Research

AI Planning & Model Checking (2009–2017 · 23 papers)

Model checking as a planning engine for hybrid discrete–continuous systems.

Where the field stood. By the late 2000s automated planning worked well in purely discrete domains, but real systems — plants, vehicles, physical processes — mix continuous dynamics with discrete decisions: existing PDDL+ planners struggled with them, and 'universal' planning (finding a valid policy from every state, not a single plan) was considered computationally prohibitive.

Our contribution. The key insight is to use model checking — the exhaustive formal verification of a system's states — as a planning engine: this gave rise to UPMurphi (ICAPS 2009), a discretisation-based universal planner for PDDL+, later extended with heuristics in DiNo (IJCAI 2016) together with the planning group at King's College London. In parallel, theoretical results on the synthesis of strong plans with optimal cost in non-deterministic domains.

Why it matters. It makes it possible to automatically generate robust control policies for hybrid systems, from chemical benchmarks to planetary rovers. It is the methodological root of the whole journey: the formal rigour of this phase lives on in data quality and, today, in the geometric analysis of neural networks.

Open-source tools: UPMurphi, DiNo

• An AI Planning System for Data Cleaning (ECML-PKDD 2017, 2017). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Shows that data cleaning can be framed as an AI planning problem: the system automatically finds the optimal sequence of corrections. A bridge between the group's two souls: planning and data quality. Contribution: Frames data cleaning as an AI planning problem: the planner automatically synthesises the optimal sequence of corrections. The formal bridge between the planning and data quality research lines. [read the paper · full reference]
• PDDL+ Planning with Temporal Pattern Databases (AAAI 2017, 2017). Wiktor Mateusz Piotrowski, Maria Fox, Derek Long, Daniele Magazzeni, Fabio Mercorio. In plain terms: Improves PDDL+ planning (hybrid discrete/continuous systems) with temporal pattern databases that guide the search for a solution. Contribution: Introduces temporal pattern databases as a heuristic for PDDL+ planning: informed search guidance for hybrid domains. [read the paper · full reference]
• Heuristic Planning for Hybrid Systems (AAAI 2016, 2016). Wiktor Mateusz Piotrowski, Maria Fox, Derek Long, Daniele Magazzeni, Fabio Mercorio. In plain terms: Presents heuristic planning for hybrid systems at AAAI: how to find plans in domains where the world evolves both continuously (physics, processes) and discretely at the same time. Contribution: Heuristic planning for hybrid systems, presented at AAAI: a new level of scalability for domains with continuous dynamics. [read the paper · DiNo]
• Heuristic Planning for PDDL+ Domains (AAAI 2016, 2016). Wiktor Mateusz Piotrowski, Maria Fox, Derek Long, Daniele Magazzeni, Fabio Mercorio. In plain terms: The workshop version of the work on DiNo: heuristic planning for PDDL+ domains. Contribution: The workshop version of the work on DiNo: heuristics for PDDL+ planning. [read the paper · full reference · DiNo]
• Heuristic Planning for PDDL+ Domains (IJCAI 2016, 2016). Wiktor Mateusz Piotrowski, Maria Fox, Derek Long, Daniele Magazzeni, Fabio Mercorio. In plain terms: DiNo, developed with the planning group at King's College London, is the heuristic planner for PDDL+ (hybrid discrete-continuous) domains: it extends UPMurphi with heuristics that multiply its scalability. Contribution: DiNo adds informed heuristics (a hybrid TRPG) to the UPMurphi engine: the first heuristic planner for full PDDL+, developed with the KCL Planning Group, multiplying the scale of tractable problems. [read the paper · full reference · DiNo]
• Synthesis of Cost-Optimal Strong Plans in Non-Deterministic Domains (Int. J. AI Tools, 2015). Giuseppe Della Penna, Benedetto Intrigila, Daniele Magazzeni, Fabio Mercorio. In plain terms: How to synthesise cost-optimal 'strong' plans (guaranteeing success no matter what happens) in non-deterministic domains: theoretical and algorithmic results. Contribution: Algorithms for synthesising cost-optimal strong plans in non-deterministic domains: worst-case success guarantees at minimum cost. [read the paper · full reference]
• UPMurphi Released: PDDL+ Planning for Hybrid Systems (2nd Workshop on Model Checking and Automated Planning (MO…, 2015). Giuseppe Della Penna, Benedetto Intrigila, Daniele Magazzeni, Fabio Mercorio. In plain terms: The public release of UPMurphi: the universal PDDL+ planner based on model checking, packaged and documented for the community, with its discretise-and-validate approach to planning in hybrid discrete–continuous domains. Contribution: Tool release at the MOCHAP workshop that made the ICAPS 2009 planner available to the planning community. [read the paper · full reference · UPMurphi]
• Accurate Data Cleansing through Model Checking and Machine Learning Techniques (DATA 2014, 2014). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Combines model checking and machine learning for more accurate data cleansing: formal methods detect the inconsistencies, ML learns how to fix them. Contribution: Combines model checking (to detect inconsistencies) with machine learning (to learn how to fix them), achieving higher accuracy than either approach alone. [read the paper · full reference]
• Planning meets Data Cleansing (ICAPS 2014, 2014). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: 'Planning meets Data Cleansing', presented at ICAPS: data cleansing framed as an automated planning problem, with the planner synthesising the correction sequences. The conceptual bridge between the first and second phases of the group's research. Contribution: 'Planning meets Data Cleansing' at ICAPS: the first formulation of data cleansing as an automated planning problem, with automatic synthesis of correction sequences. The hinge paper between the formal and the data-driven phases of the research path. [read the paper · full reference]
• Towards data cleansing via planning (Intelligenza Artificiale, 2014). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Extended journal version of the idea of tackling data cleansing through automated planning. Contribution: Journal extension of the planning-based approach to data cleansing. [read the paper · full reference]
• Model checking for universal planning in deterministic and non-deterministic domains (AI Communications, 2013). Fabio Mercorio. In plain terms: Journal article (derived from the PhD thesis): model checking as an engine for universal planning in deterministic and non-deterministic domains — finding not a plan, but a policy that works from every state. Contribution: Consolidates the thesis contribution in journal form: model checking as a universal planning engine — policies valid from every state, in both deterministic and non-deterministic domains. [read the paper · full reference]
• Business Model Design as a Temporal Planning Problem: Preliminary Results (5th Italian Workshop on Planning and Scheduling at AIxIA …, 2013). Daniele Magazzeni, Fabio Mercorio, Balbir Barn, Tony Clark, Franco Raimondi, Vinay Kulkarni. In plain terms: Can designing a business model be treated as an AI planning problem? The paper encodes business-model choices as a temporal planning domain, so that a planner can synthesise and compare viable strategies over time. Contribution: Preliminary formulation of business-model design as temporal planning — an unconventional application of the group's planning machinery. [read the paper · full reference]
• Can Planning meet Data Cleansing? (5th Italian Workshop on Planning and Scheduling at AIxIA …, 2013). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: The workshop paper that first asked the question: can a planner clean data? Inconsistencies in longitudinal records become states to escape, and cleansing actions become planning operators. Contribution: First formulation of data cleansing as automated planning — the direct precursor of 'Planning meets Data Cleansing' (ICAPS 2014). [read the paper · full reference]
• A universal planning system for hybrid domains (Applied Intelligence, 2012). Giuseppe Della Penna, Daniele Magazzeni, Fabio Mercorio. In plain terms: A universal planning system for hybrid domains: it handles continuous dynamics and non-determinism, generating robust policies for real-world systems. Contribution: Universal planning system for hybrid domains: robust policies under continuous dynamics and non-determinism. [read the paper · full reference]
• Model checking for the analysis and control of complex and non-deterministic systems (Tesi di dottorato, Univ. dell'Aquila, 2012). Fabio Mercorio. In plain terms: The PhD thesis: model checking for the analysis and control of complex, non-deterministic systems — the theoretical foundation of the planning and data quality research lines. Contribution: The PhD thesis that grounds the whole path: model checking for the analysis and control of complex, non-deterministic systems.
• Towards the Use of Model Checking for Performing Data Consistency Evaluation and Cleansing (ICIQ 2012, 2012). Mario Mezzanzanica, Mirko Cesarini, Fabio Mercorio, Roberto Boselli. In plain terms: Proposes using model checking to evaluate and cleanse data consistency: formal software verification techniques applied to information archives. Contribution: Proposes model checking for evaluating and cleansing data consistency: formal software verification transferred to information archives. [full reference]
• Cost-optimal Strong Planning in Non-deterministic Domains (ICINCO 2011, 2011). Giuseppe Della Penna, Fabio Mercorio, Benedetto Intrigila, Daniele Magazzeni, Enrico Tronci. In plain terms: Cost-optimal 'strong' planning in non-deterministic domains: algorithms that guarantee reaching the goal at minimum worst-case cost. Contribution: Algorithms for cost-optimal strong planning in non-deterministic domains. [read the paper · full reference]
• Data Quality through Model Checking Techniques (IDA 2011, 2011). Mario Mezzanzanica, Roberto Boselli, Mirko Cesarini, Fabio Mercorio. In plain terms: The first work to bring model checking into data quality: formally verifying the consistency of longitudinal data archives. The founding act of the group's DQ research line. Contribution: The founding act of the data quality line: the first application of model checking to consistency verification of longitudinal archives. [read the paper · full reference]
• A PDDL+ Benchmark Problem: The Batch Chemical Plant (ICAPS 2010, 2010). Giuseppe Della Penna, Benedetto Intrigila, Daniele Magazzeni, Fabio Mercorio. In plain terms: Proposes the batch chemical plant as a benchmark problem for PDDL+ planning: a realistic testbed for hybrid planners. Contribution: Proposes the batch chemical plant as a PDDL+ benchmark: a reference problem for comparing hybrid planners. [read the paper · full reference]
• Planning for Autonomous Planetary Vehicles (ICAS 2010, 2010). Giuseppe Della Penna, Benedetto Intrigila, Daniele Magazzeni, Fabio Mercorio. In plain terms: Applies automated planning to autonomous planetary vehicles: how to plan the operations of a rover with limited resources. Contribution: Applies automated planning to planetary rovers: managing operations under resource constraints. [read the paper · full reference]
• Resource-Optimal Planning For An Autonomous Planetary Vehicle (arXiv (preprint), 2010). Giuseppe Della Penna, Benedetto Intrigila, Daniele Magazzeni, Fabio Mercorio. In plain terms: Preprint version of the work on resource-optimal planning for an autonomous planetary vehicle. [Preprint] Contribution: Preprint version of the work on resource-optimal planning for autonomous planetary vehicles. [read the paper · full reference]
• Planning for Continuous Domains (The AI*IA Doctoral Consortium, Brescia (Italy) December 1-3, 2010). Fabio Mercorio. In plain terms: A doctoral-consortium presentation of the research programme: planning in continuous and hybrid domains via discretisation and model checking, with validation of the resulting plans against the continuous dynamics. Contribution: Early statement of the discretise-and-validate approach that UPMurphi turned into a working universal planner. [read the paper · full reference]
• UPMurphi: A Tool for Universal Planning on PDDL+ Problems (ICAPS 2009, 2009). Giuseppe Della Penna, Daniele Magazzeni, Fabio Mercorio, Benedetto Intrigila. In plain terms: UPMurphi, the first tool of the research path: a universal planner for PDDL+ problems based on model checking, able to handle continuous dynamics through discretisation. Presented at ICAPS 2009, it is still cited today as a reference for hybrid planning. Contribution: UPMurphi: the first model-checking-based universal planner for PDDL+, built on a discretise-and-validate approach. It opened the way to planning over continuous domains and remains a reference in hybrid planning. [read the paper · full reference · UPMurphi]

Data Quality & Formal Methods (2011–2017 · 17 papers)

Data quality as a provable property: formal verification and repair of longitudinal data.

Where the field stood. In the early 2010s data cleansing was dominated by ad-hoc rules and static integrity constraints: for longitudinal data (working careers, clinical records), where consistency depends on the whole sequence of events, systematic and verifiable methods were missing.

Our contribution. The original idea is twofold: (1) formally model the admissible behaviour of the data and use model checking to uncover inconsistencies (IDA 2011); (2) cast data repair as an automated planning problem, letting the planner synthesise the optimal sequence of corrective actions ('Planning meets Data Cleansing', ICAPS 2014). Data quality as a provable property, not a heuristic.

Why it matters. Applied to the administrative archives of the Lombardy Region (ARIFL projects) and to healthcare data: automatic verification and cleansing of millions of working-career records. It is the bridge that took the group from formal methods to labour data — and hence to Labour Market Intelligence.

