Roberto Dessì

About

I'm a Research Scientist at Not Diamond, where I work on model routing for AI agents — building systems that decide which model to call on which task to balance cost and capability over long-horizon, dynamic workloads.

Previously, I was a Member of Technical Staff at Samaya AI, working on retrieval for an LLM-based knowledge discovery platform. I completed my PhD cum laude at Universitat Pompeu Fabra in Barcelona in 2023, as part of a joint industrial doctorate with Meta AI (FAIR) in Paris, supervised by Marco Baroni. My thesis was on language emergence in multi-agent systems. Before that, I studied Cognitive Science at the University of Trento and SISSA, and earned a B.S.Eng. in Computer Science Engineering from La Sapienza in Rome.

My research interests span large language models, multi-agent systems, and emergent communication. You can find a more detailed CV here.

News

Sep 2025Joined Not Diamond as a Research Scientist, working on model routing for coding agents.
Oct 2025Our work on self-supervised representations of dolphin vocalizations (Dolph2Vec) accepted at the AI for Non-Human Animal Communication workshop at NeurIPS 2025.
Oct 2025Model Merging Improves Zero-Shot Generalization in Bioacoustic Foundation Models (D. Marincione, D. Crisostomi, R. Dessì, E. Rodolà, E. Rossi) accepted at the same workshop.
2025New papers in TMLR 2025 (Referential communication in heterogeneous communities of pre-trained visual deep networks; I Want to Break Free!) and *SEM 2025 (Explanations explained).
Jun 2024Joined Samaya AI as a Member of Technical Staff.
Nov 2023Defended my PhD at Universitat Pompeu Fabra (cum laude).
Sep 2023Toolformer accepted at NeurIPS 2023 as an Oral.
Feb 2023Cross-Domain Image Captioning with Discriminative Finetuning accepted at CVPR 2023.

Experience

Sep 2025 – Present

Research Scientist · Not Diamond

Jun 2024 – Present

Member of Technical Staff · Samaya AI

Dec 2023 – May 2024

ML Engineer, Intern · Samaya AI

Nov 2019 – May 2023

Research Scientist, PhD student · Meta AI (FAIR), Paris

Aug 2019 – Oct 2019

Research Scientist, Intern · Amazon — Alexa TTS

Jul 2018 – Oct 2018

Research Scientist, Intern · Fondazione Bruno Kessler — MT group

Jun 2016 – Dec 2016

Research Engineer, Intern · Xerox Research Centre Europe

Education

2019 – 2023

PhD, Language Sciences · Universitat Pompeu Fabra, Barcelona

cum laude, Joint industrial PhD with Meta AI (FAIR), Paris.

2018 – 2019

Diploma in Cognitive Neuroscience · SISSA, Trieste

2017 – 2019

MSc, Cognitive Science · CIMeC, University of Trento

Final grade: 110/110 cum laude. Recipient of the UniTrento–SISSA scholarship.

2013 – 2017

B.S.Eng., Computer Science Engineering · University of Rome, La Sapienza

Final grade: 108/110. Spring 2017 semester at the University of Buffalo.

Selected Publications

Toolformer: Language Models Can Teach Themselves to Use Tools T. Schick, J. Dwivedi-Yu, R. Dessì, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, T. Scialom NeurIPS 2023Oral

Augmented Language Models: a Survey G. Mialon, R. Dessì, M. Lomeli, C. Nalmpantis, R. Pasunuru, R. Raileanu, B. Rozière, T. Schick, J. Dwivedi-Yu, A. Celikyilmaz, E. Grave, Y. LeCun, T. Scialom TMLR 2023

Cross-Domain Image Captioning with Discriminative Finetuning R. Dessì, M. Bevilacqua, E. Gualdoni, N. C. Rakotonirina, F. Franzon, M. Baroni CVPR 2023

Communication breakdown: On the low mutual intelligibility between human and neural captioning R. Dessì, E. Gualdoni, F. Franzon, G. Boleda, M. Baroni EMNLP 2022

Interpretable Agent Communication from Scratch (with a generic visual processor emerging on the side) R. Dessì, E. Kharitonov, M. Baroni NeurIPS 2021

CNNs found to jump around more skillfully than RNNs: Compositional generalization in seq2seq convolutional networks R. Dessì, M. Baroni ACL 2019

