Paolo Mandica

I am an AI Research Scientist at the Samsung AI Center in Warsaw, Poland, where I focus on the Safety and Alignment of large multimodal models.

I hold a Ph.D. in Computer Science from Sapienza University of Rome, where I was part of the Perception and Intelligence Lab led by Prof. Fabio Galasso. My research experience includes a year-long internship at Panasonic North America in Mountain View, and a visiting researcher position at UC Berkeley in Trevor Darrell’s lab at BAIR. I also served as Lead Researcher and Founding Member of ItalAI, an Italian startup dedicated to AI research.

I earned a Master’s degree in Data Science from Sapienza University of Rome, graduating in 2021 with top honors (110/110 cum laude). Before that, I completed my Bachelor’s degree in Computer Science and Engineering at Roma Tre University in 2019. During my undergraduate program, I gained practical experience in Cloud Engineering through an internship at Claranet IT, where I developed a serverless application for monitoring event-driven architectures on AWS.

Outside of research, I enjoy spending time in nature, reading, photography, running, playing guitar, and working out at the gym.

GitHub LinkedIn Scholar CV

Get in contact with me at paolomandica.p [at] gmail [dot] com

Last updated September 2025, from Warsaw, Poland.

Now

September 20, 2025 — Our paper Representation-based Broad Hallucination Detectors Fail to Generalize Out of Distribution has been accepted to EMNLP 2025 Findings!

View past updates →

Research

I currently work on AI Safety for LLMs and VLMs, focusing on hallucination detection, model reliability, and trustworthy evaluation. The focus of my past research has been on deep learning for computer vision and uncertainty estimation, with interests spanning self-supervised learning, multimodal systems, active learning, and hyperbolic neural networks. I have hands-on experience in action recognition, domain adaptation, and semantic segmentation, often under low-label or cross-domain scenarios.

Representation-based Broad Hallucination Detectors Fail to Generalize Out of Distribution
Z. Dubanowska, M. Żelaszczyk, M. Brzozowski, Paolo Mandica, M. Karpowicz
EMNLP 2025 Findings

[arXiv]
HALO: Hyperbolic Active Learning for Semantic Segmentation under Domain Shift
L. Franco*, Paolo Mandica*, K. Kallidromitis, YT Li, D. Guillory, T. Darrell, F. Galasso
ICML 2024

[arXiv] [OpenReview] [Github]
Hyperbolic Learning with Multimodal Large Language Models
Paolo Mandica*, L. Franco*, K. Kallidromitis, S. Petryk, F. Galasso
ECCV 2024 - Beyond Euclidean Workshop

[arXiv]
HYSP: HYperbolic Self-Paced Learning for Self-Supervised Skeleton-based Action Representations
L. Franco*, Paolo Mandica*, B. Munjal, F. Galasso
ICLR 2023

[arXiv] [OpenReview] [Github]

* indicates equal contribution

Activities

Teaching assistant for these Master's degree courses in Data Science at Sapienza University:

Reviewer for the following conferences and journals:

Participated at the these events:

ECCV 2024 - European Conference on Computer Vision - Milan, Italy
ICML 2024 - International Conference on Machine Learning - Vienna, Austria
CVPR 2023 - Computer Vision and Pattern Recognition Conference - Vancouver, Canada
ICVSS 2022 - International Computer Vision Summer School - Scicli (RG), Italy