• An AI Planning System for Data Cleaning (ECML-PKDD 2017, 2017). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Shows that data cleaning can be framed as an AI planning problem: the system automatically finds the optimal sequence of corrections. A bridge between the group's two souls: planning and data quality. Contribution: Frames data cleaning as an AI planning problem: the planner automatically synthesises the optimal sequence of corrections. The formal bridge between the planning and data quality research lines. [read the paper · full reference]
• A Model-Based Approach for Developing Data Cleansing Solutions (ACM J. Data & Information Quality, 2015). Mario Mezzanzanica, Roberto Boselli, Mirko Cesarini, Fabio Mercorio. In plain terms: A model-based approach to developing data cleansing solutions: formalising the expected quality of the data in order to derive the cleaning activities automatically. Contribution: A model-based methodology for deriving data cleansing solutions from a formal specification of the expected data quality. [read the paper · full reference]
• A model-based evaluation of data quality activities in KDD (Inf. Processing & Management, 2015). Mario Mezzanzanica, Roberto Boselli, Mirko Cesarini, Fabio Mercorio. In plain terms: Uses formal methods to evaluate data quality activities in knowledge discovery processes: how much data cleansing improves (or distorts) the analyses downstream. Contribution: Formally quantifies the impact of data quality activities on KDD processes: how much cleansing improves (or distorts) downstream analyses. [read the paper · full reference]
• A Policy-Based Cleansing and Integration Framework for Labour and Healthcare Data (LNCS (Springer), 2014). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: A policy-based framework for cleansing and integrating labour and healthcare data: declarative rules that drive the correction of longitudinal administrative archives. Contribution: Policy-based framework for cleansing and integrating labour and healthcare data: declarative rules for correcting longitudinal archives. [read the paper · full reference]
• Accurate Data Cleansing through Model Checking and Machine Learning Techniques (DATA 2014, 2014). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Combines model checking and machine learning for more accurate data cleansing: formal methods detect the inconsistencies, ML learns how to fix them. Contribution: Combines model checking (to detect inconsistencies) with machine learning (to learn how to fix them), achieving higher accuracy than either approach alone. [read the paper · full reference]
• Are the Methodologies for Producing Linked Open Data Feasible for Public Administrations? (DATA 2014, 2014). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Assesses whether the methodologies for producing Linked Open Data are actually feasible for Italian public administrations: a critical analysis with recommendations. Contribution: Critical analysis of the feasibility of Linked Open Data methodologies for Italian public administrations, with practical recommendations. [read the paper]
• Data Quality on KDD: a Real-life Scenario (SEBD 2014, 2014). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Data quality in KDD processes told through a real-life scenario: the administrative labour market archives of the Lombardy Region. Contribution: Brings a real-life scenario (the Lombardy Region labour market archives) into the debate on data quality in KDD processes. [read the paper · full reference]
• Improving Data Cleansing Accuracy - A Model-based Approach (DATA 2014, 2014). Mario Mezzanzanica, Roberto Boselli, Mirko Cesarini, Fabio Mercorio. In plain terms: Improves the accuracy of data cleansing through an approach based on formal models of the expected behaviour of the data. Contribution: Model-based approach that improves cleansing accuracy by modelling the expected behaviour of the data. [read the paper · full reference]
• Longitudinal data consistency verification using formal methods (Int. J. Information Quality, 2014). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Verifies the consistency of longitudinal data (working careers) using formal methods: each trajectory is checked against a model of admissible behaviour. Contribution: Formal verification of longitudinal data consistency: each career checked against a model of admissible behaviour. [read the paper · full reference]
• Planning meets Data Cleansing (ICAPS 2014, 2014). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: 'Planning meets Data Cleansing', presented at ICAPS: data cleansing framed as an automated planning problem, with the planner synthesising the correction sequences. The conceptual bridge between the first and second phases of the group's research. Contribution: 'Planning meets Data Cleansing' at ICAPS: the first formulation of data cleansing as an automated planning problem, with automatic synthesis of correction sequences. The hinge paper between the formal and the data-driven phases of the research path. [read the paper · full reference]
• Towards data cleansing via planning (Intelligenza Artificiale, 2014). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Extended journal version of the idea of tackling data cleansing through automated planning. Contribution: Journal extension of the planning-based approach to data cleansing. [read the paper · full reference]
• Automatic Synthesis of Data Cleansing Activities (DATA 2013, 2013). Mario Mezzanzanica, Roberto Boselli, Mirko Cesarini, Fabio Mercorio. In plain terms: Automatically synthesises data cleansing activities from a formal model of the admissible evolution of the data. Contribution: Automatic synthesis of cleansing activities from a formal model of the admissible evolution of the data. [read the paper · full reference]
• Inconsistency Knowledge Discovery for Longitudinal Data Management: A Model-Based Approach (HCI-KDD 2013, 2013). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Discovers inconsistencies in longitudinal data through a model-based approach, applied to administrative archives. Contribution: Model-based discovery of inconsistencies in longitudinal data, applied to administrative archives. [read the paper · full reference]
• Can Planning meet Data Cleansing? (5th Italian Workshop on Planning and Scheduling at AIxIA …, 2013). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: The workshop paper that first asked the question: can a planner clean data? Inconsistencies in longitudinal records become states to escape, and cleansing actions become planning operators. Contribution: First formulation of data cleansing as automated planning — the direct precursor of 'Planning meets Data Cleansing' (ICAPS 2014). [read the paper · full reference]
• Data Quality Sensitivity Analysis on Aggregate Indicators (DATA 2012, 2012). Mario Mezzanzanica, Roberto Boselli, Mirko Cesarini, Fabio Mercorio. In plain terms: Analyses how sensitive aggregate statistical indicators are to the quality of the underlying data: small errors in the microdata can distort published statistics. Contribution: Sensitivity analysis of aggregate indicators to microdata quality: quantifies how errors propagate into published statistics. [read the paper · full reference]
• Towards the Use of Model Checking for Performing Data Consistency Evaluation and Cleansing (ICIQ 2012, 2012). Mario Mezzanzanica, Mirko Cesarini, Fabio Mercorio, Roberto Boselli. In plain terms: Proposes using model checking to evaluate and cleanse data consistency: formal software verification techniques applied to information archives. Contribution: Proposes model checking for evaluating and cleansing data consistency: formal software verification transferred to information archives. [full reference]
• Data Quality through Model Checking Techniques (IDA 2011, 2011). Mario Mezzanzanica, Roberto Boselli, Mirko Cesarini, Fabio Mercorio. In plain terms: The first work to bring model checking into data quality: formally verifying the consistency of longitudinal data archives. The founding act of the group's DQ research line. Contribution: The founding act of the data quality line: the first application of model checking to consistency verification of longitudinal archives. [read the paper · full reference]

Labour Market Intelligence (2014–present · 41 papers)

Turning millions of online job ads into real-time labour-market statistics for Europe.

Where the field stood. In the mid-2010s labour statistics came from slow, expensive sample surveys; online job advertisements (OJAs) were a real-time goldmine, but nobody knew how to classify them reliably against official taxonomies (ISCO/ESCO), across languages and at continental scale.

Our contribution. The group built the full pipeline: WoLMIS (2018) for the automatic classification of job ads, NEO (AAAI 2021) for discovering emerging occupations, SkiLLMo and SkiLLens for extracting and mapping skills onto ESCO with Transformers, up to 'Terminator Economy' (IJCAI 2025), which measures the exposure of jobs to AI task by task. All of it scaled to 28 countries within the Cedefop/Eurostat projects, where Mercorio leads the research unit.

Why it matters. This is the European infrastructure (Web Intelligence Hub) that lets Eurostat, Cedefop and national governments read labour demand in real time: which occupations are growing, which skills are needed, where AI is reshaping jobs. Awarded by the Accademia dei Lincei (2019); TerminatorEconomy.com was covered by Repubblica, Wired and Radio24.

