What happens when AI agents are asked to debate the same question several times? Do they follow similar paths, or can small differences lead to very different outcomes?
In this talk, we explore how multiple Large Language Models interact, change their opinions, and move toward agreement or disagreement during a debate. By repeating the same conversations across different models, we can study how stable and reproducible their collective behavior really is.
Using a simple geometric representation of opinions and a dynamical model, we show that different AI models can behave in surprisingly different ways, even when starting from the same conditions.
The goal is not only to understand what answer AI agents reach, but how they get there.
Speaker: Leonardo Mascagni – collaboratore ICT Lab (il gruppo di ricerca interdisciplinare dedicato a Intelligence, Complexity and Technology) e futuro dottorando
