Francesco Sovrano

Francesco Sovrano, Giulia Vilone, Michael Lognoul

Explainable AI (XAI) has evolved in response to expectations and regulations, such as the EU AI Act, which introduces regulatory requirements on AI-powered systems. However, a persistent gap remains between existing XAI methods and society's legal requirements, leaving practitioners without clear guidance on how to approach compliance in the EU market. To bridge this gap, we study model-agnostic XAI methods and relate their interpretability features to the requirements of the AI Act. We then propose a qualitative-to-quantitative scoring framework: qualitative expert assessments of XAI properties are aggregated into a regulation-specific compliance score. This helps practitioners identify when XAI solutions may support legal explanation requirements while highlighting technical issues that require further research and regulatory clarification.

Francesco Sovrano, Gabriele Dominici, Alberto Bacchelli

Prompt-induced cognitive biases are changes in a general-purpose AI (GPAI) system's decisions caused solely by biased wording in the input (e.g., framing, anchors), not task logic. In software engineering (SE) decision support (where problem statements and requirements are natural language) small phrasing shifts (e.g., popularity hints or outcome reveals) can push GPAI models toward suboptimal decisions. We study this with PROBE-SWE, a dynamic benchmark for SE that pairs biased and unbiased versions of the same SE dilemmas, controls for logic and difficulty, and targets eight SE-relevant biases (anchoring, availability, bandwagon, confirmation, framing, hindsight, hyperbolic discounting, overconfidence). We ask whether prompt engineering mitigates bias sensitivity in practice, focusing on actionable techniques that practitioners can apply off-the-shelf in real environments. Testing common strategies (e.g., chain-of-thought, self-debiasing) on cost-effective GPAI systems, we find no statistically significant reductions in bias sensitivity on a per-bias basis. We then adopt a Prolog-style view of the reasoning process: solving SE dilemmas requires making explicit any background axioms and inference assumptions (i.e., SE best practices) that are usually implicit in the prompt. So, we hypothesize that bias-inducing features short-circuit assumptions elicitation, pushing GPAI models toward biased shortcuts. Building on this, we introduce an end-to-end method that elicits best practices and injects axiomatic reasoning cues into the prompt before answering, reducing overall bias sensitivity by 51% on average (p < .001). Finally, we report a thematic analysis that surfaces linguistic patterns associated with heightened bias sensitivity, clarifying when GPAI use is less advisable for SE decision support and where to focus future countermeasures.

Francesco Sovrano, Lidia Losavio, Giulia Vilone et al.

Symbolic regression aims to replace black-box predictors with concise analytical expressions that can be inspected and validated in scientific machine learning. Kolmogorov-Arnold Networks (KANs) are well suited to this goal because each connection between adjacent units (an "edge") is parametrised by a learnable univariate function that can, in principle, be replaced by a symbolic operator. In practice, however, symbolic extraction is a bottleneck: the standard KAN-to-symbol approach fits operators to each learned edge function in isolation, making the discrete choice sensitive to initialisation and non-convex parameter fitting, and ignoring how local substitutions interact through the full network. We study in-context symbolic regression for operator extraction in KANs, and present two complementary instantiations. Greedy in-context Symbolic Regression (GSR) performs greedy, in-context selection by choosing edge replacements according to end-to-end loss improvement after brief fine-tuning. Gated Matching Pursuit (GMP) amortises this in-context selection by training a differentiable gated operator layer that places an operator library behind sparse gates on each edge; after convergence, gates are discretised (optionally followed by a short in-context greedy refinement pass). We quantify robustness via one-factor-at-a-time (OFAT) hyper-parameter sweeps and assess both predictive error and qualitative consistency of recovered formulas. Across several experiments, greedy in-context symbolic regression achieves up to 99.8% reduction in median OFAT test MSE.