Open-source tools: Terminator-Economy

• SkiLLens: Recognising and Mapping Novel Skills from Millions of Job Ads Across Europe Using Language Models (EACL 2026, 2026). Alessia De Santo, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani. In plain terms: A system that reads millions of European job ads and discovers the new skills the market is asking for (e.g. new tools, new practices), automatically linking them to the European ESCO taxonomy with the help of LLMs and human review. Designed for those who shape training and skills policies. Contribution: The first multilingual human-in-the-loop pipeline to close the entire loop: extraction of candidate skills from national job-ad corpora, entity linking to ESCO via encoders and refinement with decoder LLMs, validated both quantitatively and qualitatively. [read the paper · full reference]
• VEUCTOR: Training and selecting best vector space models from online job ads for European countries (Information Sciences, 2026). Emilio Colombo, Simone D'Amico, Fabio Mercorio, Mario Mezzanzanica. In plain terms: VEUCTOR trains and compares over 3,000 word embedding models on 4.5 million job ads across 28 European countries, and releases the best ones already aligned with each other. It demonstrates that the choice of embedding concretely changes the results of labour market analyses. Contribution: Unprecedented scale: 3,000+ vector space models trained and selected on 4.5M job ads across 28 countries, with the first systematic evidence that the choice of embedding alters the conclusions of labour market analyses. [read the paper · full reference]
• Terminator Economy: Assessing jobs and tasks exposure to AI (ECML-PKDD 2026, 2026). Fabio Mercorio, Mario Mezzanzanica, Daniele Potertì, Antonio Serino, Andrea Seveso. In plain terms: How exposed is each job — and each single task — to today's AI? The paper presents the Terminator Economy framework at ECML-PKDD: open-source LLMs assess AI's current ability to perform job tasks, producing the TEAI (Task Exposure to AI) and TRAI (Task Replacement by AI) indices, validated with human evaluators. About one third of U.S. employment turns out to be highly exposed — mostly high-skill jobs — yet exposure correlates with employment and wage growth, suggesting AI complements more than it replaces. Contribution: Brings the TEAI/TRAI framework to the European ML community as a reproducible, open benchmark: a task-level, LLM-based measure of AI exposure that can be recomputed over time to monitor AI's progress on the labour market. [full reference · Terminator-Economy]
• Using Language Models to Extract and Map Emerging Skills to the ESCO Taxonomy (Discover Artificial Intelligence, 2026). Alessia De Santo, Navid Nobani, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: New skills appear in job ads long before they enter official taxonomies. SkiLLink is a two-stage pipeline that (i) spots emerging skills in online job ads using embeddings and language models, and (ii) maps them onto the ESCO taxonomy via embedding-based recommendations and LLMs — validated by labour-market experts across several EU countries and stress-tested against thirteen different distance metrics. Contribution: An end-to-end, expert-validated pipeline for discovering emerging skills and aligning them with ESCO, with a robustness analysis showing the alignment is stable across similarity measures. [read the paper · full reference]
• JobSet: Synthetic Job Advertisements Dataset for Labour Market Intelligence (ACM SAC 2025, 2025). Samuele Colombo, Simone D'Amico, Lorenzo Malandri, Fabio Mercorio, Andrea Seveso. In plain terms: JobGen uses LLMs to generate synthetic yet realistic job advertisements, calibrated on the real distributions of a European project. The result is JobSet, an open dataset that makes it possible to train and test skill extraction systems without privacy concerns. Contribution: JobGen: the first LLM-based generator of synthetic job ads calibrated on the real distributions of a European project, producing the open JobSet dataset for skill intelligence. [read the paper · full reference]
• SkiLLMo: Normalized ESCO Skill Extraction through Transformer Models (ACM SAC 2025, 2025). Lorenzo Malandri, Fabio Mercorio, Antonio Serino. In plain terms: SkiLLMo uses Transformer models to extract skills from the text of job advertisements and normalise them against the European ESCO taxonomy, making job ads written in vastly different ways comparable. Contribution: Skill extraction normalised directly onto ESCO via Transformers: it merges recognition and standardisation into a single step, where previous methods kept them separate. [read the paper · full reference]
• Taxonomy Expansion through Collaborative LLM Mapping (ACM SAC 2025, 2025). Simone D'Amico, Alessia De Santo, Mario Mezzanzanica, Fabio Mercorio. In plain terms: TAXMAP has three LLMs collaborate to propose new terms to add to a taxonomy (here, ESCO), with final human validation. On 40,000 digital terms collected from the web, it proposed 924 additions, 81% of which were judged correct by experts. Contribution: TAXMAP has three generative LLMs collaborate as a 'committee' for taxonomy expansion with final human validation: 81% PPV on 40,000 real-world terms for ESCO, drastically reducing the expert workload. [read the paper · full reference]
• Towards the Terminator Economy: Assessing Job Exposure to AI Through LLMs (IJCAI 2025, 2025). Emilio Colombo, Fabio Mercorio, Mario Mezzanzanica, Antonio Serino. In plain terms: How exposed is each job to AI? The paper builds a pipeline in which LLMs themselves assess, task by task, what they can already do of real occupations. The result is TerminatorEconomy.com, cited by Repubblica, Wired, and Radio24. Contribution: A two-stage methodology that uses LLMs themselves to assess, task by task, the exposure of occupations to AI: an empirical, reproducible measure where before there were only expert estimates, released as an explorable tool (TerminatorEconomy.com). [read the paper · full reference · Terminator-Economy]
• Alignment of Multilingual Embeddings to Estimate Job Similarities in Online Labour Market (IEEE DSAA 2024, 2024). Simone D'Amico, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Filippo Pallucchini. In plain terms: How can occupations be compared across countries that speak different languages? The paper aligns multilingual embeddings of job advertisements to estimate the similarity between jobs in different national markets. Contribution: Estimates similarities between occupations across different markets and languages by aligning multilingual embeddings trained on real job ads: a practical method for cross-country comparison. [read the paper · full reference]
• Enriching Skill Taxonomies through Vector Space Models (IEEE BigData 2024, 2024). Simone D'Amico, Alessia De Santo, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Shows how to enrich skill taxonomies (such as ESCO) using vector space models trained on job advertisements, surfacing new terms from the bottom up. Contribution: Shows how to use vector space models to enrich skill taxonomies with terms that emerge from the bottom up, on real European data. [read the paper · full reference]
• Towards the Terminator Economy: Assessing Job Exposure to AI through LLMs (arXiv (preprint), 2024). Emilio Colombo, Fabio Mercorio, Mario Mezzanzanica, Antonio Serino. In plain terms: Preprint version of the 'Terminator Economy' study on job exposure to AI. [Preprint] Contribution: Preprint version of the Terminator Economy study on job exposure to AI. [read the paper · full reference · Terminator-Economy]
• GraphLMI: A data driven system for exploring labor market information through graph databases (Multimedia Tools & Applications, 2022). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: GraphLMI turns the labour market into a navigable graph: occupations, skills and job ads become queryable nodes and relations, making it possible to explore supply and demand through visual queries. Contribution: GraphLMI brings the graph paradigm into Labour Market Intelligence: occupations, skills and job ads as queryable nodes, enabling relational queries that tabular systems cannot express. [read the paper · full reference]
• JoTA: Aligning Multilingual Job Taxonomies through Word Embeddings (Student Abstract) (AAAI 2022, 2022). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: JoTA aligns job taxonomies across different languages using word embeddings, making it possible to compare national classifications without manual translations. [Student abstract] Contribution: JoTA aligns multilingual job taxonomies via word embeddings, removing the need for manual mappings between national classifications. [read the paper]
• WETA: Automatic taxonomy alignment via word embeddings (Computers in Industry, 2022). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: WETA automatically aligns two taxonomies (e.g. a national one and ESCO) by exploiting word embeddings: it finds correspondences between categories without manual mappings. Contribution: WETA automates taxonomy alignment through embeddings: correspondences between categories are discovered from data rather than compiled by hand. [read the paper · full reference]
• Skills Mismatch, Automation, and Training: Evidence from 17 European Countries Using Survey Data and Online Job Ads (CESifo Forum, 2022). Yuchen Guo, Christina Langer, Fabio Mercorio, Francesco Trentini. In plain terms: What do skill mismatch and automation look like when you observe 17 European countries through both official surveys and millions of online job ads? This policy study combines the two lenses to show where demand and training are misaligned, and how automation pressure interacts with the skills workers actually have. Contribution: Cross-country evidence on skills mismatch and automation that pairs survey data with OJA-based indicators — an example of Labour Market Intelligence informing the policy debate (CESifo Forum). [read the paper · full reference]
• A Human-AI Teaming Approach for Incremental Taxonomy Learning from Text (IJCAI 2021, 2021). Andrea Seveso, Fabio Mercorio, Mario Mezzanzanica. In plain terms: A human-AI 'teaming' approach to growing a taxonomy from text: the system proposes new concepts and where to place them, the expert validates, and the cycle repeats incrementally. Contribution: Formalises incremental taxonomy learning as a cyclic human-AI collaboration: the system proposes, the expert validates, and the taxonomy grows without rebuilding from scratch. [read the paper]
• A Method for Taxonomy-Aware Embeddings Evaluation (Student Abstract) (AAAI 2021, 2021). Navid Nobani, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: How do you choose the right embedding when the data has a taxonomic structure? This work proposes a taxonomy-aware evaluation metric. [Student abstract] Contribution: A taxonomy-aware embedding evaluation metric: it measures how well hierarchies are preserved, not just word analogies. [read the paper]
• MEET-LM: A method for embeddings evaluation for taxonomic data in the labour market (Computers in Industry, 2021). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani. In plain terms: MEET-LM is a method for assessing how well an embedding model preserves the relations of a labour market taxonomy: it helps choose the right model before building analyses and systems on top of it. Contribution: MEET-LM specialises taxonomy-based embedding evaluation for the labour domain: an operational criterion for choosing the model on which to build European LMI systems. [read the paper · full reference]
• NEO: A System for Identifying New Emerging Occupation from Job Ads (AAAI 2021, 2021). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: NEO discovers emerging occupations by reading online job ads: when an occupation appears that the official taxonomy does not yet cover (e.g. 'prompt engineer'), NEO detects it and suggests where it should be placed. Contribution: NEO is the first system that detects emerging occupations from online job ads and suggests where to place them in the official taxonomy, turning the updating of classifications from a multi-year expert process into a data-driven pipeline. [read the paper · full reference]
• Skills2Graph: Processing million Job Ads to face the Job Skill Mismatch Problem (IJCAI 2021, 2021). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: Skills2Graph processes millions of job ads to tackle the mismatch between the skills companies demand and those workers offer, representing them in a queryable graph. Contribution: Skills2Graph tackles skill mismatch by processing millions of job ads into a queryable skill-occupation graph at European scale. [read the paper · full reference]
• Skills2Job: A Recommender System that Encodes Job Offer Embeddings on Graph Databases (Student Abstract) (AAAI 2021, 2021). Andrea Seveso, Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Skills2Job recommends job offers by encoding job ads as embeddings within a graph database. [Student abstract] Contribution: Skills2Job encodes job offer embeddings directly in the graph database: recommendation that combines vector semantics with relational structure. [read the paper]
• Skills2Job: A recommender system that encodes job offer embeddings on graph databases (Applied Soft Computing, 2021). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: Extended version of Skills2Job: a recommender system that combines semantic embeddings and graph databases to suggest the job offers that best match a person's skill profile. Contribution: Extended version of Skills2Job: the embedding-on-graph integration for job recommendation, validated on real-world data. [read the paper · full reference]
• TaxoRef: Embeddings Evaluation for AI-driven Taxonomy Refinement (ECML-PKDD 2021, 2021). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani. In plain terms: TaxoRef assesses which embedding model best represents an existing taxonomy and then uses it to refine that taxonomy: it points out where the hierarchy is inconsistent and how to fix it. Presented at ECML-PKDD. Contribution: TaxoRef reverses the perspective: embeddings are not only evaluated against the taxonomy, they become the instrument for refining it, flagging hierarchical inconsistencies and proposing corrections. [read the paper · full reference · TaxoRef]
• MEET: A Method for Embeddings Evaluation for Taxonomic Data (ICDM Workshops 2020, 2020). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani. In plain terms: MEET introduces the method for evaluating embeddings on taxonomic data, later specialised for the labour market (MEET-LM): it measures how faithfully the vector space respects concept hierarchies. Contribution: MEET introduces the framework for evaluating embeddings on taxonomic data, subsequently adopted throughout the entire follow-up line of research. [read the paper · full reference]
• NEO: A Tool for Taxonomy Enrichment with New Emerging Occupations (ISWC 2020, 2020). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: The first presentation of NEO to the Semantic Web community: the tool that enriches taxonomies with new occupations emerging from online job ads. Contribution: First presentation of NEO to the Semantic Web community: taxonomy enrichment with emerging occupations drawn from real job advertisements. [read the paper · full reference]
• Big Data as Fuel of Skill Intelligence (Encyclopedia of Big Data Technologies, 2020). Mario Mezzanzanica, Fabio Mercorio. In plain terms: An encyclopedia entry that defines 'skill intelligence': how big data from online job ads can be turned into timely, granular knowledge about the skills the labour market actually demands, and what it takes — methodologically — to do it right. Contribution: Consolidates the group's OJA methodology into a reference definition of skill intelligence for the Encyclopedia of Big Data Technologies. [read the paper · full reference]
• Big Data Enables Labor Market Intelligence (Encyclopedia of Big Data Technologies, 2019). Mario Mezzanzanica, Fabio Mercorio. In plain terms: The 'Big Data Enables Labor Market Intelligence' entry in the Encyclopedia of Big Data Technologies: it defines the field of LMI and establishes its core concepts and reference architectures. Contribution: The encyclopedia entry that formally defines Labour Market Intelligence: the concepts, architectures, and research agenda of the field. [read the paper · full reference]
• Towards Labour Market Intelligence through Topic Modelling (HICSS 2019, 2019). Francesco Colace, Massimo De Santo, Marco Lombardi, Fabio Mercorio, Mario Mezzanzanica, Francesco Pascale. In plain terms: Applies topic modelling to job advertisements to surface the latent themes of labour demand, with a focus on the interpretability of the results. Contribution: Applies topic modelling to labour demand while keeping the topics interpretable: an early bridge between LMI and explainability. [read the paper · full reference]
• AI meets labor market: Exploring the link between automation and skills (Information Economics and Policy, 2019). Emilio Colombo, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Which skills protect a job from automation? Analysing millions of online job ads, the paper links the skill content of occupations to their automation risk, showing that digital and soft skills reshape — rather than simply destroy — the demand for work. Contribution: One of the first large-scale, OJA-based analyses of the automation–skills link (Information Economics and Policy): an empirical bridge between Labour Market Intelligence and the economics of automation, and the conceptual ancestor of the Terminator Economy line. [read the paper · full reference]
• Classifying online Job Advertisements through Machine Learning (Future Generation Computer Systems, 2018). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Shows at scale how to classify online job advertisements with machine learning against standard taxonomies: one of the foundational works of the LMI research line. Contribution: Demonstrates, at real-world scale, the ML classification of job advertisements onto standard taxonomies: foundational evidence for OJA-based labour market observatories. [read the paper · full reference]
• Multimedia story creation on social networks (Future Generation Computer Systems, 2018). Flora Amato, Aniello Castiglione, Fabio Mercorio, Mario Mezzanzanica, Vincenzo Moscato, Antonio Picariello, Giancarlo Sperlì. In plain terms: A system that automatically creates multimedia 'stories' by aggregating social network content around events and narratives. Contribution: Automatic generation of multimedia 'stories' by aggregating social content around events: data-driven narrative composition. [read the paper · full reference]
• Skills in demand for ICT and statistical occupations: Evidence from web-based job vacancies (Statistical Analysis & Data Mining, 2018). Pietro Giorgio Lovaglio, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Analyses the demand for ICT and statistical occupations through web job vacancies: which skills the market actually asks for, with quantitative evidence on Italy. Contribution: The first systematic quantification of the demand for ICT and statistical occupations in Italy through web vacancies, with evidence on the skills required. [read the paper · full reference]
• WoLMIS: a labor market intelligence system for classifying web job vacancies (J. Intelligent Information Systems, 2018). Roberto Boselli, Mirko Cesarini, Stefania Marrara, Fabio Mercorio, Mario Mezzanzanica, Gabriella Pasi, Marco Viviani. In plain terms: WoLMIS is the Labour Market Intelligence system that automatically classifies web job vacancies onto standard taxonomies: the architecture that laid the foundations for the collaboration with Cedefop and for scaling up to the European level. Contribution: WoLMIS is the reference architecture for the automatic classification of job vacancies onto standard taxonomies: the system that made LMI credible at national and then European scale (Cedefop). [read the paper · full reference]
• A language modelling approach for discovering novel labour market occupations from the web (WI 2017, 2017). Stefania Marrara, Gabriella Pasi, Marco Viviani, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica, Marco Pappagallo. In plain terms: Uses language modelling techniques to discover, from the web, new occupations not yet codified in official classifications: a precursor of NEO. Contribution: Uses language modelling to discover occupations not yet codified, straight from the web: the idea that would later mature into NEO. [read the paper · full reference]
• Labour Market Intelligence for Supporting Decision Making (Proceedings of the 25th Italian Symposium on Adv, 2017). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: An Italian overview of Labour Market Intelligence as decision support: data, methods, and use cases for policy makers. Contribution: Systematises LMI as a decision-support instrument for policy makers: data, methods, and use cases.
• Un'Analisi del Mondo del Lavoro e un Modello Predittivo per Potenziali Nuove Occupazioni (An Analysis of the Job Market and a Predictive Model for Potential New Jobs) (IIR 2017, 2017). Gabriella Pasi, Mirko Cesarini, Stefania Marrara, Fabio Mercorio, Marco Viviani, Mario Mezzanzanica, Marco Pappagallo. In plain terms: An analysis of the Italian labour market and a predictive model for identifying potential new occupations, presented to the Italian IR community. Contribution: A predictive model for identifying potential new occupations in the Italian market: data-driven anticipation of occupational change.
• Using Machine Learning for Labour Market Intelligence (ECML-PKDD 2017, 2017). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: The ECML-PKDD paper that introduced the European ML community to the use of machine learning for Labour Market Intelligence at real-world scale: classifying job vacancies for labour market observatories. Contribution: Brings Labour Market Intelligence to the European ML community (ECML-PKDD): the application case that demonstrates the value of machine learning for labour market observatories. [read the paper · full reference]
• Analysing the Relevance of ICT Skills on occupations in Web Job Vacancies (Digital (R)Evolution and Its Effects on Labour: Opportuni…, 2016). Silvia Dusi, Matteo Fontana, Fabio Mercorio, Mario Mezzanzanica. In plain terms: An early demonstration of what job-ad data can reveal: which ICT skills are requested across occupations in web job vacancies, and how digital skills spread beyond strictly technical roles. Contribution: Early empirical evidence of ICT-skill demand from web vacancies — groundwork for the taxonomy-based skill extraction developed in the following years. [full reference]
• Challenge: Processing web texts for classifying job offers (IEEE ICSC 2015, 2015). Flora Amato, Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica, Vincenzo Moscato, Fabio Persia, Antonio Picariello. In plain terms: The 'challenge' of processing web texts to classify job offers: problems, datasets and first solutions. Contribution: Defines the challenge of processing web texts for job offer classification: problems, datasets, first solutions. [read the paper · full reference]
• Classification of Web Job Advertisements: A Case Study (SEBD 2015, 2015). Flora Amato, Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica, Vincenzo Moscato, Fabio Persia, Antonio Picariello. In plain terms: A case study on classifying web job advertisements: the early experiments that launched the group's LMI research line. Contribution: The group's first experiments on classifying web job ads: the empirical starting point of the LMI research line. [read the paper · full reference]
• A Policy-Based Cleansing and Integration Framework for Labour and Healthcare Data (LNCS (Springer), 2014). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: A policy-based framework for cleansing and integrating labour and healthcare data: declarative rules that drive the correction of longitudinal administrative archives. Contribution: Policy-based framework for cleansing and integrating labour and healthcare data: declarative rules for correcting longitudinal archives. [read the paper · full reference]

Graph Databases & Network Analysis (2017–2022 · 8 papers)

Science and the labour market as queryable graphs: communities, paths, bridging skills.

