Reproduce “Evaluating Cooperation in LLM Social Groups through Elected Leadership” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ryan Faulkner, Anushka Deshpande, David Guzman Piedrahita, Joel Z. Leibo, Zhijing Jin

An investigation into multi-agent Large Language Model (LLM) social groups demonstrated that democratically elected leadership improves group sustainability and social welfare in common-pool resource management. Groups with elected leaders showed a 55.4% gain in Social Welfare and a 128.6% increase in Survival Time compared to no-leader scenarios.

Key insights

  • Elected leadership significantly improves sustainability and social welfare in LLM groups managing common-pool resources, outperforming both fixed leadership and no leadership scenarios.
  • LLM voters exhibit rational preferences, consistently favoring group-rewarding (Prosocial and Altruistic) leader personas, which correlates with improved resource survival.
  • Self-interested agents, even when not elected, can exert significant 'Importance Centrality' (influential outsiders), shifting group deliberation towards self-rewarding considerations, an effect termed 'Losing Voice'.
  • Leader personas display distinct rhetorical profiles aligned with their SVOs, with deceptive prompting notably altering altruistic leaders' persuasive strategies (e.g., increased Logos, decreased Ethos).

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.