Graph Feedback Controls Consensus and Clique Formation in Open-Weight Language-Model Populations

Short summary

This study examines how runtime interaction graphs affect convention formation in open-weight language-model populations (1.1B-32B parameters) using a naming-game protocol. Homophilous threshold-similarity routing amplifies fragmentation, while bridge-seeking routing can repair it when memory is available. In homogeneous populations like Qwen2.5-32B, retained history shifts fragmented dynamics toward consensus, achieving stable behavioral and state consensus across all well-mixed settings.

• •Interaction graph topology controls consensus vs fragmentation in multi-agent LM populations
• •Homophilous routing amplifies fragmentation; bridge-seeking routing repairs it with memory
• •Qwen2.5-32B reaches full consensus in all retained-history well-mixed settings while threshold-similarity fails in 189 settings

Generated with AI, which can make mistakes.