Where the field stood. Complex relational data (scientific collaborations, labour markets) was analysed with traditional databases, ill-suited to querying paths, communities and multi-hop relationships; graph databases were emerging, but mature scientific use cases were lacking.

Our contribution. GraphDBLP (2018) turns the entire computer-science bibliography into a queryable graph (700,000+ authors); DICO (IEEE TETC 2021) discovers scientific communities in it at scale; GraphLMI (2022) applies the same paradigm to the labour market, connecting occupations, skills and job ads.

Why it matters. It answers questions a tabular database cannot: 'who connects two research communities?', 'which skills bridge different occupations?'. The Skills2Job systems also use graphs to recommend job offers.

Open-source tools: GraphDBLP

• GraphLMI: A data driven system for exploring labor market information through graph databases (Multimedia Tools & Applications, 2022). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: GraphLMI turns the labour market into a navigable graph: occupations, skills and job ads become queryable nodes and relations, making it possible to explore supply and demand through visual queries. Contribution: GraphLMI brings the graph paradigm into Labour Market Intelligence: occupations, skills and job ads as queryable nodes, enabling relational queries that tabular systems cannot express. [read the paper · full reference]
• DICO: A Graph-DB Framework for Community Detection on Big Scholarly Data (IEEE TETC, 2021). Fabio Mercorio, Mario Mezzanzanica, Vincenzo Moscato, Antonio Picariello, Giancarlo Sperlí. In plain terms: DICO discovers the scientific communities hidden in big scholarly data using graph databases: who collaborates with whom, around which topics, with algorithms that scale to millions of publications. Contribution: DICO scales community detection over big scholarly data inside a native graph database: scientific communities discovered across millions of nodes without intermediate data extraction. [read the paper · full reference]
• Skills2Job: A Recommender System that Encodes Job Offer Embeddings on Graph Databases (Student Abstract) (AAAI 2021, 2021). Andrea Seveso, Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Skills2Job recommends job offers by encoding job ads as embeddings within a graph database. [Student abstract] Contribution: Skills2Job encodes job offer embeddings directly in the graph database: recommendation that combines vector semantics with relational structure. [read the paper]
• Skills2Job: A recommender system that encodes job offer embeddings on graph databases (Applied Soft Computing, 2021). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: Extended version of Skills2Job: a recommender system that combines semantic embeddings and graph databases to suggest the job offers that best match a person's skill profile. Contribution: Extended version of Skills2Job: the embedding-on-graph integration for job recommendation, validated on real-world data. [read the paper · full reference]
• A Tool for Researchers: Querying Big Scholarly Data Through Graph Databases (ECML-PKDD 2019, 2019). Fabio Mercorio, Mario Mezzanzanica, Vincenzo Moscato, Antonio Picariello, Giancarlo Sperlì. In plain terms: A demo of the tool that lets researchers query big scholarly data as a graph: collaboration paths, communities, and the evolution of research topics. Contribution: A demonstration tool that makes big scholarly data queryable as a graph by researchers, with no database expertise required. [read the paper · full reference · GraphDBLP]
• GraphDBLP Released: Querying the Computer Scientists Network as a Graph (Proceedings of the 26th Italian Symposium on Adv, 2018). Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica, Vincenzo Moscato, Antonio Picariello. In plain terms: The Italian presentation of the GraphDBLP release: the network of computer scientists made queryable as a graph. Contribution: The public release of GraphDBLP, presented to the Italian database community. [read the paper · full reference · GraphDBLP]
• GraphDBLP: a system for analysing networks of computer scientists through graph databases - GraphDBLP (Multimedia Tools & Applications, 2018). Mario Mezzanzanica, Fabio Mercorio, Mirko Cesarini, Vincenzo Moscato, Antonio Picariello. In plain terms: GraphDBLP turns DBLP (the computer science bibliography) into a graph database: 700,000+ authors and millions of relationships, queryable for collaboration paths, skills, and venues. Released as an open-source tool. Contribution: GraphDBLP is the first complete transformation of DBLP into an open-source graph database: 700,000+ authors queryable by collaboration paths, skills, and venues. [read the paper · full reference · GraphDBLP]
• A Pipeline for Multimedia Twitter Analysis through Graph Databases: Preliminary Results (DATA 2017, 2017). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica, Alessandro Vaccarino. In plain terms: A pipeline for analysing multimedia Twitter streams through graph databases: first results on storing and querying tweets as a graph. Contribution: A pipeline for analysing multimedia Twitter streams via graph databases: storing and querying tweets as a graph. [read the paper · full reference]

Embeddings & Taxonomy Learning (2018–present · 26 papers)

Evaluating and aligning word embeddings to keep official taxonomies alive.

Where the field stood. With word2vec and its successors, words became vectors — but how do you choose the right model for your domain? And how do you keep official taxonomies (ESCO, ISCO) up to date, hand-curated by experts and chronically behind the market's real language?

Our contribution. A family of methods for evaluating embeddings against taxonomic structures (MEET, TaxoRef, vec2best with the PCE metric) and for using them in taxonomy learning: enrichment (NEO), alignment across languages and classifications (WETA, JoTA, SeNSe), up to systematic model selection across 28 countries (VEUCTOR, 3,000+ trained models). More recently, a reference survey on cross-lingual alignment (ACM CSUR 2026) and XAI for embeddings.

Why it matters. These are the tools that keep European taxonomies 'alive': they surface emerging terms bottom-up, align classifications across countries, and tell you which vector model to use for each task. Most of it is open source (vec2best, SeNSe, TaxoRef).