Other Publications

A. Bruera, R. Dessì. Sparse Autoencoders Reveal Representations of Cognitively Motivated Semantic Categories in LLMs, Under review
L. Zhou, S. Shah, E. Rodolà, R. Dessì. Hard or Just Unreached? Diagnosing the Sampling Blind Spot in Math-Reasoning Difficulty Estimation, Under review
F. Mustun, C. Semenzin, R. Dessì. P. Robin Guerrero, P. Orhan, A. Emanuelli, E. Rossi, Y. Lakretz, G. de Polavieja, G. Sumbre. OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations, Under review
L. Ferrarotti, G. M. Campedelli, R. Dessì, A. Baronchelli, G. Iacca, K. M. Carley, A. Pentland, B. Lepri. Generative AI collective behavior needs an interactionist paradigm, Under review
Y. Aluma, Z. Baron, R. Barrett, C. Baumgartner, G. Beguš, S. Bhattacharya, M. M. Bronstein, S. Dahan, O. Davis, S. De Haas, J. Defoe, J. DelPreto, R. Dessì, R. Diamant, J. Gatesy, K. George, S. Gero, D. Gibbons, D. Gibbons, S. Gil, S. Goldwasser, D. F. Gruber, O. Harve, A. Hernandez, M. Ishay, N. Jadhav, L. Kc, A. Kenny, A. Leitao, M. Lucas, A. Maalouf, P. Malkin, Y. Mevorach, S. Pagani, O. Paradise, G. Petri, S. Poetto, E. Rossi, D. Rus, M. Salino-Hugg, A. Santoro, P. Sharma, D. Tchernov, A. Torralba, P. Tønnesen, D. M. Vogt, R. J. Wood. Description of a collaborative sperm whale birth and shifts in coda vocal styles during key events Scientific Reports (2026)
M. Mahaut, F. Franzon, R. Dessì, M. Baroni. Referential communication in heterogeneous communities of pre-trained visual deep networks, TMLR 2025 (earlier version at AAMAS 2023)
C. Semenzin, F. Mustun, R. Dessì, P. Orhan, G. de Polavieja, Y. Lakretz, G. Sumbre. Dolph2Vec: Self-Supervised Representations of Dolphin Vocalizations, AI for Non-Human Animal Communication workshop @ NeurIPS 2025
A. Zaninello, R. Dessì, M. Nissim, B. Magnini. Explanations explained. Influence of Free-text Explanations on LLMs and the Role of Implicit Knowledge, *SEM 2025
G. M. Campedelli, N. Penzo, M. Stefan, R. Dessì, M. Guerini, B. Lepri, J. Staiano. I Want to Break Free! Persuasion and Anti-Social Behavior of LLMs in Multi-Agent Settings with Social Hierarchy, TMLR 2025
S. Sapora, C. Baziotis, R. Dessì, F. Petroni, M. Bevilacqua. QCR: Quantised Codebooks for Retrieval, arXiv
S. Goodarzitaeme, N. Kagita, D. Minn, S. Wang, R. Dessì, S. Toshniwal, A. Williams, J. Lanchantin, K. Sinha. Robustness of Named-Entity Replacements for In-Context Learning, Findings of EMNLP 2023
N. C. Rakotonirina, R. Dessì, F. Petroni, S. Riedel, M. Baroni. Can Discrete Information Extraction Prompts Generalize Across Language Models? ICLR 2023
R. Chaabouni, R. Dessì, E. Kharitonov. Can Transformers Jump Around Right in Natural Language? Assessing Performance Transfer from SCAN, BlackBoxNLP @ EMNLP 2021
R. Dessì, D. Bouchacourt, D. Crepaldi, M. Baroni. Focus on What's Informative and Ignore What's Not: Communication Strategies in a Referential Game, Emergent Communication workshop @ NeurIPS 2019
M. A. Di Gangi, M. Negri, R. Cattoni, R. Dessì, M. Turchi. Enhancing Transformer for End-to-End Speech-to-Text Translation, MT Summit 2019
M. A. Di Gangi, R. Dessì, R. Cattoni, M. Negri, M. Turchi. Fine-tuning on Clean Data for End-to-End Speech Translation: FBK @ IWSLT 2018, IWSLT 2018

Selected Invited Talks

Nov 2023 — Talk at the Data Science and AI seminar, Chalmers University of Technology, Gothenburg.
Jul 2023 — Talk at the SARDINE lab, Instituto Superior Técnico, Lisbon.
Jun 2023 — Talk at the ILLC, University of Amsterdam.
Jun 2023 — Talk at the Berkeley NLP Group.
Jun 2023 — Talk on Emergent Communication at the second CETI workshop on Decoding Communication in Nonhuman Species, Simons Institute, UC Berkeley.
Apr 2023 — Talk at LIGM, École des Ponts ParisTech.
Mar 2023 — Talk at the Cognitive Science seminar, Queen Mary University.
Jun 2022 — Talk on emergent communication at the Reinforcement Learning Symposium, University of Bristol.

Community Service

Tutorial on Emergent Language-Based Coordination in Deep Multi-Agent Systems at EMNLP 2022, with Marco Baroni and Angeliki Lazaridou.
Reviewer for ICML (2026, 2025, 2023); NeurIPS (2026, 2025, 2023, 2022, 2021); ICLR (2026, 2024, 2022 — Highlighted Reviewer); ACL Rolling Review (NAACL 2022, EMNLP 2022); COLM 2026, 2025; LREC-COLING 2024; Frontiers in Robotics and AI.
Tutorials reviewer for NAACL-HLT / ACL / EMNLP 2025.
Organizer of the 4th (NeurIPS 2020) and 5th (ICLR 2022) Emergent Communication workshops.
(Update: former maintainer, the repo is now unsupported) Maintainer of EGG: Emergence of lanGuage in Games, the main public repository for language emergence experiments (300+ stars).