Francesco Sovrano, Alberto Bacchelli

Natural language explanations produced by large language models (LLMs) are often persuasive, but not necessarily scrutable: users cannot easily verify whether the claims in an explanation are supported by evidence. In XAI, this motivates a focus on faithfulness and traceability, i.e., the extent to which an explanation's claims can be grounded in, and traced back to, an explicit source. We study these desiderata in retrieval-augmented generation (RAG) for programming education, where textbooks provide authoritative evidence. We benchmark six LLMs on 90 Stack Overflow questions grounded in three programming textbooks and quantify source faithfulness via source adherence metrics. We find that non Retrieval-Augmented Generation (RAG) models have median source adherence of 0%, while baseline RAG systems still exhibit low median adherence (22-40%, depending on the model). Motivated by Achinstein's illocutionary theory of explanation, we introduce illocutionary macro-planning as a descriptive design principle for source-faithful explanations and instantiate it with chain-of-illocution prompting (CoI), which expands a query into implicit explanatory questions that drive retrieval. Across models, CoI yields statistically significant gains (up to 63%) in source adherence, although absolute adherence remains moderate and the gains are weak or non-significant for some models. A user study with 165 retained participants (220 recruited) indicates that these gains do not harm satisfaction, relevance, or perceived correctness.

Francesco Sovrano, Gabriele Dominici, Marc Langheinrich

A key goal of explainable AI (XAI) is to express the decision logic of large language models (LLMs) in symbolic form and link it to internal mechanisms. Global rule-extraction methods typically learn symbolic surrogates without grounding rules in model circuitry, while mechanistic interpretability can connect behaviors to neuron sets but often depends on hand-crafted hypotheses and expensive neuron-level interventions. We introduce MechaRule, a pipeline that grounds rule extraction in LLM circuits by efficiently localizing sparse neurons called agonists, whose activation neutralization disrupts rule-related behaviors. MechaRule rests on two empirical observations. First, within a fixed baseline/flip regime, sparse agonist effects can be approximately monotone and saturating: a few dominant neuron activations can overtop weaker ones at coarse scales, while overlapping neurons flip many of the same examples. This motivates viewing localization as adaptive group testing driven by a regime-conditional strength predicate with confidence-guided conservative pruning, yielding Theta(k log(N/k) + k) interventions over N candidates when k << N neurons are agonists under the monotone-overtopping abstraction. Second, agonists emerge more reliably when ablations are verified through data splits aligned with close-to-faithful rule behavior; spectral splits remain a useful rule-free fallback, while unfaithful splits degrade localization. Empirically, overtopping appears mainly in learned, task-aligned regimes: on arithmetic and jailbreak tasks across Qwen2 and GPT-J, MechaRule recalls 96.8% of high-effect brute-force agonists in completed comparisons, and suppressing localized agonists reduces arithmetic accuracy and jailbreak success by up to 71.1% and 8.8%, respectively.

Francesco Sovrano, Michael Lognoul, Giulia Vilone

Significant investment and development have gone into integrating Artificial Intelligence (AI) in medical and healthcare applications, leading to advanced control systems in medical technology. However, the opacity of AI systems raises concerns about essential characteristics needed in such sensitive applications, like transparency and trustworthiness. Our study addresses these concerns by investigating a process for selecting the most adequate Explainable AI (XAI) methods to comply with the explanation requirements of key EU regulations in the context of smart bioelectronics for medical devices. The adopted methodology starts with categorising smart devices by their control mechanisms (open-loop, closed-loop, and semi-closed-loop systems) and delving into their technology. Then, we analyse these regulations to define their explainability requirements for the various devices and related goals. Simultaneously, we classify XAI methods by their explanatory objectives. This allows for matching legal explainability requirements with XAI explanatory goals and determining the suitable XAI algorithms for achieving them. Our findings provide a nuanced understanding of which XAI algorithms align better with EU regulations for different types of medical devices. We demonstrate this through practical case studies on different neural implants, from chronic disease management to advanced prosthetics. This study fills a crucial gap in aligning XAI applications in bioelectronics with stringent provisions of EU regulations. It provides a practical framework for developers and researchers, ensuring their AI innovations advance healthcare technology and adhere to legal and ethical standards.