Open-source tools: vec2best, SeNSe, TaxoRef

• Lost in Alignment: A Survey on Cross-Lingual Alignment Methods for Contextualized Representation (ACM Computing Surveys, 2026). Filippo Pallucchini, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: A comprehensive map of methods for 'aligning' linguistic representations across languages — that is, for making a model understand that 'cane' and 'dog' occupy the same place in the space of meanings. Useful for anyone building multilingual systems. Contribution: A reference survey (ACM CSUR) that organises for the first time the cross-lingual alignment methods for contextualized representations, resolving the fragmentation of the field. [read the paper · full reference]
• VEUCTOR: Training and selecting best vector space models from online job ads for European countries (Information Sciences, 2026). Emilio Colombo, Simone D'Amico, Fabio Mercorio, Mario Mezzanzanica. In plain terms: VEUCTOR trains and compares over 3,000 word embedding models on 4.5 million job ads across 28 European countries, and releases the best ones already aligned with each other. It demonstrates that the choice of embedding concretely changes the results of labour market analyses. Contribution: Unprecedented scale: 3,000+ vector space models trained and selected on 4.5M job ads across 28 countries, with the first systematic evidence that the choice of embedding alters the conclusions of labour market analyses. [read the paper · full reference]
• Benchmarking Distributional Vector Similarity Measures: A Survey (Expert Systems, 2026). Erik Cambria, Navid Nobani, Filippo Pallucchini, Fabio Mercorio. In plain terms: Almost every NLP system compares word vectors with cosine similarity — usually by habit, not by evidence. This survey builds a taxonomy of distributional similarity measures and benchmarks 15 of them across four tasks, multiple datasets and embedding models, showing that the best measure varies substantially by task: cosine similarity is not universally optimal. Contribution: First systematic taxonomy and empirical benchmark of vector similarity measures for NLP, turning a default choice into an informed, task-dependent one. [read the paper · full reference]
• Using Language Models to Extract and Map Emerging Skills to the ESCO Taxonomy (Discover Artificial Intelligence, 2026). Alessia De Santo, Navid Nobani, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: New skills appear in job ads long before they enter official taxonomies. SkiLLink is a two-stage pipeline that (i) spots emerging skills in online job ads using embeddings and language models, and (ii) maps them onto the ESCO taxonomy via embedding-based recommendations and LLMs — validated by labour-market experts across several EU countries and stress-tested against thirteen different distance metrics. Contribution: An end-to-end, expert-validated pipeline for discovering emerging skills and aligning them with ESCO, with a robustness analysis showing the alignment is stable across similarity measures. [read the paper · full reference]
• SeNSe: embedding alignment via semantic anchors selection (Int. J. Data Science & Analytics, 2025). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Filippo Pallucchini. In plain terms: SeNSe aligns two embedding spaces (for instance, of two languages or two time periods) by automatically selecting the most reliable anchor words, with no predefined dictionaries. It beats the state of the art in building bilingual lexicons. Contribution: SeNSe selects anchors by considering the context of use (the neighbours), not just the word itself: unsupervised alignment without a priori dictionaries that outperforms the state of the art on bilingual lexicon extraction. [read the paper · full reference · SeNSe]
• SkiLLMo: Normalized ESCO Skill Extraction through Transformer Models (ACM SAC 2025, 2025). Lorenzo Malandri, Fabio Mercorio, Antonio Serino. In plain terms: SkiLLMo uses Transformer models to extract skills from the text of job advertisements and normalise them against the European ESCO taxonomy, making job ads written in vastly different ways comparable. Contribution: Skill extraction normalised directly onto ESCO via Transformers: it merges recognition and standardisation into a single step, where previous methods kept them separate. [read the paper · full reference]
• A Fistful of Vectors: A Tool for Intrinsic Evaluation of Word Embeddings (Cognitive Computation, 2024). Roberto Ascari, Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: vec2best brings together the main 'intrinsic' evaluation metrics for word embeddings in a single Python tool and proposes a unified score (PCE) for choosing the best model for one's task. Contribution: vec2best unifies intrinsic evaluation metrics in a single Python framework and proposes the aggregated PCE score: the first opportunity for standardised comparison across embeddings. [read the paper · full reference · vec2best]
• Alignment of Multilingual Embeddings to Estimate Job Similarities in Online Labour Market (IEEE DSAA 2024, 2024). Simone D'Amico, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Filippo Pallucchini. In plain terms: How can occupations be compared across countries that speak different languages? The paper aligns multilingual embeddings of job advertisements to estimate the similarity between jobs in different national markets. Contribution: Estimates similarities between occupations across different markets and languages by aligning multilingual embeddings trained on real job ads: a practical method for cross-country comparison. [read the paper · full reference]
• Enriching Skill Taxonomies through Vector Space Models (IEEE BigData 2024, 2024). Simone D'Amico, Alessia De Santo, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Shows how to enrich skill taxonomies (such as ESCO) using vector space models trained on job advertisements, surfacing new terms from the bottom up. Contribution: Shows how to use vector space models to enrich skill taxonomies with terms that emerge from the bottom up, on real European data. [read the paper · full reference]
• XAI meets LLMs: A Survey of the Relation between Explainable AI (arXiv (preprint), 2024). Erik Cambria, Lorenzo Malandri, Fabio Mercorio, Navid Nobani, Andrea Seveso. In plain terms: A survey on the relationship between XAI and LLMs: how large language models can generate explanations and how, in turn, they themselves need to be explained. [Preprint] Contribution: The first systematic mapping of the bidirectional XAI↔LLM relationship: models as generators of explanations and as objects to be explained. [read the paper · full reference]
• KRAKEN: A Novel Semantic-Based Approach for Keyphrases Extraction (IC3K 2023, 2023). Simone D'Amico, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: KRAKEN extracts keyphrases from a document based on semantics rather than mere word frequency, capturing important concepts even when they are mentioned only once. Contribution: Keyphrase extraction driven by semantics rather than frequency: it captures relevant concepts even when they appear only once. [read the paper · full reference]
• SEEDOT: Tool for Enhancing Sentiment Lexicon with Machine Learning (ECML-PKDD 2023, 2023). Vittorio Haardt, Lorenzo Malandri, Fabio Mercorio, Luca Porcelli. In plain terms: SEEDOT automatically enriches sentiment lexicons (the lists of positive/negative words) using machine learning, reducing the manual annotation effort. Contribution: SEEDOT automates the enrichment of sentiment lexicons with ML, cutting down the manual annotation required by traditional lexicon-based approaches. [read the paper · full reference]
• Embeddings Evaluation Using a Novel Measure of Semantic Similarity (Cognitive Computation, 2022). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani. In plain terms: Proposes a new measure of semantic similarity for evaluating word embeddings, checking how faithfully they preserve the relations of a reference taxonomy. Contribution: A new semantic similarity measure for embedding evaluation, anchored to taxonomic relations rather than human similarity judgements alone. [read the paper · full reference]
• JoTA: Aligning Multilingual Job Taxonomies through Word Embeddings (Student Abstract) (AAAI 2022, 2022). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: JoTA aligns job taxonomies across different languages using word embeddings, making it possible to compare national classifications without manual translations. [Student abstract] Contribution: JoTA aligns multilingual job taxonomies via word embeddings, removing the need for manual mappings between national classifications. [read the paper]
• WETA: Automatic taxonomy alignment via word embeddings (Computers in Industry, 2022). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: WETA automatically aligns two taxonomies (e.g. a national one and ESCO) by exploiting word embeddings: it finds correspondences between categories without manual mappings. Contribution: WETA automates taxonomy alignment through embeddings: correspondences between categories are discovered from data rather than compiled by hand. [read the paper · full reference]
• A Method for Taxonomy-Aware Embeddings Evaluation (Student Abstract) (AAAI 2021, 2021). Navid Nobani, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: How do you choose the right embedding when the data has a taxonomic structure? This work proposes a taxonomy-aware evaluation metric. [Student abstract] Contribution: A taxonomy-aware embedding evaluation metric: it measures how well hierarchies are preserved, not just word analogies. [read the paper]
• MEET-LM: A method for embeddings evaluation for taxonomic data in the labour market (Computers in Industry, 2021). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani. In plain terms: MEET-LM is a method for assessing how well an embedding model preserves the relations of a labour market taxonomy: it helps choose the right model before building analyses and systems on top of it. Contribution: MEET-LM specialises taxonomy-based embedding evaluation for the labour domain: an operational criterion for choosing the model on which to build European LMI systems. [read the paper · full reference]
• NEO: A System for Identifying New Emerging Occupation from Job Ads (AAAI 2021, 2021). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: NEO discovers emerging occupations by reading online job ads: when an occupation appears that the official taxonomy does not yet cover (e.g. 'prompt engineer'), NEO detects it and suggests where it should be placed. Contribution: NEO is the first system that detects emerging occupations from online job ads and suggests where to place them in the official taxonomy, turning the updating of classifications from a multi-year expert process into a data-driven pipeline. [read the paper · full reference]
• Skills2Graph: Processing million Job Ads to face the Job Skill Mismatch Problem (IJCAI 2021, 2021). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: Skills2Graph processes millions of job ads to tackle the mismatch between the skills companies demand and those workers offer, representing them in a queryable graph. Contribution: Skills2Graph tackles skill mismatch by processing millions of job ads into a queryable skill-occupation graph at European scale. [read the paper · full reference]
• Skills2Job: A Recommender System that Encodes Job Offer Embeddings on Graph Databases (Student Abstract) (AAAI 2021, 2021). Andrea Seveso, Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Skills2Job recommends job offers by encoding job ads as embeddings within a graph database. [Student abstract] Contribution: Skills2Job encodes job offer embeddings directly in the graph database: recommendation that combines vector semantics with relational structure. [read the paper]
• Skills2Job: A recommender system that encodes job offer embeddings on graph databases (Applied Soft Computing, 2021). Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: Extended version of Skills2Job: a recommender system that combines semantic embeddings and graph databases to suggest the job offers that best match a person's skill profile. Contribution: Extended version of Skills2Job: the embedding-on-graph integration for job recommendation, validated on real-world data. [read the paper · full reference]
• TaxoRef: Embeddings Evaluation for AI-driven Taxonomy Refinement (ECML-PKDD 2021, 2021). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani. In plain terms: TaxoRef assesses which embedding model best represents an existing taxonomy and then uses it to refine that taxonomy: it points out where the hierarchy is inconsistent and how to fix it. Presented at ECML-PKDD. Contribution: TaxoRef reverses the perspective: embeddings are not only evaluated against the taxonomy, they become the instrument for refining it, flagging hierarchical inconsistencies and proposing corrections. [read the paper · full reference · TaxoRef]
• MEET: A Method for Embeddings Evaluation for Taxonomic Data (ICDM Workshops 2020, 2020). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani. In plain terms: MEET introduces the method for evaluating embeddings on taxonomic data, later specialised for the labour market (MEET-LM): it measures how faithfully the vector space respects concept hierarchies. Contribution: MEET introduces the framework for evaluating embeddings on taxonomic data, subsequently adopted throughout the entire follow-up line of research. [read the paper · full reference]
• A Tool for Researchers: Querying Big Scholarly Data Through Graph Databases (ECML-PKDD 2019, 2019). Fabio Mercorio, Mario Mezzanzanica, Vincenzo Moscato, Antonio Picariello, Giancarlo Sperlì. In plain terms: A demo of the tool that lets researchers query big scholarly data as a graph: collaboration paths, communities, and the evolution of research topics. Contribution: A demonstration tool that makes big scholarly data queryable as a graph by researchers, with no database expertise required. [read the paper · full reference · GraphDBLP]
• GraphDBLP: a system for analysing networks of computer scientists through graph databases - GraphDBLP (Multimedia Tools & Applications, 2018). Mario Mezzanzanica, Fabio Mercorio, Mirko Cesarini, Vincenzo Moscato, Antonio Picariello. In plain terms: GraphDBLP turns DBLP (the computer science bibliography) into a graph database: 700,000+ authors and millions of relationships, queryable for collaboration paths, skills, and venues. Released as an open-source tool. Contribution: GraphDBLP is the first complete transformation of DBLP into an open-source graph database: 700,000+ authors queryable by collaboration paths, skills, and venues. [read the paper · full reference · GraphDBLP]
• Planning meets Data Cleansing (ICAPS 2014, 2014). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: 'Planning meets Data Cleansing', presented at ICAPS: data cleansing framed as an automated planning problem, with the planner synthesising the correction sequences. The conceptual bridge between the first and second phases of the group's research. Contribution: 'Planning meets Data Cleansing' at ICAPS: the first formulation of data cleansing as an automated planning problem, with automatic synthesis of correction sequences. The hinge paper between the formal and the data-driven phases of the research path. [read the paper · full reference]

eXplainable AI (2019–present · 18 papers)

Contrastive, global and conversational explanations for machine-learning models.

Where the field stood. Around 2020, XAI mostly offered local explanations (why this single prediction?) in the LIME/SHAP style; what was missing were global, contrastive explanations ('why A and not B?', 'what changed since last time?') and interfaces accessible to non-experts.

Our contribution. The contrastive-explanations line: ContrXT (Information Fusion 2022) explains how a text classifier has changed over time; MERLIN (DSS 2024) compares two models via symbolic reasoning; ConvXAI makes any explainer queryable by voice and text. Plus reference surveys on natural-language explanations (IPM 2023, with E. Cambria) and critical evaluations of post-hoc methods.

Why it matters. Tools for audit and trust: a bank can understand how its anti-fraud model changed after retraining; an operator can ask a system 'why?' without reading gradients. Real-world applications in banking (Intesa Sanpaolo), medical prosthetics, clinical documents.

Open-source tools: ContrXT, MERLIN

• Addressing the Concept Faithfulness Gap through Probe-Guided Sparse Autoencoders (EMNLP 2026 (Findings), 2026). Lorenzo Malandri, Fabio Mercorio, Antonio Serino. In plain terms: Sparse Autoencoder features often look interpretable yet fail to predict concepts — like toxicity or sentiment — that are still clearly encoded in the model's dense activations: a 'concept faithfulness gap'. ProbeSAE closes it by guiding SAE fine-tuning with linear probes, yielding sparse features that match dense-probe accuracy while staying monosemantic and interpretable — and enabling causal control that transfers to six languages from English-only training. Contribution: Introduces ProbeSAE, a probing-guided fine-tuning framework that preserves task-relevant concept structure in the sparse decomposition: more predictive and monosemantic features on TopK and JumpReLU SAEs, language-independent feature directions confirmed by cross-lingual causal ablation, and no overfitting under SAEBench. [full reference]
• SFAL: Semantic-Functional Alignment Scores for Distributional Evaluation of Auto-Interpretability in Sparse Autoencoders (EMNLP 2025, 2025). Fabio Mercorio, Filippo Pallucchini, Daniele Potertì, Antonio Serino, Andrea Seveso. In plain terms: Sparse Autoencoders decompose an LLM's 'thoughts' into interpretable features, but assessing how good the interpretations are is extremely expensive. SFAL proposes a clever shortcut: comparing each feature's semantic neighbourhood with its functional neighbourhood, yielding fast and inexpensive evaluations on large models. Contribution: SFAL replaces the costly LLM-based scoring of auto-interpretability with a distributional comparison between the semantic and functional neighbourhoods of SAE features: scalable, low-cost interpretability evaluation designed for real-world deployment. [read the paper · full reference · SFAL]
• An approach to Evaluative AI through Large Language Models (MAI-XAI@ECAI 2024, 2024). Andrea Ermellino, Lorenzo Malandri, Fabio Mercorio, Navid Nobani, Antonio Serino. In plain terms: Explores 'Evaluative AI': using LLMs not to decide in place of people but to present evidence for and against each hypothesis, supporting a more informed human judgement. Contribution: Operationalises the 'Evaluative AI' paradigm with LLMs: evidence for and against hypotheses instead of flat recommendations, moving XAI towards the support of human judgement. [read the paper · full reference]
• Augmenting XAI with LLMs: A Case Study in Banking Marketing Recommendation (World Conf. xAI 2024, 2024). Alessandro Castelnovo, Roberto Depalmas, Fabio Mercorio, Nicol`o Mombelli, Daniele Potertì, Antonio Serino, Andrea Seveso, Salvatore Sorrentino, Laura Viola. In plain terms: A case study in banking: the explanations produced by a marketing recommendation system are translated by LLMs into language that operators can understand, combining XAI and language generation. Contribution: A real-world banking case: the first study to use LLMs to translate technical XAI explanations into operational language for marketing, evaluating their usefulness with actual users. [read the paper · full reference]
• Model-contrastive explanations through symbolic reasoning (Decision Support Systems, 2024). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: MERLIN explains the difference between two models (or two versions of the same model) through symbolic reasoning: it produces human-readable rules stating 'model A decides this way, model B that way'. Useful for auditing and updating systems in production. Contribution: MERLIN introduces 'model-contrastive' explanations via symbolic reasoning: readable rules describing the behavioural difference between any two models — a problem (change auditing) that local XAI does not address. [read the paper · full reference · MERLIN]
• XAI meets LLMs: A Survey of the Relation between Explainable AI (arXiv (preprint), 2024). Erik Cambria, Lorenzo Malandri, Fabio Mercorio, Navid Nobani, Andrea Seveso. In plain terms: A survey on the relationship between XAI and LLMs: how large language models can generate explanations and how, in turn, they themselves need to be explained. [Preprint] Contribution: The first systematic mapping of the bidirectional XAI↔LLM relationship: models as generators of explanations and as objects to be explained. [read the paper · full reference]
• A survey on XAI and natural language explanations (Inf. Processing & Management, 2023). Erik Cambria, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani. In plain terms: One of the reference surveys on natural language explanations in XAI: why textual explanations are more accessible than charts and numbers, and which techniques exist to generate them. Contribution: A reference survey (with Erik Cambria) on natural language explanations: it defines the sub-field of natural language explanations within XAI. Among the group's most cited works. [read the paper · full reference]
• ConvXAI: a System for Multimodal Interaction with Any Black-box Explainer (Cognitive Computation, 2023). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani. In plain terms: ConvXAI lets users 'converse' with any black-box model explainer: they ask questions in natural language and the system answers by combining text and charts. A step towards XAI for non-experts. Contribution: ConvXAI is the first system that makes any black-box explainer conversational, combining natural language dialogue with multimodal answers (text + charts). [read the paper · full reference]
• Leveraging Group Contrastive Explanations for Handling Fairness (World Conf. xAI 2023, 2023). Alessandro Castelnovo, Nicole Inverardi, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: Uses group contrastive explanations to uncover unfair treatment: by comparing how a model treats different groups of people, hidden discrimination comes to light. Contribution: The first use of group contrastive explanations as a fairness diagnostic tool: discrimination emerges by comparing explanations across groups. [read the paper · full reference]
• A Survey on XAI for Cyber Physical Systems in Medicine (IEEE MetroXRAINE 2022, 2022). Nicola Alimonda, Luca Guidotto, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Giovanni Tosi. In plain terms: A survey on XAI applied to cyber-physical systems in medicine: how to explain the decisions of devices that bring together sensors, software and the human body. Contribution: The first survey devoted to XAI for cyber-physical systems in medicine, where explainability meets sensors and actuators on the human body. [read the paper]
• ContrXT: Generating contrastive explanations from any text classifier (Information Fusion, 2022). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani, Andrea Seveso. In plain terms: ContrXT answers a new question: 'how has my text classifier changed compared to before?'. It compares two versions of a model and generates global contrastive explanations ('it used to decide this way, now it decides that way'). Among the group's most cited XAI works. Contribution: ContrXT defines a new problem — global 'time-contrastive' explanation: not why the model decides as it does, but how and why it has changed with respect to a previous version. The first model-agnostic method for auditing change in text classifiers. [read the paper · full reference · ContrXT]
• FFTree: A flexible tree to handle multiple fairness criteria (Inf. Processing & Management, 2022). Alessandro Castelnovo, Andrea Cosentini, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: FFTree is a 'flexible' decision tree that lets you choose and combine multiple fairness criteria at once: transparent by construction and adaptable to the regulatory context. Born from the collaboration with Intesa Sanpaolo. Contribution: FFTree is the first 'fairness-flexible' decision tree: it allows multiple fairness criteria to be chosen and combined at training time, while remaining interpretable by design. Born from a real need in the banking sector. [read the paper · full reference]
• The Good, the Bad, and the Explainer: A Tool for Contrastive Explanations of Text Classifiers (IJCAI 2022, 2022). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani, Andrea Seveso. In plain terms: Demo of ContrXT presented at IJCAI: the tool that explains how and why a text classifier has changed over time. Contribution: IJCAI demo of the ContrXT tool: the time-contrastive explanation system made usable by the community. [read the paper · full reference · ContrXT]
• XAI for myo-controlled prosthesis: Explaining EMG data for hand gesture classification (Knowledge-Based Systems, 2022). Noemi Gozzi, Lorenzo Malandri, Fabio Mercorio, Alessandra Pedrocchi. In plain terms: Applies XAI to hand prostheses controlled by muscle signals (EMG): explaining which signal patterns drive gesture classification helps build more reliable and personalised prostheses. Contribution: The first systematic application of XAI to myo-controlled prostheses: it explains which EMG patterns drive gesture classification, with implications for reliability and clinical personalisation. [read the paper · full reference]
• Contrastive Explanations of Text Classifiers as a Service (2022 Conference of the North American Chapter of the Asso…, 2022). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: The ContrXT engine — which explains how a text classifier has changed between two versions — packaged as a ready-to-use web service: upload two models, get back a time-contrastive explanation in natural language, no XAI expertise required. Contribution: Demonstrates contrastive model-comparison as a service at NAACL, making the Information Fusion 2022 method accessible to practitioners through a simple API and interface. [read the paper · full reference]
• Towards an Explainer-agnostic Conversational XAI (IJCAI 2021, 2021). Navid Nobani, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Towards a conversational XAI that is independent of the explainer: an architecture for interacting with any explanation system, asking follow-up questions in natural language. Contribution: An architecture for explainer-agnostic conversational XAI: the follow-up dialogue works on top of any existing explanation system. [read the paper]
• A Symbolic Approach to Generating Contrastive Explanations for Black Box Classifiers (AAAI-21, 2021). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: The first public presentation of the idea behind ContrXT: encode the decision logic of two versions of a black-box text classifier into symbolic structures (Binary Decision Diagrams) and reason over them to say, in natural language, what changed and why. Contribution: Introduces symbolic, time-contrastive explanations for black-box classifiers at AAAI — the seed of the ContrXT line later consolidated in Information Fusion and IJCAI. [full reference]
• Towards Labour Market Intelligence through Topic Modelling (HICSS 2019, 2019). Francesco Colace, Massimo De Santo, Marco Lombardi, Fabio Mercorio, Mario Mezzanzanica, Francesco Pascale. In plain terms: Applies topic modelling to job advertisements to surface the latent themes of labour demand, with a focus on the interpretability of the results. Contribution: Applies topic modelling to labour demand while keeping the topics interpretable: an early bridge between LMI and explainability. [read the paper · full reference]

