State Encodings Shape Collective Dynamics in Language-Model Agents

A recent preprint on arXiv (2608.06968) examines how language-model agents' encoding of environmental states affects their group behavior. The researchers conducted a circular-synchronization experiment where each agent only received a summary of its neighbors' relative phases, deciding to advance, stay, or retard. While the physical system remained constant, the state encoding varied between low-order circular moments and a histogram. In GPT, the moment encoding achieved synchronization in all 6 seeds, whereas the histogram encoding did not synchronize in any. Claude replicated the effect but in the opposite direction. Replaying identical fields significantly altered the probabilities of each agent's decisions across GPT, Claude, and Gemini. The results indicate that state encodings are model-specific mechanisms influencing collective outcomes, with implications for managing multi-agent AI systems.