Francesco Sovrano, Adam Bauer, Alberto Bacchelli

Recent advancements in artificial intelligence have enabled processing of larger inputs, leading everyday software developers to increasingly rely on chat-based large language models (LLMs) like GPT-3.5 and GPT-4 to detect vulnerabilities across entire files, not just within functions. This new development practice requires researchers to urgently investigate whether commonly used LLMs can effectively analyze large file-sized inputs, in order to provide timely insights for software developers and engineers about the pros and cons of this emerging technological trend. Hence, the goal of this paper is to evaluate the effectiveness of several state-of-the-art chat-based LLMs, including the GPT models, in detecting in-file vulnerabilities. We conducted a costly investigation into how the performance of LLMs varies based on vulnerability type, input size, and vulnerability location within the file. To give enough statistical power to our study, we could only focus on the three most common (as well as dangerous) vulnerabilities: XSS, SQL injection, and path traversal. Our findings indicate that the effectiveness of LLMs in detecting these vulnerabilities is strongly influenced by both the location of the vulnerability and the overall size of the input. Specifically, regardless of the vulnerability type, LLMs tend to significantly (p < .05) underperform when detecting vulnerabilities located toward the end of larger files, a pattern we call the 'lost-in-the-end' effect. Finally, to further support software developers and practitioners, we also explored the optimal input size for these LLMs and presented a simple strategy for identifying it, which can be applied to other models and vulnerability types. Eventually, we show how adjusting the input size can lead to significant improvements in LLM-based vulnerability detection, with an average recall increase of over 37% across all models.

Francesco Sovrano, Michaël Lognoul, Alberto Bacchelli

Compliance with the European Union's Platform-to-Business (P2B) Regulation is challenging for online platforms, and assessing their compliance can be difficult for public authorities. This is partly due to the lack of automated tools for assessing the information (e.g., software documentation) platforms provide concerning ranking transparency. Our study tackles this issue in two ways. First, we empirically evaluate the compliance of six major platforms (Amazon, Bing, Booking, Google, Tripadvisor, and Yahoo), revealing substantial differences in their documentation. Second, we introduce and test automated compliance assessment tools based on ChatGPT and information retrieval technology. These tools are evaluated against human judgments, showing promising results as reliable proxies for compliance assessments. Our findings could help enhance regulatory compliance and align with the United Nations Sustainable Development Goal 10.3, which seeks to reduce inequality, including business disparities, on these platforms.

Francesco Sovrano, Gabriele Dominici, Rita Sevastjanova et al.

Human cognitive biases in software engineering can lead to costly errors. While general-purpose AI (GPAI) systems may help mitigate these biases due to their non-human nature, their training on human-generated data raises a critical question: Do GPAI systems themselves exhibit cognitive biases? To investigate this, we present the first dynamic benchmarking framework to evaluate data-induced cognitive biases in GPAI within software engineering workflows. Starting with a seed set of 16 hand-crafted realistic tasks, each featuring one of 8 cognitive biases (e.g., anchoring, framing) and corresponding unbiased variants, we test whether bias-inducing linguistic cues unrelated to task logic can lead GPAI systems from correct to incorrect conclusions. To scale the benchmark and ensure realism, we develop an on-demand augmentation pipeline relying on GPAI systems to generate task variants that preserve bias-inducing cues while varying surface details. This pipeline ensures correctness (88--99% on average, according to human evaluation), promotes diversity, and controls reasoning complexity by leveraging Prolog-based reasoning and LLM-as-a-judge validation. It also verifies that the embedded biases are both harmful and undetectable by logic-based, unbiased reasoners. We evaluate leading GPAI systems (GPT, LLaMA, DeepSeek) and find a consistent tendency to rely on shallow linguistic heuristics over deep reasoning. All systems exhibit cognitive biases (ranging from 5.9% to 35% across types), with bias sensitivity increasing sharply with task complexity (up to 49%), highlighting critical risks in real-world software engineering deployments.