Fairness & Bias (2021–2023 · 4 papers)

Operational tools for fair algorithmic decisions, beyond single-metric fairness.

Where the field stood. Fairness metrics were many and mutually conflicting; practical tools imposed a single criterion, ignored intersectional bias (e.g. woman AND foreign-born) and treated fairness as a static property, to be checked once.

Our contribution. FFTree (IPM 2022) is a decision tree that lets you choose and combine multiple fairness criteria, transparent by construction; 'Towards Fairness Through Time' introduces fairness monitoring across the model life-cycle; other works use group contrastive explanations to uncover discrimination and propose fair rankings under intersectional bias.

Why it matters. Born from the collaboration with the banking sector (industrial PhD with Intesa Sanpaolo), this line provides operational tools for those who must guarantee fair algorithmic decisions in credit, hiring and rankings.

• Leveraging Group Contrastive Explanations for Handling Fairness (World Conf. xAI 2023, 2023). Alessandro Castelnovo, Nicole Inverardi, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: Uses group contrastive explanations to uncover unfair treatment: by comparing how a model treats different groups of people, hidden discrimination comes to light. Contribution: The first use of group contrastive explanations as a fairness diagnostic tool: discrimination emerges by comparing explanations across groups. [read the paper · full reference]
• Preserving Utility in Fair Top-k Ranking with Intersectional Bias (BIAS@ECIR 2023, 2023). Nicola Alimonda, Alessandro Castelnovo, Riccardo Crupi, Fabio Mercorio, Mario Mezzanzanica. In plain terms: How to order a ranking (e.g. candidates for a position) fairly towards groups affected by intersectional bias, while losing as little as possible in terms of the ranking's own utility. Contribution: A top-k re-ranking algorithm that handles intersectional bias (combinations of protected attributes) while minimising utility loss: previous methods dealt with one attribute at a time. [read the paper · full reference]
• FFTree: A flexible tree to handle multiple fairness criteria (Inf. Processing & Management, 2022). Alessandro Castelnovo, Andrea Cosentini, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica. In plain terms: FFTree is a 'flexible' decision tree that lets you choose and combine multiple fairness criteria at once: transparent by construction and adaptable to the regulatory context. Born from the collaboration with Intesa Sanpaolo. Contribution: FFTree is the first 'fairness-flexible' decision tree: it allows multiple fairness criteria to be chosen and combined at training time, while remaining interpretable by design. Born from a real need in the banking sector. [read the paper · full reference]
• Towards Fairness Through Time (ECML-PKDD 2021, 2021). Alessandro Castelnovo, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Andrea Cosentini. In plain terms: Fairness is not static: a model that is fair today can discriminate tomorrow as the data change. The paper introduces the monitoring of fairness over time in production ML systems. Contribution: Introduces the temporal dimension into fairness: monitoring equity throughout the model's life cycle, because a system that is fair at deployment can discriminate after data drift. [read the paper · full reference]

LLM Evaluation & Benchmarks (2023–present · 23 papers)

Measuring what LLMs really understand about Italian language, culture and jobs.

Where the field stood. After ChatGPT, LLM evaluation was dominated by English benchmarks, often poorly translated: for Italian there were no serious measures of linguistic and, above all, cultural understanding, and impact assessments (e.g. on jobs) were speculative.

Our contribution. ITALIC (NAACL 2025): 10,000 questions from real public-service exams on Italian language and culture; the INVALSI suite (ECML-PKDD 2025) on school competences; BEEP on driving-licence quizzes; the contribution to the national CALAMITA initiative. On the applied side: Terminator Economy (IJCAI 2025) uses LLMs to estimate the exposure of jobs to AI; TAXMAP makes them collaborate to expand ESCO; RE-FIN uses them to enrich financial data without hallucinations.

Why it matters. It tells us how much LLMs really understand Italian — and Italy (with results covered by La Repubblica) — and turns them from hype into measurable tools for concrete tasks: taxonomies, finance, labour policy.

