DATABANK

Jonathan Elsworth Eicher[govsim]

Models now saturate cooperation in GovSim

gpt-6-astra has now saturated cooperation in the GovSim benchmark when original models at most only survived in half of the scenarios.

Baseline runs that survive to month twelve, by model

Piatti et al. reported "all but the most powerful LLM agents fail to achieve a sustainable equilibrium" but their best model of the time, GPT-4o, only reached that equilibrium 53% of the time. With an extension it has become clear that we have reached the peak cooperation with gpt-6-astra which survives in all baseline scenarios.

785 runs

GovSim: Emergence of Sustainable Cooperation in a Society of LLM Agents

Five LLM agents share a common-pool resource and must agree on limits to avoid collapse. A replication of Piatti et al. (NeurIPS 2024) across three scenarios. This finding uses baseline treatment in all three scenarios, every model, completed full-scenario runs.

Cite this finding

BibTeX
@misc{eicher-2026-models-now-saturate,
  author = {Jonathan Elsworth Eicher},
  title = {Models now saturate cooperation in GovSim},
  year = {2026},
  howpublished = {Multi-Agent Databank},
  url = {https://databank.antimemetics.institute/findings/cooperation-saturates-beyond-the-paper},
  note = {Retrieved 2026-10-19}
}