Alessandra Stramiglio, Andrea Schimmenti, Valentina Pasqual et al.

Text Implicitness has always been challenging in Natural Language Processing (NLP), with traditional methods relying on explicit statements to identify entities and their relationships. From the sentence "Zuhdi attends church every Sunday", the relationship between Zuhdi and Christianity is evident for a human reader, but it presents a challenge when it must be inferred automatically. Large language models (LLMs) have proven effective in NLP downstream tasks such as text comprehension and information extraction (IE). This study examines how textual implicitness affects IE tasks in pre-trained LLMs: LLaMA 2.3, DeepSeekV1, and Phi1.5. We generate two synthetic datasets of 10k implicit and explicit verbalization of biographic information to measure the impact on LLM performance and analyze whether fine-tuning implicit data improves their ability to generalize in implicit reasoning tasks. This research presents an experiment on the internal reasoning processes of LLMs in IE, particularly in dealing with implicit and explicit contexts. The results demonstrate that fine-tuning LLM models with LoRA (low-rank adaptation) improves their performance in extracting information from implicit texts, contributing to better model interpretability and reliability.

Francesco Sovrano, Salvatore Sapienza, Monica Palmirani et al.

This study discusses the interplay between metrics used to measure the explainability of the AI systems and the proposed EU Artificial Intelligence Act. A standardisation process is ongoing: several entities (e.g. ISO) and scholars are discussing how to design systems that are compliant with the forthcoming Act and explainability metrics play a significant role. This study identifies the requirements that such a metric should possess to ease compliance with the AI Act. It does so according to an interdisciplinary approach, i.e. by departing from the philosophical concept of explainability and discussing some metrics proposed by scholars and standardisation entities through the lenses of the explainability obligations set by the proposed AI Act. Our analysis proposes that metrics to measure the kind of explainability endorsed by the proposed AI Act shall be risk-focused, model-agnostic, goal-aware, intelligible & accessible. This is why we discuss the extent to which these requirements are met by the metrics currently under discussion.

Francesco Sovrano, Fabio Vitali

We propose a new method for generating explanations with AI and a tool to test its expressive power within a user interface. In order to bridge the gap between philosophy and human-computer interfaces, we show a new approach for the generation of interactive explanations based on a sophisticated pipeline of AI algorithms for structuring natural language documents into knowledge graphs, answering questions effectively and satisfactorily. With this work we aim to prove that the philosophical theory of explanations presented by Achinstein can be actually adapted for being implemented into a concrete software application, as an interactive and illocutionary process of answering questions. Specifically, our contribution is an approach to frame illocution in a computer-friendly way, to achieve user-centrality with statistical question answering. Indeed, we frame the illocution of an explanatory process as that mechanism responsible for anticipating the needs of the explainee in the form of unposed, implicit, archetypal questions, hence improving the user-centrality of the underlying explanatory process. Therefore, we hypothesise that if an explanatory process is an illocutionary act of providing content-giving answers to questions, and illocution is as we defined it, the more explicit and implicit questions can be answered by an explanatory tool, the more usable its explanations. We tested our hypothesis with a user-study involving more than 60 participants, on two XAI-based systems, one for credit approval and one for heart disease prediction. The results showed that increasing the illocutionary power of an explanatory tool can produce statistically significant improvements on effectiveness. This, combined with a visible alignment between the increments in effectiveness and satisfaction, suggests that our understanding of illocution can be correct, giving evidence in favour of our theory.