Open-source tools: ITALIC, INVALSI-Eval-Suite

• Diagnosing LLM Fragility: An Empirical Study on Reasoning Entropy and Malicious Outputs (EMNLP 2026, 2026). Fabio Mercorio, Daniele Potertì, Andrea Seveso. In plain terms: Chain-of-Thought traces are increasingly used as a safety window into a model's reasoning — but reasoning is a stochastic branching process, not an auditable path. The paper introduces Bifurcation Entropy (H_BE), a step-level metric that measures how widely a model's reasoning can branch at each step, and shows it pinpoints the fragile steps: across five reasoning models in high-stakes scenarios, top-quartile steps are exploited 6.2× more often than bottom-quartile ones under adversarial injection. Contribution: Extends Kernel Language Entropy from final answers to intermediate reasoning steps (H_BE) and validates it with In-Distribution Attacks, a probing protocol that injects the model's own high-entropy continuations — establishing H_BE as a principled signal to concentrate safety monitoring on the few reasoning steps that carry most of the alignment risk. [full reference]
• Addressing the Concept Faithfulness Gap through Probe-Guided Sparse Autoencoders (EMNLP 2026 (Findings), 2026). Lorenzo Malandri, Fabio Mercorio, Antonio Serino. In plain terms: Sparse Autoencoder features often look interpretable yet fail to predict concepts — like toxicity or sentiment — that are still clearly encoded in the model's dense activations: a 'concept faithfulness gap'. ProbeSAE closes it by guiding SAE fine-tuning with linear probes, yielding sparse features that match dense-probe accuracy while staying monosemantic and interpretable — and enabling causal control that transfers to six languages from English-only training. Contribution: Introduces ProbeSAE, a probing-guided fine-tuning framework that preserves task-relevant concept structure in the sparse decomposition: more predictive and monosemantic features on TopK and JumpReLU SAEs, language-independent feature directions confirmed by cross-lingual causal ablation, and no overfitting under SAEBench. [full reference]
• Safe-Unsafe Concept Separation Emerges from a Single Direction in Language Models Activation Space (EACL 2026, 2026). Andrea Ermellino, Lorenzo Malandri, Fabio Mercorio, Antonio Serino. In plain terms: Inside a language model, 'safe' and 'unsafe' concepts separate along a single geometric direction. The paper shows how to find the exact point where this separation is greatest and use it as an internal 'alarm': a guardrail more robust than traditional filters, without touching the model. Validated across 16 languages and 3 tasks (safety, prompt injection, toxicity). Contribution: Shows that safe/unsafe separation emerges along a single linear direction in the activation space, and that a linear classifier on the optimal layer is enough for a guardrail more robust than generative filters — with no fine-tuning and no changes to the weights. [read the paper · full reference]
• SkiLLens: Recognising and Mapping Novel Skills from Millions of Job Ads Across Europe Using Language Models (EACL 2026, 2026). Alessia De Santo, Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani. In plain terms: A system that reads millions of European job ads and discovers the new skills the market is asking for (e.g. new tools, new practices), automatically linking them to the European ESCO taxonomy with the help of LLMs and human review. Designed for those who shape training and skills policies. Contribution: The first multilingual human-in-the-loop pipeline to close the entire loop: extraction of candidate skills from national job-ad corpora, entity linking to ESCO via encoders and refinement with decoder LLMs, validated both quantitatively and qualitatively. [read the paper · full reference]
• Terminator Economy: Assessing jobs and tasks exposure to AI (ECML-PKDD 2026, 2026). Fabio Mercorio, Mario Mezzanzanica, Daniele Potertì, Antonio Serino, Andrea Seveso. In plain terms: How exposed is each job — and each single task — to today's AI? The paper presents the Terminator Economy framework at ECML-PKDD: open-source LLMs assess AI's current ability to perform job tasks, producing the TEAI (Task Exposure to AI) and TRAI (Task Replacement by AI) indices, validated with human evaluators. About one third of U.S. employment turns out to be highly exposed — mostly high-skill jobs — yet exposure correlates with employment and wage growth, suggesting AI complements more than it replaces. Contribution: Brings the TEAI/TRAI framework to the European ML community as a reproducible, open benchmark: a task-level, LLM-based measure of AI exposure that can be recomputed over time to monitor AI's progress on the labour market. [full reference · Terminator-Economy]
• A Benchmark to Evaluate LLMs' Proficiency on Italian Student Competencies (ECML-PKDD 2025, 2025). Fabio Mercorio, Mario Mezzanzanica, Daniele Potertì, Antonio Serino, Andrea Seveso. In plain terms: How well do LLMs perform on the INVALSI tests taken by Italian students? The paper builds a benchmark of 405 questions across 6 school levels and measures where the models excel and where they stumble in Italian. Contribution: The first benchmark that uses official INVALSI tests to measure LLMs on the competencies certified by the Italian school system, across 6 levels of education. [read the paper · full reference · INVALSI-Eval-Suite]
• Can Role Vectors Affect LLM Behaviour? (EMNLP 2025, 2025). Daniele Potertì, Andrea Seveso, Fabio Mercorio. In plain terms: Instead of telling an LLM 'act like a chemist' in the prompt, one can inject a 'role vector' directly into its internal activations. The paper builds 29 role vectors and shows that intervening inside the model works better than persona prompting. Contribution: Introduces 'role vectors': 29 role directions extracted from the activations, with the first comparative evidence that intervening on internal representations (activation addition / directional ablation) outperforms persona prompting. [read the paper · full reference]
• Challenging the Abilities of Large Language Models in Italian: a Community Initiative (arXiv (preprint), 2025). Malvina Nissim, Danilo Croce, Viviana Patti, Pierpaolo Basile, Giuseppe Attanasio, Elio Musacchio, Matteo Rinaldi, Federico Borazio, Maria Francis, Jacopo Gili, Daniel Scalena, Bego~na Altuna, Ekhi Azurmendi, Valerio Basile, Luisa Bentivogli, Arianna Bisazza, Marianna Bolognesi, Dominique Brunato, Tommaso Caselli, Silvia Casola, Maria Cassese, Mauro Cettolo, Claudia Collacciani, Leonardo De Cosmo, Maria Pia di Buono, Andrea Esuli, Julen Etxaniz, Chiara Ferrando, Alessia Fidelangeli, Simona Frenda, Achille Fusco, Marco Gaido, Andrea Galassi, Federico Galli, Luca Giordano, Mattia Goffetti, Itziar Gonzalez-Dios, Lorenzo Gregori, Giulia Grundler, Sandro Iannaccone, Chunyang Jiang, Moreno La Quatra, Francesca Lagioia, Soda Marem Lo, Marco Madeddu, Bernardo Magnini, Raffaele Manna, Fabio Mercorio, Paola Merlo, Arianna Muti, Vivi Nastase, Matteo Negri, Dario Onorati, Elena Palmieri, Sara Papi, Lucia C. Passaro, Giulia Pensa, Andrea Piergentili, Daniele Potertì, Giovanni Puccetti, Federico Ranaldi, Leonardo Ranaldi, Andrea Amelio Ravelli, Martina Rosola, Elena Sofia Ruzzetti, Giuseppe Samo, Andrea Santilli, Piera Santin, Gabriele Sarti, Giovanni Sartor, Beatrice Savoldi, Antonio Serino, Andrea Seveso, Lucia Siciliani, Paolo Torroni, Rossella Varvara, Andrea Zaninello, Asya Zanollo, Fabio Massimo Zanzotto, Kamyar Zeinalipour, Andrea Zugarini. In plain terms: CALAMITA is the Italian community initiative (80+ contributors) for evaluating LLMs in Italian: over 20 tasks and nearly 100 subtasks, with a centralised evaluation pipeline. The paper reports the methodology, results, and lessons learned. [Preprint] Contribution: Contribution to the national CALAMITA initiative: the largest federation of evaluation tasks for Italian (80+ contributors, 20+ tasks), with a centralised pipeline and a 'rolling' benchmark. [read the paper]
• Designing Role Vectors to Improve LLM Inference Behaviour (arXiv (preprint), 2025). Daniele Potertì, Andrea Seveso, Fabio Mercorio. In plain terms: Preprint version of the study on 'role vectors': how to steer an LLM's behaviour by modifying its internal representations rather than the prompt. [Preprint] Contribution: Preprint version of the role vectors contribution: steering behaviour via internal representations rather than prompts. [read the paper]
• ITALIC: An Italian Culture-Aware Natural Language Benchmark (NAACL 2025, 2025). Andrea Seveso, Daniele Potertì, Edoardo Federici, Mario Mezzanzanica, Fabio Mercorio. In plain terms: ITALIC is a benchmark of 10,000 multiple-choice questions on Italian language and culture, built from real public examinations. It measures how well LLMs truly understand Italian and Italy: the results reveal significant cultural gaps even in the best models. Contribution: The first large-scale benchmark (10,000 questions from real public examinations) measuring LLMs' cultural — not just linguistic — understanding of Italian. [read the paper · full reference · ITALIC]
• RE-FIN: Retrieval-based Enrichment for Financial data (COLING 2025, 2025). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Filippo Pallucchini. In plain terms: RE-FIN automatically enriches financial sentences (often short and full of implicit assumptions) with knowledge retrieved from a document base, rewriting them in more explicit terms. It improves financial sentiment analysis while reducing the risk of hallucinations. Contribution: RE-FIN constrains the LLM to rewrite using only propositions retrieved from the knowledge base: text enrichment with a structurally reduced risk of hallucination, plus a selection algorithm that preserves meaning. [read the paper · full reference]
• SAFE: A Sparse Autoencoder-Based Framework for Robust Query Enrichment and Hallucination Mitigation in LLMs (EMNLP 2025, 2025). Samir Abdaljalil, Filippo Pallucchini, Andrea Seveso, Hasan Kurban, Fabio Mercorio, Erchin Serpedin. In plain terms: SAFE uses Sparse Autoencoders — tools that 'open up' an LLM's internal representations — to recognise when the model is about to hallucinate and enrich the query accordingly. It improves accuracy by up to 29%. Contribution: The first systematic integration of Sparse Autoencoders and hallucination detection into a single 'hallucination-aware' query enrichment framework, with gains of up to 29.45%. [read the paper · full reference]
• SAFE: A Sparse Autoencoder-Based Framework for Robust Query Enrichment and Hallucination Mitigation in LLMs (arXiv (preprint), 2025). Samir Abdaljalil, Filippo Pallucchini, Andrea Seveso, Hasan Kurban, Fabio Mercorio, Erchin Serpedin. In plain terms: Preprint version of SAFE: detecting and mitigating LLM hallucinations using Sparse Autoencoders. [Preprint] Contribution: Preprint version of SAFE: hallucination detection and mitigation via Sparse Autoencoders. [read the paper · full reference]
• SFAL: Semantic-Functional Alignment Scores for Distributional Evaluation of Auto-Interpretability in Sparse Autoencoders (EMNLP 2025, 2025). Fabio Mercorio, Filippo Pallucchini, Daniele Potertì, Antonio Serino, Andrea Seveso. In plain terms: Sparse Autoencoders decompose an LLM's 'thoughts' into interpretable features, but assessing how good the interpretations are is extremely expensive. SFAL proposes a clever shortcut: comparing each feature's semantic neighbourhood with its functional neighbourhood, yielding fast and inexpensive evaluations on large models. Contribution: SFAL replaces the costly LLM-based scoring of auto-interpretability with a distributional comparison between the semantic and functional neighbourhoods of SAE features: scalable, low-cost interpretability evaluation designed for real-world deployment. [read the paper · full reference · SFAL]
• Taxonomy Expansion through Collaborative LLM Mapping (ACM SAC 2025, 2025). Simone D'Amico, Alessia De Santo, Mario Mezzanzanica, Fabio Mercorio. In plain terms: TAXMAP has three LLMs collaborate to propose new terms to add to a taxonomy (here, ESCO), with final human validation. On 40,000 digital terms collected from the web, it proposed 924 additions, 81% of which were judged correct by experts. Contribution: TAXMAP has three generative LLMs collaborate as a 'committee' for taxonomy expansion with final human validation: 81% PPV on 40,000 real-world terms for ESCO, drastically reducing the expert workload. [read the paper · full reference]
• Towards the Terminator Economy: Assessing Job Exposure to AI Through LLMs (IJCAI 2025, 2025). Emilio Colombo, Fabio Mercorio, Mario Mezzanzanica, Antonio Serino. In plain terms: How exposed is each job to AI? The paper builds a pipeline in which LLMs themselves assess, task by task, what they can already do of real occupations. The result is TerminatorEconomy.com, cited by Repubblica, Wired, and Radio24. Contribution: A two-stage methodology that uses LLMs themselves to assess, task by task, the exposure of occupations to AI: an empirical, reproducible measure where before there were only expert estimates, released as an explorable tool (TerminatorEconomy.com). [read the paper · full reference · Terminator-Economy]
• An approach to Evaluative AI through Large Language Models (MAI-XAI@ECAI 2024, 2024). Andrea Ermellino, Lorenzo Malandri, Fabio Mercorio, Navid Nobani, Antonio Serino. In plain terms: Explores 'Evaluative AI': using LLMs not to decide in place of people but to present evidence for and against each hypothesis, supporting a more informed human judgement. Contribution: Operationalises the 'Evaluative AI' paradigm with LLMs: evidence for and against hypotheses instead of flat recommendations, moving XAI towards the support of human judgement. [read the paper · full reference]
• Augmenting XAI with LLMs: A Case Study in Banking Marketing Recommendation (World Conf. xAI 2024, 2024). Alessandro Castelnovo, Roberto Depalmas, Fabio Mercorio, Nicol`o Mombelli, Daniele Potertì, Antonio Serino, Andrea Seveso, Salvatore Sorrentino, Laura Viola. In plain terms: A case study in banking: the explanations produced by a marketing recommendation system are translated by LLMs into language that operators can understand, combining XAI and language generation. Contribution: A real-world banking case: the first study to use LLMs to translate technical XAI explanations into operational language for marketing, evaluating their usefulness with actual users. [read the paper · full reference]
• BEEP - BEst DrivEr's License Performer: A CALAMITA Challenge (CLiC-it 2024, 2024). Fabio Mercorio, Daniele Potertì, Antonio Serino, Andrea Seveso. In plain terms: BEEP puts LLMs to the test on the Italian driving licence quizzes: a CALAMITA challenge that measures their grasp of rules, exceptions, and linguistic traps. Contribution: Turns the official Italian driving licence quizzes into a CALAMITA challenge: a benchmark where rules, exceptions, and linguistic traps trip up even the best models. [read the paper · full reference]
• Disce aut Deficere: Evaluating LLMs Proficiency on the INVALSI Italian Benchmark (arXiv (preprint), 2024). Fabio Mercorio, Mario Mezzanzanica, Daniele Potertì, Antonio Serino, Andrea Seveso. In plain terms: Preprint version of the INVALSI benchmark for LLMs: measuring model proficiency on Italian school assessments. [Preprint] Contribution: Preprint version of the INVALSI benchmark for LLMs. [read the paper · full reference · INVALSI-Eval-Suite]
• Towards the Terminator Economy: Assessing Job Exposure to AI through LLMs (arXiv (preprint), 2024). Emilio Colombo, Fabio Mercorio, Mario Mezzanzanica, Antonio Serino. In plain terms: Preprint version of the 'Terminator Economy' study on job exposure to AI. [Preprint] Contribution: Preprint version of the Terminator Economy study on job exposure to AI. [read the paper · full reference · Terminator-Economy]
• XAI meets LLMs: A Survey of the Relation between Explainable AI (arXiv (preprint), 2024). Erik Cambria, Lorenzo Malandri, Fabio Mercorio, Navid Nobani, Andrea Seveso. In plain terms: A survey on the relationship between XAI and LLMs: how large language models can generate explanations and how, in turn, they themselves need to be explained. [Preprint] Contribution: The first systematic mapping of the bidirectional XAI↔LLM relationship: models as generators of explanations and as objects to be explained. [read the paper · full reference]
• Marrying LLMs with Domain Expert Validation for Causal Graph Generation (short paper) (AIABI@AI*IA 2023, 2023). Alessandro Castelnovo, Riccardo Crupi, Fabio Mercorio, Mario Mezzanzanica, Daniele Potertì, Daniele Regoli. In plain terms: Makes LLMs and domain experts work together to build causal graphs: the model proposes cause-effect relations, the expert validates them. A practical workflow for causal discovery. Contribution: An LLM + expert-validation workflow for causal graph generation: the model proposes, the expert verifies — a practical trade-off between automation and causal reliability.

Mechanistic Interpretability & AI Safety (2025–present · 8 papers)

Opening the LLM black box: SAE features, steering vectors, linear guardrails.

Where the field stood. Mechanistic interpretability (opening the LLM 'black box' by looking at activations, directions and features) was confined to a few industrial labs; evaluating Sparse Autoencoder feature interpretations cost prohibitive amounts in LLM calls, and safety guardrails remained external filters — opaque and easy to bypass.

Our contribution. Four contributions in rapid succession: SFAL (EMNLP 2025) makes SAE auto-interpretability evaluation affordable by comparing semantic and functional neighbourhoods of features; SAFE (EMNLP 2025) uses SAEs to detect and mitigate hallucinations (+29% accuracy); Role Vectors (EMNLP 2025) steer model behaviour by acting on activations rather than prompts; the EACL 2026 work shows that safe/unsafe concepts separate along a single geometric direction, yielding a linear guardrail validated across 16 languages.

Why it matters. It is the natural evolution of the XAI line in the LLM era: understanding models from the inside to make them safer, more controllable and more reliable — with methods efficient enough to run in production, not just in the lab.

