Donato Crisostomi

Donato Crisostomi

AI Safety Fellow @ Anthropic

Biography

Hello!

I gave up on keeping an up-to-date bio that I don’t find cringe 6 months later, so I’ll just list facts.

I am the PI of a Coefficient Giving grant on modular steering vectors, and I joined Anthropic as an AI Safety Fellow in August 2026.

I got my ELLIS Ph.D. from Sapienza, advised by Prof. Rodolà and co-supervised by Prof. Liò at the University of Cambridge. I was previously at Cohere, Amazon Search and Amazon Alexa.

I am interested in mech interp, weight space techniques and model merging.

You can find my papers on Scholar, and (some of) my opinions on X.

Education

  • Ph.D. in Computer Science, 2026

    ELLIS PhD @ Sapienza University & University of Cambridge

  • MS.c. in Computer Science, 2021

    Sapienza University

    110/110 with honours | GPA 30.5/31

  • BS.c. in Computer Science, 2019

    Sapienza University

    110/110 with honours | GPA 29.6/31

Experience

  • AI Safety Fellow

    Aug 2026 - Nov 2026

    Anthropic

  • Research Scholar

    Dec 2025 - Apr 2026

    Cohere

  • Applied Scientist

    Jun 2022 - Dec 2022

    Amazon Search

  • Research Scientist

    May 2021 - Oct 2021

    Amazon Alexa

News

I’ll spend the next 4 months in London for the Anthropic AI Safety Fellowship!
Aug 3, 2026

Two papers accepted at ICML 2026! 🇰🇷 See you in Seoul!

Jun 14, 2026
I’ve started as a Research Scholar at Cohere! Working on Large Language Models for search and retrieval.
Dec 1, 2025
Attended the amazing EEML 2025 and won a best poster award! 🏅 Check it here
Jul 28, 2025
Mentored a great team at LOGML 2025 and helped them achieve the best project award! 🏆 Check it here 👀
Jul 20, 2025
Our paper “Mergenetic: a Simple Evolutionary Model Merging Library” has been accepted at ACL 2025, System Demonstration Track. Come evolve your own LLMs on Mergenetic 🧬!
May 23, 2025
Both MERGE3 and Update Your Transformer were accepted at ICML 2025 main track! 🙏
May 2, 2025
Check our new preprint MASS to see MoEs and model merging in action 💥
May 1, 2025
Audio pre-prints 🎧! Check these out to see stemmed accompaniment generation, how to generate seamless loops and activation patching for steering music generation.
Apr 20, 2025
See all

Selected Publications

The Independence Prior of SAEs Fragments Visual Concepts
Preprint
Sparse Autoencoders (SAEs) decompose model activations into sparse combinations of interpretable dictionary atoms. Although SAEs are …
Tommaso Mencattini, Giorgos Nikolaou, Donato Crisostomi, Thomas Fel, Francesco Montagna, Emanuele Rodolà, Francesco Locatello
arXiv
The Independence Prior of SAEs Fragments Visual Concepts
Language Models are Injective and Hence Invertible
ICLR 2026
Transformer components such as non-linear activations and normalization are inherently non-injective, suggesting that different inputs …
Giorgios Nikolaou, Tommaso Mencattini, Donato Crisostomi, Santilli, Yannis Panagakis, Emanuele Rodolà
Cite arXiv Thread
Language Models are Injective and Hence Invertible
MASS: MoErging through Adaptive Subspace Selection
ICLR 2026
Model merging has recently emerged as a lightweight alternative to ensembling, combining multiple fine-tuned models into a single set …
Donato Crisostomi, Alessandro Zirilli, Antonio Andrea Gargiulo, Maria Sofia Bucarelli, Simone Scardapane, Fabrizio Silvestri, Iacopo Masi, Emanuele Rodolà
Cite arXiv
MASS: MoErging through Adaptive Subspace Selection
Task Singular Vectors: Reducing Task Interference in Model Merging
CVPR 2025
Task Arithmetic has emerged as a simple yet effective method to merge models without additional training. However, by treating entire …
Antonio Andrea Gargiulo, Donato Crisostomi, Maria Sofia Bucarelli, Simone Scardapane, Emanuele Rodolà
Cite arXiv GitHub Tweeprint
Task Singular Vectors: Reducing Task Interference in Model Merging
See all