Francesco Sovrano, Fabio Vitali, Monica Palmirani

The European Union (EU) through the High-Level Expert Group on Artificial Intelligence (AI-HLEG) and the General Data Protection Regulation (GDPR) has recently posed an interesting challenge to the eXplainable AI (XAI) community, by demanding a more user-centred approach to explain Automated Decision-Making systems (ADMs). Looking at the relevant literature, XAI is currently focused on producing explainable software and explanations that generally follow an approach we could term One-Size-Fits-All, that is unable to meet a requirement of centring on user needs. One of the causes of this limit is the belief that making things explainable alone is enough to have pragmatic explanations. Thus, insisting on a clear separation between explainabilty (something that can be explained) and explanations, we point to explanatorY AI (YAI) as an alternative and more powerful approach to win the AI-HLEG challenge. YAI builds over XAI with the goal to collect and organize explainable information, articulating it into something we called user-centred explanatory discourses. Through the use of explanatory discourses/narratives we represent the problem of generating explanations for Automated Decision-Making systems (ADMs) into the identification of an appropriate path over an explanatory space, allowing explainees to interactively explore it and produce the explanation best suited to their needs.

Francesco Sovrano, Alex Raymond, Amanda Prorok

Human environments are often regulated by explicit and complex rulesets. Integrating Reinforcement Learning (RL) agents into such environments motivates the development of learning mechanisms that perform well in rule-dense and exception-ridden environments such as autonomous driving on regulated roads. In this paper, we propose a method for organising experience by means of partitioning the experience buffer into clusters labelled on a per-explanation basis. We present discrete and continuous navigation environments compatible with modular rulesets and 9 learning tasks. For environments with explainable rulesets, we convert rule-based explanations into case-based explanations by allocating state-transitions into clusters labelled with explanations. This allows us to sample experiences in a curricular and task-oriented manner, focusing on the rarity, importance, and meaning of events. We label this concept Explanation-Awareness (XA). We perform XA experience replay (XAER) with intra and inter-cluster prioritisation, and introduce XA-compatible versions of DQN, TD3, and SAC. Performance is consistently superior with XA versions of those algorithms, compared to traditional Prioritised Experience Replay baselines, indicating that explanation engineering can be used in lieu of reward engineering for environments with explainable features.

Francesco Sovrano, Fabio Vitali

Explainable AI was born as a pathway to allow humans to explore and understand the inner working of complex systems. However, establishing what is an explanation and objectively evaluating explainability are not trivial tasks. This paper presents a new model-agnostic metric to measure the Degree of Explainability of information in an objective way. We exploit a specific theoretical model from Ordinary Language Philosophy called the Achinstein's Theory of Explanations, implemented with an algorithm relying on deep language models for knowledge graph extraction and information retrieval. To understand whether this metric can measure explainability, we devised a few experiments and user studies involving more than 190 participants, evaluating two realistic systems for healthcare and finance using famous AI technology, including Artificial Neural Networks and TreeSHAP. The results we obtained are statistically significant (with P values lower than .01), suggesting that our proposed metric for measuring the Degree of Explainability is robust in several scenarios, and it aligns with concrete expectations.

Francesco Sovrano, Fabio Vitali

We propose a new method for explanations in Artificial Intelligence (AI) and a tool to test its expressive power within a user interface. In order to bridge the gap between philosophy and human-computer interfaces, we show a new approach for the generation of interactive explanations based on a sophisticated pipeline of AI algorithms for structuring natural language documents into knowledge graphs, answering questions effectively and satisfactorily. Among the mainstream philosophical theories of explanation we identified one that in our view is more easily applicable as a practical model for user-centric tools: Achinstein's Theory of Explanation. With this work we aim to prove that the theory proposed by Achinstein can be actually adapted for being implemented into a concrete software application, as an interactive process answering questions. To this end we found a way to handle the generic (archetypal) questions that implicitly characterise an explanatory processes as preliminary overviews rather than as answers to explicit questions, as commonly understood. To show the expressive power of this approach we designed and implemented a pipeline of AI algorithms for the generation of interactive explanations under the form of overviews, focusing on this aspect of explanations rather than on existing interfaces and presentation logic layers for question answering. We tested our hypothesis on a well-known XAI-powered credit approval system by IBM, comparing CEM, a static explanatory tool for post-hoc explanations, with an extension we developed adding interactive explanations based on our model. The results of the user study, involving more than 100 participants, showed that our proposed solution produced a statistically relevant improvement on effectiveness (U=931.0, p=0.036) over the baseline, thus giving evidence in favour of our theory.

