Seminari di Ingegneria Informatica - Roberto Dessi': Towards language-based communication in AI systems: from emergent protocols to agentic AI
Speaker:
Roberto Dessi' (Not Diamond)
Data dell'evento:
Monday, 19 October, 2026 - 11:30
Luogo:
DIAG - Sapienza, Room B203 - Zoom Link: https://uniroma1.zoom.us/j/85485829872
Contatto:
Stefano Leonardi - leonardi@diag.uniroma1.it
Speaker: Roberto Dessi' (Not Diamond)
Title: Towards language-based communication in AI systems: from emergent protocols to agentic AI
When: Monday, October 19, 11:30 AM
Where: DIAG - Sapienza, Room B203
Zoom Link: https://uniroma1.zoom.us/j/85485829872
Abstract: Large language models are fluent, but fluency is not the same as being able to interact. This talk looks at AI agents through the lens of communication: how they coordinate with other agents, how their language lands with human partners, and how they reach beyond their parameters into external tools and knowledge. I first show that neural networks trained from scratch on a referential game develop a discrete communication protocol that is both effective and interpretable to humans. I then turn a setup with humans in the loop: model-generated language that looks natural can fail badly with human listeners, and optimizing the speaker for its listener recovers language that works. Finally, I present Toolformer, in which language models teach themselves to call external tools, a result that helped launch the current field of agentic AI. I close with a proposal for making AI agents more interaction-ready: treating interaction as a first-class objective throughout the training of language models, rather than a final-stage patch.
Biosketch: Roberto Dessì is a founding Research Scientist at Not Diamond, where he works on evaluation and model routing for long-horizon coding agents. Previously he was a Member of Technical Staff at Samaya AI, building production retrieval-augmented LLM systems. He completed an industrial PhD cum laude between Meta AI (FAIR Paris) and Universitat Pompeu Fabra, supported by the Meta–UPF industrial-doctorate scholarship. His research spans emergent communication, human–model interaction, and tool-augmented language models; he is a co-author of Toolformer, a contribution that helped launch the recent field of agentic AI. His work has been published at leading AI and NLP venues, including NeurIPS, ICLR, CVPR, ACL, EMNLP, and TMLR.
He holds an MSc in Cognitive Science from the University of Trento, a Diploma in Neuroscience from SISSA, and a B.Eng in Computer Science Engineering from Sapienza University of Rome.
He has co-organized two editions of the Emergent Communication workshop (NeurIPS, ICLR) and delivered a tutorial on emergent communication at EMNLP.