By placing up to 100,000 agents from four LLMs on different networks, physicists show that topology dictates the speed of consensus and vulnerability to manipulation.
As language models are deployed in interacting groups, their collective behavior also depends on who is talking to whom. Roberto Passaro, Ariel Flint, Romualdo Pastor-Satorras, Andrea Baronchelli, and their colleagues played the "naming game," a classic in the physics of social systems, policies extracted from four LLMs, spread over 1,000 to 100,000 agents. Where previous studies assumed that anyone can intersect with everyone, they tested fully connected, regular, and highly heterogeneous random networks.
Hubs, the collective's weak point
The structure of the network changes the speed at which consensus emerges, sometimes its very existence, and more rarely the winning convention. The most sparse networks slow down agreement by up to twelve times, but, by maintaining stable neighborhoods, they sometimes make it possible to achieve agreement where complete mixing fails.
On the security side, the authors measured the critical mass of "converted" agents needed to switch an established convention. On the most heterogeneous network, it is a median of 3% if we target hubs, 16% at random and 35% for poorly connected nodes: targeting hubs requires six times fewer agents. For most models, networked populations switch over more easily than well-mixed populations, so the latter give an overly optimistic picture of the robustness of multi-agent systems.
Read the full article on the original source
Go to the source →