Applied Scientist II · Amazon Alexa AI
Building large language and vision-language models. PhD researcher in AI at MICC, University of Florence.
I'm an Applied Scientist II at Amazon Alexa AI in Turin, Italy, where I work on large language models that make Alexa smarter for millions of users worldwide.
I'm also a PhD researcher in Artificial Intelligence at MICC, University of Florence, advised by Prof. Andrew D. Bagdanov and Prof. Marco Bertini. Before joining Alexa, I was an Applied Scientist Intern with the Amazon Rufus team in London, fine-tuning, evaluating and deploying multimodal LLMs.
I study multimodal vision-language models such as CLIP and large language models: how their representations align across modalities, and how to adapt them efficiently without forgetting what they already know.
Understanding and bridging modality gaps in vision-language models.
Fine-tuning, evaluating and deploying LLMs and MLLMs at scale.
Improving zero-shot generalization without labeled data.
Learning incrementally while preventing catastrophic forgetting.
Efficient model adaptation with minimal supervision.
Two papers accepted at NeurIPS 2026: WASP (main track) and GeoBiaset (Evaluations & Datasets track).
Joined Amazon Alexa AI in Turin as an Applied Scientist II, working on LLMs for Alexa.
IsoCLIP, on intra-modal alignment in CLIP, is accepted at CVPR 2026.
SpectralGCD, on generalized category discovery, is accepted at ICLR 2026.
Started an Applied Scientist Internship with the Amazon Rufus team in London (until Dec 2025).
Cross the Gap, on intra-modal misalignment in CLIP, is accepted at ICLR 2025.
Presented KDPL, on unsupervised prompt learning, at ECCV 2024 in Milan.
KDPL source code is available.
Presented RE-tune at the NeurIPS 2023 Medical Imaging workshop.
Selected papers at top-tier AI venues. The full list is on Google Scholar. * equal contribution
NeurIPS 2026 Main track
Builds weakly aligned MRI–ultrasound pairs from unpaired fetal cohorts via optimal transport on clinical metadata, enabling cross-modal alignment without perfectly paired data.
NeurIPS 2026 Evaluations & Datasets
A counterfactual benchmark that measures demographic bias in world-level image geolocalization models.
Paper & dataset coming soon
CVPR 2026 Main conference
Shows how CLIP projectors induce distinct inter- and intra-modal operators, and finds an approximately isotropic subspace where image and text are well aligned.
ICLR 2026 Main conference
Represents images as mixtures over semantic concepts using CLIP cross-modal similarities, with spectral concept selection for generalized category discovery.
ICLR 2025 Main conference · First author
Intra-modal similarities from off-the-shelf VLMs are misaligned; solving image-to-image and text-to-text retrieval inter-modally via modality inversion significantly improves performance.
ECCV 2024 Main conference · First author
Unsupervised prompt learning that needs no labels or even class names, improving generalization across domains, datasets and base-to-novel classes by distilling from stronger models.
NeurIPS 2023 Workshop Medical Imaging meets NeurIPS · First author
Continual multi-label chest X-ray classification with frozen biomedical VLMs and lightweight adaptors trained from engineered text prompts.
Jul 2026 — Present · Turin, Italy
Working on cutting-edge large language models to make Alexa smarter for millions of users worldwide.
Jul 2025 — Dec 2025 · London, UK
Fine-tuned, evaluated and deployed multimodal LLMs for Amazon Rufus, impacting millions of customers.
Nov 2023 — Present
Multimodal vision-language models, incremental and prompt learning. Supervisors: A. D. Bagdanov, M. Bertini.
Sep 2021 — Jul 2023
Thesis: RE-Tune, incremental fine-tuning of biomedical VLMs. 110/110 cum laude.
Sep 2018 — Sep 2021
Thesis: AI-driven generation of Scarlatti sonatas. 105/110.
Jan 2024 — Jan 2026
Interactive lessons on C/C++ and Python for over 200 bachelor students.
Apr 2024 — Sep 2024
Few-shot CLIP fine-tuning that preserves zero-shot ability by distilling learned prompts into LoRA adapters.
Jan 2020 — Nov 2020
Mentoring students on exam projects, internships and career development.
NeurIPS 2024 · ICLR 2025 · NeurIPS 2025 · CVPR 2026
IEEE Transactions on Multimedia · IEEE Transactions on Circuits and Systems for Video Technology
Two pages with everything in one place. Updated October 2026.
Music has always been a big part of my life. I played the transverse flute for ten years in local bands and orchestras, sang with the Coro Giovanile Effetti Sonori (national competition winner), and I'm the guitarist and songwriter of I Green Clouds (provincial competition winner). I also enjoy strategic and cooperative games.
Based in Turin, Italy. Research affiliation: MICC, University of Florence. Email is the best way to reach me.
Institutional email
marco.mistretta@unifi.itPersonal email
marcomistretta99@gmail.comGoogle Scholar
Marco MistrettaGitHub
marcomistrettaX (Twitter)
@mistretta_marco