Joined April 2024
- Tweets 149
- Following 193
- Followers 151
- Likes 140
32 Photos and videos
May 18
๐ Two papers at #SIGIR2026!
๐ฑ CATS: Clustering Thompson Sampling for smarter negative mining in RecSys. 20% NDCG@10.
โ๏ธ Load-sensitive Selective Pruning: dynamic embedding dimensionality for dense retrieval under load. 400% queries served.
See you in Melbourne! #IR #RecSys
1
4
182
Mar 27
Giulia Iadisernia is a first-year PhD student in Data Science at Sapienza, supervised by Fabrizio Silvestri a.
Her research focuses on the #trustworthiness and #explainability of LLMs.
She also collaborates with the Applied Research Team at the Bank of Italy.
1
3
203
Alessia Borghini is a first-year PhD student in the National Doctorate in AI supervised by Fabrizio Silvestri.
Her research focuses on Mechanistic Interpretability for #AISafety, aiming to understand LLMs and improve alignment and reliability.
Fell free to reach out!!
4
8
460
Feb 26
Your VLM gives the same answer before and after a tiny image change. So it's robust, right?
Wrong.
In our new paper, we show that VLMs can preserve their predictions while their internal representations drift to regions normally occupied by completely unrelated images.
๐งต๐
2
6
241
Feb 16
Gavriel Di Nepi is a first-year PhD student in Data Science at Sapienza under the supervision of prof. @fabreetseo . His research focuses on human-inspired memory for #AI to boost performance and reasoning, in collaboration with @bakerhughesco. #RSTLess #Sapienza
1
5
172
Meet Maria Vittoria Vestini, 1st-year PhD student in Data Science at Sapienza University of Rome, supervised by @fabreetseo. Her research explores interactions and learning dynamics between large and small language models.
Learn more: rstless.it/
#RSTLess #Sapienza
2
5
221
Directional Sheaf Hypergraph Networks: Unifying Learning on Directed and Undirected Hypergraphs accepted at #ICLR2026!!
Work done by Emanuele Mule, Stefano Fiorini, @AntonioPuri00, @_FedeSiciliano_ , Stefano Coniglio and @fabreetseo.
Paper: arxiv.org/pdf/2510.04727
2
3
8
263
By combining the expressive power of sheaves with a principled treatment of directionality, our method achieves consistent relative accuracy improvements over prior methods.
1
34
- We introduce the notion of Directed Cellular Sheaves over directed hypergraphs
- From this construction, we derive the Directed Sheaf Hypergraph Laplacian, a complex-valued Hermitian operator that unifies and generalizes existing graph and hypergraph convolutional operators.
3
81
Jan 30
๐๐ก๐ ๐๐๐ฃ๐จ๐ซ๐ข๐ญ๐ฒ ๐๐จ๐ญ๐ข๐ง๐ ๐๐๐ซ๐๐๐ข๐ ๐ฆ ๐๐ก๐ข๐๐ญ: ๐๐ก๐๐ง ๐๐จ๐ฉ๐ฎ๐ฅ๐๐ซ ๐๐๐๐ญ๐ฌ ๐๐ฉ๐ญ๐ข๐ฆ๐๐ฅ accepted at #AISTATS2026! Work done by @AntonioPuri00, @mariasofiabuc, Anil Kumar Nelakanti, @Andrea_Bacciu, @amantrac, @fabreetseo! Paper arxiv.org/pdf/2502.12581
1
6
11
222
Jan 30
It presents theoretical and empirical results answering to the question: "๐ถ๐๐ ๐๐๐๐๐๐๐ก๐ฆ ๐๐๐ก๐๐๐ ๐๐โ๐๐๐ฃ๐ ๐กโ๐ ๐กโ๐๐๐๐๐ก๐๐๐๐๐๐ฆ ๐๐๐ก๐๐๐๐ ๐๐๐๐๐ ๐๐ ๐ก๐๐๐๐ก๐๐๐ ๐๐๐๐๐ ๐๐๐ ๐ข๐๐๐๐ ๐คโ๐๐ก ๐๐๐๐๐๐ก๐๐๐๐ ?"
1
3
81
Jan 28
1/5 Composable Sparse Subnetworks via the Maximum-Entropy Principle will be presented at #ICLR2026 ๐
๐ openreview.net/forum?id=IHwxโฆ
Itโs about circuits, sparsity, and compositionality, with a training-time loss that shapes how circuits form. Congrats @frances55037016 !
4
1
6
466
Jan 28
5/5
This offers a complementary perspective on mechanistic interpretability:
from finding circuits to shaping and composing them during learning.
Looking forward to discussions at #ICLR!
Thanks to Samuele Fonio, Simone Monaco Nicola Saccomanno and @fabreetseo !
1
76
Jan 28
4/5
Crucially, these circuits are composable.
We show they can be recombined into generalist models, behaving like modular building blocks rather than isolated explanations.
26
Jan 28
3/5 We introduce a training-time objective that steers circuit formation toward specific concepts.
After training, we use pruning to extract class-specific circuits: functional subnetworks specialized for a subset of classes.
25
Jan 28
2/5 Most mechanistic interpretability work finds circuits post hoc, after training a model.
We ask a different question:
can we control which concepts shape circuits during training?
30
Jan 26
Paper accepted at #AISTATS2026!!
๐๐๐ง๐ค ๐๐ข๐๐ญ๐ข๐ง๐ ๐๐ง๐ ๐๐๐ง๐๐จ๐ฆ ๐๐จ๐ง-๐๐ข๐ง๐๐๐ซ ๐๐๐ฉ๐ฌ: once the size of a one-hidden-layer neural network exceeds a threshold, the hidden representations are linearly independent with high probability.
#RSTLess #Sapienza
1
1
6
168
Jan 26
Work done by: Andrea Drago, @mariasofiabuc , @frances55037016 , Marius Michetti, @_FedeSiciliano_ , @fabreetseo and Luca Becchetti.
1
1
55
Jan 19
Meet @Ali_Ghasemi7899 ๐ New PhD of our group under the supervision of @fabreetseo working on Efficient LLMs & LLM Agents exploring methods to optimize LLMs. He holds a MSc in Artificial Intelligence and Robotics (graduated with full marks and honors). #RSTLess #Sapienza #AI #LLM
1
5
136