CV

Research Scientist Internship (PhD Intern)

IBM Research

Natural Language Processing — LLM Factuality

• Created a novel benchmark for multimodal scientific fact-checking, targeting LLM long-form generation
• Coordinated a multi-team human annotation pipeline, overseeing annotation guidelines, quality control, and IAA
• Conducted systematic analysis exposing critical factuality fragilities in multimodal LLMs across scientific domains

Research Scientist Internship (Master Thesis)

IBM Research

Natural Language Processing — Named Entity Recognition

• Leveraged LLM distillation to improve smaller-sized models for zero-shot NER
• Produced two peer-reviewed publications — OpenBioNER (NAACL 2025) and ZeroNER (ACL 2025)
• Contributed to the IBM zshot library for zero-shot NER model inference

PhD

University of Bologna

Natural Language Processing

• Focus: LLMs, RAG, Information Extraction, Benchmarking, Low-resource NLP
• Supervisor: Prof. Gianluca Moro — Research Group: UniboNLP

MSc

University of Bologna

Artificial Intelligence

• Grade: 110/110 with honors

BSc

University of Bologna

Computer Science

• Grade: 110/110 with honors

LLMs (Almost) Never Abstain Under Medical Uncertainty

We introduce MedQAbstain, a benchmark for medical abstention under uncertainty, revealing that state-of-the-art LLMs systematically overcommit, rarely abstaining even when the question itself is hidden.

ReMedQA: Are We Done With Medical Multiple-Choice Benchmarks?

We show that high LLM accuracy in medical MCQA masks severe inconsistency. We propose novel metrics to evaluate true reliability across MCQA formats.

What do you call a dog that is incontrovertibly true? Dogma: Testing LLM Generalization through Humor

We introduce Phunny, a novel benchmark using uncontaminated English puns, revealing that LLMs struggle with generalization even on simple tasks, consistently underperforming the human baseline.

Can Large Language Models Win the International Mathematical Games?

We introduce MathGames, a novel multimodal benchmark of age-graded math problems from an international competition, showing that frontier LLMs underperform compared to humans, including 11-year-olds.

OpenBioNER: Lightweight Open-Domain Biomedical NER Through Entity Type Description

We introduce a 110M BERT model that leverages descriptions for zero-shot Biomedical NER, outperforming GPT-4o, specialized LLMs, and GLiNER by up to 10% F1.

To Generate or to Retrieve? On the Effectiveness of Artificial Contexts for Medical Open-Domain Question Answering

We introduce MedGENIE, the first generate-then-read framework for open-domain medical QA, demonstrating the effectiveness of generated over retrieved contexts and significantly improving LLM RAG performance.