LLM safety researcher
University of Modena and Reggio Emilia
When should a model refuse, ask or abstain?
I build evaluations of how language models handle sensitive requests, incomplete information and the evidence we give them. I am open to research roles in LLM safety, evaluation and interpretability.
@giosullutrone
Context-driven over-refusalPinned
Can a document make a harmless task look unsafe?
Can a document make a model refuse a harmless request? We examine how the supplied context changes refusal across question answering, translation, and summarization.
Giovanni Sullutrone, Riccardo A. Vigliermo, Sonia Bergamaschi, Luca Sala
Privacy & database accessPinned
When should a database question be refused?
A healthcare schema may contain personal or sensitive fields even when a query is legitimate. We examine how prompts and user-permission context affect refusal and usable database access.
SEBD 2025; proceedings published March 2026.
Giovanni Sullutrone, Luca Sala, Lisa Trigiante, Sonia Bergamaschi
Retrieval in digital librariesPinned
Which passages stay out of view?
We examine whether embedding models make historically sensitive passages harder to retrieve, using questions grounded in a collection of religious texts.
Giovanni Sullutrone, Riccardo Amerigo Vigliermo, Luca Sala, Sonia Bergamaschi
Interactive Text-to-SQL evaluationPinned
What should an agent ask before it writes SQL?
A database question may be ambiguous or impossible to answer. ABISS evaluates how agents clarify a question, explain why a request cannot be answered, and turn useful clarification into correct SQL.
Giovanni Sullutrone, Luca Sala, Sania Aftar, Georgia Koutrika, Sonia Bergamaschi
Conversational database systems
How do we make conversation with a database work?
As database systems become more conversational, query accuracy is only part of the evaluation. This paper examines user interaction, trustworthy behavior, and the economic sustainability of LLM-based Text-to-SQL.
Luca Sala, Giovanni Sullutrone, Sonia Bergamaschi
Arabic-script OCR & digital libraries
Which pages belong in an OCR dataset?
Proposes an Arabic title-page dataset and evaluates frontispiece classification. The planned workflow combines OCR and linguistic review to prepare text for recognition and cataloguing research.
Journal article developing the IRCDL 2025 extended abstract.
Riccardo Amerigo Vigliermo, Giovanni Sullutrone, Sonia Bergamaschi, Luca Sala
Semi-automatic document cataloguing
How do historical documents become discoverable?
A platform that helps specialists create bibliographic records for non-Latin documents. It combines OCR, language models, and vision-language models to extract metadata from Arabic-script material.
Sonia Bergamaschi, Federico Ruozzi, Domenico Ciccarello, Riccardo Amerigo Vigliermo, Luca Sala, Amina El Ganadi, Giovanni Sullutrone, Sania Aftar
OCR & cataloguing research
What can a page tell a library?
An Arabic-script document resource for OCR, page classification, and cataloguing. Its pipeline identifies frontispieces, indexes, and ISBN pages, then extracts structured text for bibliographic metadata.
Online 2025; proceedings citation 2026. Dataset access is by research request.
Luca Sala, Riccardo Amerigo Vigliermo, Giovanni Sullutrone, Sonia Bergamaschi
Dataset design · Extended abstract
What would a richer archive need?
A proposal for an Arabic frontispiece dataset, combining vision-language page classification, OCR, and expert correction to improve the metadata used to discover historical documents.
Riccardo Amerigo Vigliermo, Giovanni Sullutrone, Sonia Bergamaschi, Luca Sala
Language models for OCR correction
Can GPT repair what OCR gets wrong?
Tests GPT-4 and GPT-4 Turbo as post-processing tools for Arabic OCR. Across prompting and generation settings, the models remain unreliable at correcting Arabic-script recognition errors.
Second place, Best Short Paper Award.
Luca Sala, Giovanni Sullutrone, Sonia Bergamaschi, Riccardo Amerigo Vigliermo
LLMs & real-world databases
How far is a question from a useful query?
Examines practical limits of LLM-based database querying: scalability, latency, hallucinations, representative benchmarks, and knowledge acquisition.
Luca Sala, Giovanni Sullutrone, Sonia Bergamaschi
Doctoral research directions
What can language models open up in an archive?
Outlines my initial PhD research directions for multilingual historical documents, connecting Text-to-SQL, topic modelling, and knowledge graphs with transparent access in Digital Maktaba.
The person behind the papers
Discussions with researchers or users—and things I strongly agree or disagree with in the literature—spark ideas for my papers.
I’m completing a PhD in Information and Communication Technologies at the University of Modena and Reggio Emilia, supervised by Prof. Sonia Bergamaschi. My research within the ITSERR project focuses on how LLM systems behave around safety boundaries, incomplete information, sensitive data, and external evidence.
I’m particularly interested in whether a refusal is justified, how retrieval changes a model’s decision, and whether an agent recognizes what it needs from a user before acting. I have also worked with Dr. Georgia Koutrika at DARE Lab on Text-to-SQL evaluation.
LLM evaluation and benchmarking · Retrieval-augmented generation · Text-to-SQL · Multi-agent data generation and validation · Synthetic data
Information and Communication Technologies · DBGroup
Department of Engineering “Enzo Ferrari”
Supervisor: Prof. Sonia Bergamaschi
PhD cycle ends October 2026 · defence expected early 2027
12 April–12 July 2025 · Athens, Greece
Text-to-SQL evaluation with Dr. Georgia Koutrika
Computer Engineering · Artificial Intelligence and Intelligent Robotics
110/110 cum laude
Computer Engineering · 110/110
Fisciano, Italy
Machine learning for public transportation monitoring and prediction
Let’s talk research