Seminar: Graduate Seminar
On Artificial Agents that Learn to Communicate
Date:
February,18,2026
Start Time:
10:30 - 11:30
Location:
506, New Zisapel Building
Zoom:
Zoom link
Add to:
Lecturer:
Boaz Carmeli
Research Areas:
| We investigate how artificial agents develop and use communication protocols in task-oriented settings, with a particular focus on the relationship between language structure, compositionality, and generalization. Using referential games as a unifying experimental paradigm, this work studies communication that emerges both between neural agents and within large pre-trained visionโlanguage models. Across several contributions, we examine when and how compositional structure arises, how it can be measured, and whether it is necessary for successful coordination and generalization. Our findings show that strong task performance and generalization can emerge even in the absence of classical compositionality, while alternative mechanisms, such as structured reuse, decomposition strategies, and task-specific language variants, play a central role. Taken together, these results challenge standard assumptions about the role of natural-language-like structure in learned communication and suggest a broader view of communication as an adaptive interface shaped by task demands rather than linguistic form. |
| Boaz Carmeli is a PhD candidate at the Technion and a Research Staff Member at IBM Research Haifa. His research focuses on emergent communication, compositionality, and task-oriented language in neural and visionโlanguage models, with an emphasis on referential games as a framework for studying communication and generalization. He holds B.Sc. and M.Sc. degrees from the Technion and has extensive experience in industrial research, including leadership roles in applied machine learning, natural language processing, and AI-driven healthcare systems.
Ph.D. student Under the supervision of Prof. Ron Meir.
|