Francesco Sovrano, Fabio Vitali, Monica Palmirani

Through the General Data Protection Regulation (GDPR), the European Union has set out its vision for Automated Decision- Making (ADM) and AI, which must be reliable and human-centred. In particular we are interested on the Right to Explanation, that requires industry to produce explanations of ADM. The High-Level Expert Group on Artificial Intelligence (AI-HLEG), set up to support the implementation of this vision, has produced guidelines discussing the types of explanations that are appropriate for user-centred (interactive) Explanatory Tools. In this paper we propose our version of Explanatory Narratives (EN), based on user-centred concepts drawn from ISO 9241, as a model for user-centred explanations aligned with the GDPR and the AI-HLEG guidelines. Through the use of ENs we convert the problem of generating explanations for ADM into the identification of an appropriate path over an Explanatory Space, allowing explainees to interactively explore it and produce the explanation best suited to their needs. To this end we list suitable exploration heuristics, we study the properties and structure of explanations, and discuss the proposed model identifying its weaknesses and strengths.

Francesco Sovrano, Monica Palmirani, Fabio Vitali

The main goal of this research is to produce a useful software for United Nations (UN), that could help to speed up the process of qualifying the UN documents following the Sustainable Development Goals (SDGs) in order to monitor the progresses at the world level to fight poverty, discrimination, climate changes. In fact human labeling of UN documents would be a daunting task given the size of the impacted corpus. Thus, automatic labeling must be adopted at least as a first step of a multi-phase process to reduce the overall effort of cataloguing and classifying. Deep Learning (DL) is nowadays one of the most powerful tools for state-of-the-art (SOTA) AI for this task, but very often it comes with the cost of an expensive and error-prone preparation of a training-set. In the case of multi-label text classification of domain-specific text it seems that we cannot effectively adopt DL without a big-enough domain-specific training-set. In this paper, we show that this is not always true. In fact we propose a novel method that is able, through statistics like TF-IDF, to exploit pre-trained SOTA DL models (such as the Universal Sentence Encoder) without any need for traditional transfer learning or any other expensive training procedure. We show the effectiveness of our method in a legal context, by classifying UN Resolutions according to their most related SDGs.

Francesco Sovrano

Our work is a simple extension of the paper "Exploration by Random Network Distillation". More in detail, we show how to efficiently combine Intrinsic Rewards with Experience Replay in order to achieve more efficient and robust exploration (with respect to PPO/RND) and consequently better results in terms of agent performances and sample efficiency. We are able to do it by using a new technique named Prioritized Oversampled Experience Replay (POER), that has been built upon the definition of what is the important experience useful to replay. Finally, we evaluate our technique on the famous Atari game Montezuma's Revenge and some other hard exploration Atari games.

Andrea Asperti, Daniele Cortesi, Francesco Sovrano

Rogue is a famous dungeon-crawling video-game of the 80ies, the ancestor of its gender. Rogue-like games are known for the necessity to explore partially observable and always different randomly-generated labyrinths, preventing any form of level replay. As such, they serve as a very natural and challenging task for reinforcement learning, requiring the acquisition of complex, non-reactive behaviors involving memory and planning. In this article we show how, exploiting a version of A3C partitioned on different situations, the agent is able to reach the stairs and descend to the next level in 98% of cases.