RECOVER's accuracy in identifying requirements-relevant dialogue and the quality of generated requirements in terms of correctness, completeness, and actionability. Second, at the entire conversation level, the evaluation assesses the overall usefulness and effectiveness of RECOVER in synthesizing comprehensive system requirements from full stakeholder discussions. Empirical evaluation of RECOVER shows promising performance, with generated requirements demonstrating satisfactory correctness, completeness, and actionability. The results also highlight the potential of automating requirements elicitation from conversations as an aid that enhances efficiency while maintaining human oversight

cs.SE↗

arXiv2025

A Catalog of Fairness-Aware Practices in Machine Learning Engineering

Machine learning's widespread adoption in decision-making processes raises concerns about fairness, particularly regarding the treatment of sensitive features and potential discrimination against minorities. The software engineering community has responded by developing fairness-oriented metrics, empirical studies, and approaches. However, there remains a gap in understanding and categorizing practices for engineering fairness throughout the machine learning lifecycle. This paper presents a novel catalog of practices for addressing fairness in machine learning derived from a systematic mapping study. The study identifies and categorizes 28 practices from existing literature, mapping them onto different stages of the machine learning lifecycle. From this catalog, the authors extract actionable items and implications for both researchers and practitioners in software engineering. This work aims to provide a comprehensive resource for integrating fairness considerations into the development and deployment of machine learning systems, enhancing their reliability, accountability, and credibility.

cs.SE↗

arXiv2025

Contextual Fairness-Aware Practices in ML: A Cost-Effective Empirical Evaluation

As machine learning (ML) systems become central to critical decision-making, concerns over fairness and potential biases have increased. To address this, the software engineering (SE) field has introduced bias mitigation techniques aimed at enhancing fairness in ML models at various stages. Additionally, recent research suggests that standard ML engineering practices can also improve fairness; these practices, known as fairness-aware practices, have been cataloged across each stage of the ML development life cycle. However, fairness remains context-dependent, with different domains requiring customized solutions. Furthermore, existing specific bias mitigation methods may sometimes degrade model performance, raising ongoing discussions about the trade-offs involved. In this paper, we empirically investigate fairness-aware practices from two perspectives: contextual and cost-effectiveness. The contextual evaluation explores how these practices perform in various application domains, identifying areas where specific fairness adjustments are particularly effective. The cost-effectiveness evaluation considers the trade-off between fairness improvements and potential performance costs. Our findings provide insights into how context influences the effectiveness of fairness-aware practices. This research aims to guide SE practitioners in selecting practices that achieve fairness with minimal performance costs, supporting the development of ethical ML systems.

cs.SE↗

arXiv2024

Data Preparation for Fairness-Performance Trade-Offs: A Practitioner-Friendly Alternative?

As machine learning (ML) systems are increasingly adopted across industries, addressing fairness and bias has become essential. While many solutions focus on ethical challenges in ML, recent studies highlight that data itself is a major source of bias. Pre-processing techniques, which mitigate bias before training, are effective but may impact model performance and pose integration difficulties. In contrast, fairness-aware Data Preparation practices are both familiar to practitioners and easier to implement, providing a more accessible approach to reducing bias. Objective. This registered report proposes an empirical evaluation of how optimally selected fairness-aware practices, applied in early ML lifecycle stages, can enhance both fairness and performance, potentially outperforming standard pre-processing bias mitigation methods. Method. To this end, we will introduce FATE, an optimization technique for selecting 'Data Preparation' pipelines that optimize fairness and performance. Using FATE, we will analyze the fairness-performance trade-off, comparing pipelines selected by FATE with results by pre-processing bias mitigation techniques.

cs.SE↗

arXiv2024

From Expectation to Habit: Why Do Software Practitioners Adopt Fairness Toolkits?

As the adoption of machine learning (ML) systems continues to grow across industries, concerns about fairness and bias in these systems have taken center stage. Fairness toolkits, designed to mitigate bias in ML models, serve as critical tools for addressing these ethical concerns. However, their adoption in the context of software development remains underexplored, especially regarding the cognitive and behavioral factors driving their usage. As a deeper understanding of these factors could be pivotal in refining tool designs and promoting broader adoption, this study investigates the factors influencing the adoption of fairness toolkits from an individual perspective. Guided by the Unified Theory of Acceptance and Use of Technology (UTAUT2), we examined the factors shaping the intention to adopt and actual use of fairness toolkits. Specifically, we employed Partial Least Squares Structural Equation Modeling (PLS-SEM) to analyze data from a survey study involving practitioners in the software industry. Our findings reveal that performance expectancy and habit are the primary drivers of fairness toolkit adoption. These insights suggest that by emphasizing the effectiveness of these tools in mitigating bias and fostering habitual use, organizations can encourage wider adoption. Practical recommendations include improving toolkit usability, integrating bias mitigation processes into routine development workflows, and providing ongoing support to ensure professionals see clear benefits from regular use.

cs.SE↗