Updates

Villanova.AI released Villanova2: a dense 2.4B and a 15.5B mixture of experts for Europe's languages, with open weights and documented open data. All four models cite Multi-SFT, our paper on permissively licensed multilingual instruction data.

In the publications The models

Tildea opened in early access: paste a job posting and it writes a resume for it from what you have really done, and leaves out anything you have not.

tildea.ai

Finished publishing seventeen research releases: every paper with code now has one repository and one model collection, cross-linked.

All seventeen releases

At CLiC-it 2026 in Palermo, the YUKAI! agent I built won Best Fooling AI Agent at Turing Hotel Italy: in the live Turing test in Italian, it was the agent most often mistaken for a human, especially in longer conversations. I also co-organised the competition, and the Villanova.AI team presented the Villanova-2B models.

Awards

Started as a postdoctoral researcher at DIISM, University of Siena, with Marco Gori and Marco Maggini, alongside my work at Villanova.AI and YUKAI!.

About

Taught Large Language Models for Developers at Zucchetti: twenty hours in English, from the transformer to production systems, with a lab for every lecture.

Teaching The course

The round-trip knowledge graph work was presented at ACL 2026, in the industry track.

In the publications Recording

Villanova.AI released Villanova-2B-2603, with Villanova-SFT-2603, the instruction-tuning data I built for it: 1,711,114 conversations in five languages, under Apache 2.0.

The dataset

Three papers at EACL 2026: PharmaQA.IT in the industry track, the multilingual study of plural nouns with no singular in the main conference, and the Sindhi reasoning work at the LoResLM workshop.

In the publications

Defended my Ph.D. at the University of Siena on 10 February, supervised by Marco Gori, with the grade eccellente. The thesis is open access.

The thesis

Villanova.AI released the first Villanova-2B models as a preview, a month after I joined the team.

2512 Preview

FarSense, a commonsense benchmark for Farsi, was presented at IJCNLP-AACL 2025, and Shawarma Chats, a dialogue benchmark in three Arabic varieties, at ArabicNLP 2025.

In the publications Recording

Joined Villanova.AI in Cagliari as an AI engineer, working on open European language models, and YUKAI! in Siena as a generative AI engineer.

About

Publicity co-chair of AIML Systems 2025 in Bengaluru, where the adaptive emission forecasting paper was also on the programme.

Service

Visiting researcher at the University of Gothenburg, on a multilingual study of plural nouns that have no singular, which became an EACL 2026 paper.

In the publications

Spoke at UNESCO Headquarters in Paris at LT4All 2025, the International Conference on Language Technologies for All, held under the theme of advancing humanism through language technologies.

Talks Conference

Visiting researcher at Inria in Nice, building the Detoxification Fidelity Index and a reinforcement learning pipeline for detoxification that preserves meaning.

At COLING 2025, two orals at the LoResLM workshop, on Arabic educational crosswords and on Persian multiple-choice questions, and ArabicSense at WACL-4.

In the publications

Visiting researcher at the University of Massachusetts Amherst, where I built a 22,000-pair human-verified dataset for detoxification that does not destroy the original meaning.

Gave the SAILab seminar on applications of large language models in education: crossword generation, and quizzes generated from text and images.

Seminar page

Talk on research directions, with Andrea Zugarini, at the TAILOR WebCrow workshop organised by CINI and Expert.ai.

Programme