MSc Data Science @ EPFL,
Research @ Gladia, ISTA
Research @ Gladia, ISTA
Tommaso Mencattini
Hey! I’m Tommaso, currently doing a Master’s in Data Science at EPFL.
Between lectures and problem sets, I spend my time at ISTA and GLADIA working on LLMs, mainly figuring out how to make them more interpretable and safer.
My academic path is a bit unconventional: I hold a BSc in AI and Mathematics and a BA in Philosophy. That means I probably know more math than most philosophers, more computer science than most mathematicians, and more philosophy than most computer scientists. Not entirely sure if that’s a strength or just confusing, but it definitely keeps things interesting.
news
| Jan 28, 2026 | Our paper Exploratory Causal Inference in SAEnce received an oral presentation 🏆 at ICLR 2026! |
|---|---|
| Jan 26, 2026 | Our paper Language Models Are Injective and Hence Invertible has been accepted at ICLR 2026. |
| Oct 10, 2025 | New preprint out! Language Models Are Injective and Hence Invertible We prove that LLM representations are injective, and present the first exact inversion algorithm. The announcement blew up with 4.9M views on Twitter ❌. Check it out! |
| Jul 01, 2025 | Excited to be joining ISTA this summer as a Research Intern in the Locatello Group working on multimodal foundation models and causal learning! |
| May 23, 2025 | Our paper Mergenetic: a Simple Evolutionary Model Merging Library was accepted at ACL 2025 (System Demonstrations)! |
talks
| Jan, 2026 Invited Talk | Presented Language Models Are Injective and Hence Invertible at Apple @ Aachen (S. Peitz Group) |
|---|---|
| Dec, 2025 Invited Talk | Gave a talk on Language Models Are Injective and Hence Invertible at Baker Hughes |
| Nov, 2025 Invited Talk | Discussed Language Models Are Injective and Hence Invertible at Area Science Park (University of Trieste) |
libraries
- A flexible library for merging large language models (LLMs) via evolutionary optimization. It frames model merging as a black-box optimization problem and uses techniques like genetic algorithms and smart performance estimators to search for optimal weight combinations.