Open-source tools: SFAL

• Diagnosing LLM Fragility: An Empirical Study on Reasoning Entropy and Malicious Outputs (EMNLP 2026, 2026). Fabio Mercorio, Daniele Potertì, Andrea Seveso. In plain terms: Chain-of-Thought traces are increasingly used as a safety window into a model's reasoning — but reasoning is a stochastic branching process, not an auditable path. The paper introduces Bifurcation Entropy (H_BE), a step-level metric that measures how widely a model's reasoning can branch at each step, and shows it pinpoints the fragile steps: across five reasoning models in high-stakes scenarios, top-quartile steps are exploited 6.2× more often than bottom-quartile ones under adversarial injection. Contribution: Extends Kernel Language Entropy from final answers to intermediate reasoning steps (H_BE) and validates it with In-Distribution Attacks, a probing protocol that injects the model's own high-entropy continuations — establishing H_BE as a principled signal to concentrate safety monitoring on the few reasoning steps that carry most of the alignment risk. [full reference]
• Addressing the Concept Faithfulness Gap through Probe-Guided Sparse Autoencoders (EMNLP 2026 (Findings), 2026). Lorenzo Malandri, Fabio Mercorio, Antonio Serino. In plain terms: Sparse Autoencoder features often look interpretable yet fail to predict concepts — like toxicity or sentiment — that are still clearly encoded in the model's dense activations: a 'concept faithfulness gap'. ProbeSAE closes it by guiding SAE fine-tuning with linear probes, yielding sparse features that match dense-probe accuracy while staying monosemantic and interpretable — and enabling causal control that transfers to six languages from English-only training. Contribution: Introduces ProbeSAE, a probing-guided fine-tuning framework that preserves task-relevant concept structure in the sparse decomposition: more predictive and monosemantic features on TopK and JumpReLU SAEs, language-independent feature directions confirmed by cross-lingual causal ablation, and no overfitting under SAEBench. [full reference]
• Safe-Unsafe Concept Separation Emerges from a Single Direction in Language Models Activation Space (EACL 2026, 2026). Andrea Ermellino, Lorenzo Malandri, Fabio Mercorio, Antonio Serino. In plain terms: Inside a language model, 'safe' and 'unsafe' concepts separate along a single geometric direction. The paper shows how to find the exact point where this separation is greatest and use it as an internal 'alarm': a guardrail more robust than traditional filters, without touching the model. Validated across 16 languages and 3 tasks (safety, prompt injection, toxicity). Contribution: Shows that safe/unsafe separation emerges along a single linear direction in the activation space, and that a linear classifier on the optimal layer is enough for a guardrail more robust than generative filters — with no fine-tuning and no changes to the weights. [read the paper · full reference]
• Can Role Vectors Affect LLM Behaviour? (EMNLP 2025, 2025). Daniele Potertì, Andrea Seveso, Fabio Mercorio. In plain terms: Instead of telling an LLM 'act like a chemist' in the prompt, one can inject a 'role vector' directly into its internal activations. The paper builds 29 role vectors and shows that intervening inside the model works better than persona prompting. Contribution: Introduces 'role vectors': 29 role directions extracted from the activations, with the first comparative evidence that intervening on internal representations (activation addition / directional ablation) outperforms persona prompting. [read the paper · full reference]
• Designing Role Vectors to Improve LLM Inference Behaviour (arXiv (preprint), 2025). Daniele Potertì, Andrea Seveso, Fabio Mercorio. In plain terms: Preprint version of the study on 'role vectors': how to steer an LLM's behaviour by modifying its internal representations rather than the prompt. [Preprint] Contribution: Preprint version of the role vectors contribution: steering behaviour via internal representations rather than prompts. [read the paper]
• SAFE: A Sparse Autoencoder-Based Framework for Robust Query Enrichment and Hallucination Mitigation in LLMs (EMNLP 2025, 2025). Samir Abdaljalil, Filippo Pallucchini, Andrea Seveso, Hasan Kurban, Fabio Mercorio, Erchin Serpedin. In plain terms: SAFE uses Sparse Autoencoders — tools that 'open up' an LLM's internal representations — to recognise when the model is about to hallucinate and enrich the query accordingly. It improves accuracy by up to 29%. Contribution: The first systematic integration of Sparse Autoencoders and hallucination detection into a single 'hallucination-aware' query enrichment framework, with gains of up to 29.45%. [read the paper · full reference]
• SAFE: A Sparse Autoencoder-Based Framework for Robust Query Enrichment and Hallucination Mitigation in LLMs (arXiv (preprint), 2025). Samir Abdaljalil, Filippo Pallucchini, Andrea Seveso, Hasan Kurban, Fabio Mercorio, Erchin Serpedin. In plain terms: Preprint version of SAFE: detecting and mitigating LLM hallucinations using Sparse Autoencoders. [Preprint] Contribution: Preprint version of SAFE: hallucination detection and mitigation via Sparse Autoencoders. [read the paper · full reference]
• SFAL: Semantic-Functional Alignment Scores for Distributional Evaluation of Auto-Interpretability in Sparse Autoencoders (EMNLP 2025, 2025). Fabio Mercorio, Filippo Pallucchini, Daniele Potertì, Antonio Serino, Andrea Seveso. In plain terms: Sparse Autoencoders decompose an LLM's 'thoughts' into interpretable features, but assessing how good the interpretations are is extremely expensive. SFAL proposes a clever shortcut: comparing each feature's semantic neighbourhood with its functional neighbourhood, yielding fast and inexpensive evaluations on large models. Contribution: SFAL replaces the costly LLM-based scoring of auto-interpretability with a distributional comparison between the semantic and functional neighbourhoods of SAE features: scalable, low-cost interpretability evaluation designed for real-world deployment. [read the paper · full reference · SFAL]

Education & Skills (2024–present · 3 papers)

Italian institutional tests turned into rigorous LLM benchmarks.

Where the field stood. LLM evaluation in educational settings relied on English-language academic benchmarks; measures anchored to real school systems and to competences certified by national bodies were missing.

Our contribution. The INVALSI suite (405 questions, 6 school levels) and BEEP (driving-licence quizzes) turn Italian institutional tests into rigorous LLM benchmarks, within the national CALAMITA framework.

Why it matters. It lets schools, publishers and policy makers know what LLMs can and cannot do with respect to the competences the Italian education system actually certifies.

Open-source tools: INVALSI-Eval-Suite

• A Benchmark to Evaluate LLMs' Proficiency on Italian Student Competencies (ECML-PKDD 2025, 2025). Fabio Mercorio, Mario Mezzanzanica, Daniele Potertì, Antonio Serino, Andrea Seveso. In plain terms: How well do LLMs perform on the INVALSI tests taken by Italian students? The paper builds a benchmark of 405 questions across 6 school levels and measures where the models excel and where they stumble in Italian. Contribution: The first benchmark that uses official INVALSI tests to measure LLMs on the competencies certified by the Italian school system, across 6 levels of education. [read the paper · full reference · INVALSI-Eval-Suite]
• BEEP - BEst DrivEr's License Performer: A CALAMITA Challenge (CLiC-it 2024, 2024). Fabio Mercorio, Daniele Potertì, Antonio Serino, Andrea Seveso. In plain terms: BEEP puts LLMs to the test on the Italian driving licence quizzes: a CALAMITA challenge that measures their grasp of rules, exceptions, and linguistic traps. Contribution: Turns the official Italian driving licence quizzes into a CALAMITA challenge: a benchmark where rules, exceptions, and linguistic traps trip up even the best models. [read the paper · full reference]
• Disce aut Deficere: Evaluating LLMs Proficiency on the INVALSI Italian Benchmark (arXiv (preprint), 2024). Fabio Mercorio, Mario Mezzanzanica, Daniele Potertì, Antonio Serino, Andrea Seveso. In plain terms: Preprint version of the INVALSI benchmark for LLMs: measuring model proficiency on Italian school assessments. [Preprint] Contribution: Preprint version of the INVALSI benchmark for LLMs. [read the paper · full reference · INVALSI-Eval-Suite]

Multimodal AI for Finance (2023–present · 3 papers)

Multimodal, hallucination-aware AI for financial analysis.

Where the field stood. Automatic financial analysis used almost exclusively numerical series or text alone; multimodal models were emerging but without systematisation, and financial texts — short and allusive — broke standard sentiment analysis.

Our contribution. Within the PRIN ISALDI project (of which Mercorio is PI): RE-FIN (COLING 2025) enriches financial sentences with retrieved knowledge, reducing hallucinations; the Information Fusion survey (2026) systematises multimodal models for finance; ConvXAI brings multimodal explainability into the domain.

Why it matters. Tools for more accurate and interpretable financial analysis: from knowledge-enriched sentiment analysis to criteria for choosing and combining different data modalities.

• Learning across modalities: a systematic survey of multimodal models for financial analysis (Information Fusion, 2026). Simone D'Amico, Fabio Mercorio, Navid Nobani, Giancarlo Sperlì, Carmine Ventre. In plain terms: A systematic survey of AI models that combine multiple types of data (text, numbers, charts, audio) to analyse financial markets: it takes stock of what works, what is missing, and where the field is heading. Contribution: The first coherent systematisation of multimodal models for financial analysis: a taxonomy of architectures, data, and tasks that the literature was lacking. [read the paper · full reference]
• RE-FIN: Retrieval-based Enrichment for Financial data (COLING 2025, 2025). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Filippo Pallucchini. In plain terms: RE-FIN automatically enriches financial sentences (often short and full of implicit assumptions) with knowledge retrieved from a document base, rewriting them in more explicit terms. It improves financial sentiment analysis while reducing the risk of hallucinations. Contribution: RE-FIN constrains the LLM to rewrite using only propositions retrieved from the knowledge base: text enrichment with a structurally reduced risk of hallucination, plus a selection algorithm that preserves meaning. [read the paper · full reference]
• ConvXAI: a System for Multimodal Interaction with Any Black-box Explainer (Cognitive Computation, 2023). Lorenzo Malandri, Fabio Mercorio, Mario Mezzanzanica, Navid Nobani. In plain terms: ConvXAI lets users 'converse' with any black-box model explainer: they ask questions in natural language and the system answers by combining text and charts. A step towards XAI for non-experts. Contribution: ConvXAI is the first system that makes any black-box explainer conversational, combining natural language dialogue with multimodal answers (text + charts). [read the paper · full reference]

Other Applications (2014–present · 5 papers)

The same methods, ported: clinical text, mobility, synthetic data, open government data.

Our contribution. Cross-cutting applications of the group's methods to other domains: clinical documents (eXDiL), social media and recommender systems, smart mobility, synthetic data (tertiary study in IPM 2026), open data for public administration.

Why it matters. They demonstrate the portability of the methods: explainable classification, graphs and rigorous evaluation work outside labour and finance too.

• Synthetic data generation: A tertiary study (Inf. Processing & Management, 2026). Navid Nobani, Giovanni Officioso, Filippo Pallucchini, Giancarlo Sperlì, Fabio Mercorio. In plain terms: A 'study of studies' on synthetic data generation: it analyses 17 existing surveys and shows that the field is concentrated on healthcare, evaluates privacy and diversity poorly, and is rarely reproducible. It proposes a common evaluation framework. Contribution: The first tertiary study (a review of reviews) on synthetic data generation, with DARE-4 appraisal and a domain-agnostic evaluation baseline proposed for the field. [read the paper · full reference]
• eXDiL: A Tool for Classifying and eXplaining Hospital Discharge Letters (CD-MAKE 2020, 2020). Fabio Mercorio, Mario Mezzanzanica, Andrea Seveso. In plain terms: eXDiL classifies hospital discharge letters and explains its own decisions: an early example of XAI applied to clinical documents, where transparency is non-negotiable. Contribution: eXDiL combines the classification and explanation of hospital discharge letters: XAI applied to clinical documents when the field was still in its infancy. [read the paper · full reference]
• A tool for exploring networks of computer scientists as a graph (ACM SAC 2019, 2019). Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica, Vincenzo Moscato, Antonio Picariello. In plain terms: A tool for exploring the network of computer science researchers as a navigable graph, built on top of DBLP. Contribution: An exploratory interface over the network of computer scientists, built on GraphDBLP. [read the paper · full reference · GraphDBLP]
• A Social Media Recommender System (Int. J. Multimedia Data Eng., 2018). Giancarlo Sperlì, Flora Amato, Fabio Mercorio, Mario Mezzanzanica, Vincenzo Moscato, Antonio Picariello. In plain terms: A recommender system for social media that combines multimedia content and social relationships to suggest relevant content. Contribution: A social recommender system that fuses multimedia content with the relational structure of the network. [read the paper · full reference]
• Applying the AHP to Smart Mobility Services: A Case Study (DATA 2015, 2015). Roberto Boselli, Mirko Cesarini, Fabio Mercorio, Mario Mezzanzanica. In plain terms: Applies the AHP decision-making method to smart mobility services: a case study on how to choose among intelligent transport alternatives. Contribution: Application of the AHP decision-making method to smart mobility: structured selection among service alternatives. [read the paper · full reference]